Machine Transalation.pdf

Abstract
We describe an implementation of an Example-Based Machine Translation system that
translates short sentences from Arabic to English. The system uses a large parallel
corpus, aligned at the paragraph level, and built using many parallel Arabic-English
documents from the United-Nations database.
This is a non-structural system, so examples are stored in their surface forms, with some
additional morphological and part-of-speech information. Each new input sentence is
matched to example patterns by using various levels of morphological data. Matched
fragments are transferred to English using a rough word-level alignment algorithm, which
makes use a bilingual dictionary, along with WordNet. There is no syntactical parser
involved in the matching and/or transfer processes, although we do use a shallow English
parser to help with specific situations.
Currently, the system first fragments any newly introduced input sentence and then
translates each segment separately; recombining those translations into a final coherent
form is left for future work.
We encountered several problems in the matching and the transfer steps, some of which
were solved, partially or totally, sometimes by using linguistic tools for both languages.
We discuss those problems and our proposed solutions.
The system has been implemented and automatically evaluated. Results are encouraging.

Table of Contents
1. Chapter 1 – Introduction 4
1.1 Classifying Machine Translation Systems 5
1.2 Example-Based Machine Translation (EBMT) 7
1.3 Overview of Arabic to English Machine Translator 9
1.4 Thesis Overview 13
2. Chapter 2 – Short Introduction to Arabic 15
2.1 Arabic Orthography 16
2.2 Arabic Morphology and Syntax 16
3. Chapter 3 - Natural Language Processing 19
3.1 Morphological Tools 19
3.1.1 Morphological Analysis 21
3.2 Syntax Tools 23
3.2.1 Shallow Parsing 27
4. Chapter 4 – Extracting Translation Examples 30
4.1 Preparing the Corpus 30
4.2 Parallel Text Alignment 31
4.2.1 Finding Parallel Text Anchors 32
4.2.2 DK-Vec Matching 34
4.2.3 Extracting Anchors 38
4.2.4 Extracting Corresponding Paragraphs 38
5. Chapter 5 - Translation Algorithm 41
5.1 Preprocessing and Data Preparation 41
5.2 Matching 42
5.2.1 Fragment Score Calculation 45
5.2.2 Fragments Storing 47
5.3 Transfer 48
5.3.1 Step 1 – Translation Extraction 48
5.3.2 Step 2 – Fixing the Translation 54
5.3.3 Choosing a General Fragment’s Translation 58
5.4 Recombination 59
5.4.1 The Recombination Algorithm 59
5.5 Example 62
5.5.1 Preprocessing and Data Preparation 63
5.5.2 Matching 63
5.5.3 Transfer 65
5.5.4 Recombination 68
6. Experimental Results and Conclusions 71
6.1 Results 71
6.2 Future Work 73
6.3 Conclusions 74
References 75
Appendix 81

1. Chapter 1 – Introduction
One of the oldest challenges since computers were invented is Machine-Translation
(sometimes referred as “Automatic-Translation”), that is, translating a text from one
natural language (the source-language) into another one (the target-language) using
computers. The text might be a word, or sentence or even an entire text document. Early
automatic-translation approaches focused on performing what is called “Fully Automatic
High Quality Translation (FAHQT).” The output of such systems is designed to be high-
quality coherent target-language text that exactly translates the source-language input
text. These approaches were criticized in the famous Bar Hillel report [1], issued in 1960,
claiming that developing a system for high-quality translations is utterly futile. In his
report, Bar Hillel argued that the term “High Quality” should be discarded for a system
that performs a fully automatic translation process. Such a system may be useful for tasks
that require only rough translations, or should be considered only if there is a manually
post-editing step that finalizes the entire translation process to achieve a good quality
translation.
Since then, there has been much research devoted to various levels of machine-
translation, introducing new, interesting approaches using different levels of linguistic
analysis and/or large corpora.
In spite of the criticism of FAHQT, there are many working machine-translation systems
that have achieved satisfactory results. Most existing machine-translation systems are
designed to translate texts for some predefined domains, but there are also general
systems that work without any domain restriction, producing translations that can not, for
the most of the part, compete with high-quality human translations.

1.1 Classifying Machine Translation Systems
A machine translation system can be classified by how deeply it analyzes the source-
language text. This is usually described by the classic pyramid, first presented by
Vauquois [2] and shown in Figure 1.
Word
Structure
Word
Structure
Syntactic
Structure
Syntactic
Structure
Semantic
Structure
Semantic
Structure
Interlingua
Semantic Transfer
Syntactic Transfer
Direct
Arabic Text English Text
Morphological
Analysis
Morphological
Analysis
Syntactic
Analysis
Syntactic
Analysis
Semantic
Analysis
Semantic
Analysis
Figure 1 - Classification of a machine translation system

At the bottom of the pyramid are the direct machine-translation systems that produce the
translation using only some kind of word-level analysis. Those systems are good to for
dealing with a single target-language.
Higher up on this chart are the transfer systems, which first analyze the source-language
text and create corresponding structures. The analysis can be on the syntactic level or
even on the semantic level. Once the syntactic or semantic structure of the source-
language text has been captured, the system transfers the source-language structure to the
corresponding target-language structure. This may be done using various rule-based or
corpus-based methods, as will be described later. The last step is to generate the
translated text from the transferred target-language structure.
At the apex of this taxonomy of translation methods there are the Interlingua systems,
which first translate the source-language text into some kind of “universal” intermediate
language (either a common-logic representation or even a slightly modified natural-
language, along the lines of Esperanto), and then generate the target-language text from
the universal representation. Systems that were based on this approach were sympathized
by Bar Hillel, although he actually thought that the whole idea of automatic high quality
translation approach is misguided.
Usually, transfer and Interlingua systems are designed to deal with more than one pair of
languages, while direct systems are better suited of a single language-pair translation
system.
Another dimension for the classification of a machine-translation system is its research
paradigm. A survey of current machine-translation paradigms [3] describes two major
paradigms: linguistic-based and corpus-based. Linguistic-based paradigms mostly use a

predefined set of rules, with or without some kinds of linguistic knowledge-bases, to
produce the translation. Working with such paradigms requires comprehensive
knowledge of linguistic theory for both source and target languages. Linguistic-based
paradigms may be used by all the different kind of system types – direct, transfer and
Interlingua. Corpus-based systems exploit a large parallel bilingual corpus either for
offline statistical learning, helping to predict the translation of a new given input, or for
searching already translated examples and producing a recombined translation. The latter
paradigm is referred as Example-Based Machine Translation (EBMT), and is the key
concept of the work described in this thesis.
1.2 Example-Based Machine Translation (EBMT)
The example-based paradigm (also known as “memory-based”) has become a fairly
common technique for natural language processing (NLP) applications. The main idea
behind an example-based machine translation paradigm is to emulate the way a human
translator think in some cases, as was first introduced by Nagao [4] in 1984. Example-
based machine translation systems [5-10] exploit a large bilingual translation-example
corpus to find translations for fragments of the input source-language text. This step is
called matching. The corpus is created by aligning bilingual parallel texts on a phrase,
sentence or paragraph level. Given a group of matched fragments, the next step is to
extract their possible translations from the target-language side of the corpus. This step is
called transfer. The last step is recombination, which is the generation of a complete
target language text, pasting together the translated fragments. Figure 2 presents the main
steps of an example-based machine translation system.

Source Language
Text
Target Language
Text
Bi-lingual
Corpus
Matching
Transfer
Recombination
Figure 2 – Main steps of example-based translation system
There are example-based machine translation systems that parse the translation-examples
and also store their syntactic structure. Such systems are called structural. The matching
step in structural systems is done by first analyzing the input source-language sentence to
discover its syntactic structure, and then finding translation-examples that match on the
syntactic level. Non-structural systems, on the other hand, store the translation-examples
as pair of strings, with some additional information, usually morphological and/or part-
of-speech tags. Recalling the pyramid described above, a non-structural system is
considered to be a direct system, since it translates the source-language text using only
word-level information. However, a structural system is considered to be a transfer
system, since it first analyzes the source-language text to create some sort of syntactic
structure, transfers this structure to a target-language one and finally generates the

translation of the entire input source-language text. There are also structural systems that
learn from the translation-example corpus the syntactic differences between the two
languages and build set of syntactic transfer rules. Those systems are also called
“Example-Based Syntactic Transfer Systems.”
In the matching step, the system searches for corpus fragments that match a fragment of
the input text. In some systems, the match is performed on several levels, with each level
assigned a different score. Words levels may be morphologic (stem), syntactic (POS
tags), semantic (ontology/thesaurus distance), etc. Structural systems may also try to find
syntactic matches for entire sentences. Generating translations of those fragments would
involve modifications and fixes of the translation extracted from the target-language side
of the translation-examples.
For more information on this subject, the reader may refer to the comprehensive review
of example-based machine-translation by Somers [11].
1.3 Overview of Arabic to English Machine Translator
In our work, we implemented the important parts of an example-based machine
translation system that translates short Arabic sentences to English. This is a non-
structural system, so it stores the translation-examples as textual strings, with some
additional information, as will be described later. We used the United-Nations document
inventory to build our translation-example corpus. The inventory contains a large number
of documents that were manually translated into seven languages, and Arabic is one of
them. We automatically aligned each document pair on the paragraph level, and each
parallel paragraph was taken as a translation-example.

At this point, our system performs the fragmentation, matching and transfer steps fully
automatically. The recombination step only pastes together the extracted fragments, but
it does not yet smooth out the recombined text into a fully grammatical English sentence.
The entire system was written as a Java application. Special GUI tools were built to allow
external users to interact with the system in two modes: translation mode and corpus
creation mode. Currently, those tools require the installation of the entire system, but we
are planning to develop a web-interface allowing users to interact with the system
remotely.
In this work, we concentrated mostly on the design and implementation of the Arabic–to-
English example-based machine translation system, but did not concentrate on
implementing the components efficiently. More work would be needed to reduce the total
translation time.
Several external software packages were integrated into our system. The SVM-POS
package [12] is used to find the part-of-speech tag of each Arabic word, and Brill’s tagger
is used for part-of-speech tagging English words. Buckwalter’s Arabic morphological
analyzer [13] is used for several analysis procedures on the word level. The BaseNP
Chunker is used for shallow parsing the English part of the translation-examples, as will
be explained later.
Our system is composed of several modules. Figure 3 presents a high-level view of the
system’s main modules and their relations.

Preprocessing
Matching
Transfer
Recombination
Examples
Corpus
Input: Short Arabic Sentence
Output: N Best Translations (English)
Preprocessed Text
General Fragments
General Fragments +
Translations
Figure 3 –High-level observation
Our system was designed to handle short Arabic sentences (up to 10 words). Actually, the
system’s slow performance is the reason for the length limit; larger sentences are also
possible, but their translation generation might take an unreasonably long time.
In the first module, preprocessing, the input sentence is tokenized and transliterated using
the Buckwalter transliteration system [13]. The transliterated tokens are also part-of-
speech tagged using the SVM-POS tool and analyzed using the Buckwalter
morphological analyzer. In the matching module, the system searches for already
translated fragments of the input text, by matching word by word. Words are matched on
string, stem, lemma, and part-of-speech levels, with each level of match assigned a

different score. Exact string match receives the highest score; stem match is less, etc. The
score for a fragment is calculated from the individual word-match scores, as will be
explained later. For the transfer module, we only handle fragments whose score exceeds
some predefined threshold. The first step is the extraction of the translations of the
fragments from the English version of the translation examples that have been found in
the matching module. This is done by aligning the translation example on the word level,
using a bilingual lexicon. Actually, for each Arabic word, we look up its English
equivalent in the lexicon, and expand it with synonyms from WordNet. Then, we search
the English part of the translation example for all instances on the lemma level. Finally,
we take the shortest English segment that contains the maximum number of aligned
words. A new score is produced for each proposed translation using the length of the
fragment, number of translated words and the previous matching score. After extracting a
translation of the fragment, sometimes there is a need to modify it. Recall that the match
of a corpus fragment to the source fragment can be inexact, so sometimes more work
needs to be done to get a real translation. There are several issues that we have handled
and which are presented in the following chapters of this thesis, but there remain some
more unhandled cases, which are outside the scope of this work.
The recombination module simply generates all the best combinations of the translated
fragments considering their total scores (a combination of their matching and translation
scores), but – as already mentioned – it does not smooth out the recombined text. The
final output contains the N-best combinations with their scores, where N is an adjustable
parameter. (For clarity, a separator is placed between each translated fragment.)

1.4 Thesis Overview
In Chapter 2 we give a very short introduction to some basic elements of the Arabic
language. This chapter is primarily for readers without Arabic knowledge, but
intermediate and advanced speakers of Arabic may also profit from this information.
Chapter 3 is a short review of several relevant natural-language-processing tools that we
used in our system implementation. Chapter 4 is an explanation of the corpus creation
process. The translation algorithm is described in Chapter 5 and its results and
evaluations are presented in Chapter 6. The Appendix shows several example inputs and
the corresponding system translations.

Chapter 2
Short Introduction to Arabic

2. Chapter 2 – Short Introduction to Arabic
In this chapter we introduce the relevant basic elements of the Arabic language. A non-
Arabic speaker will find this information useful, but this is far from being a
comprehensive Arabic guide. Further reading (e.g. [14]) is recommended.
Arabic is currently the sixth most widely spoken language in the world. With 22
countries that define Arabic as their official language, it is spoken by at least 250 million
people. Arabic is a Semitic language; it belongs to the same family as Hebrew, Amharic,
Maltese, and many more. It is one of the few languages that exhibit diaglossia [15],
which is the separation between the spoken languages (dialects) and the formal language.
There are Arabic dialects that are more different from each other than even French and
Spanish. However, the formal language is Modern Standard Arabic (MSA), which is used
in written texts and is spoken in formal settings.
Translating Arabic to English is not an easy task for a number of reasons. Arabic
sentences are usually long and contain only few punctuation marks. Due to the
complexity of Arabic syntax, sometimes Arabic sentences are syntactically ambiguous
and require much effort when trying to resolve such ambiguities automatically. Only in
the past several years has there been serious research on Arabic language understanding,
and some basic Arabic natural-language-processing tools have been built. There is still a
definite lack of language resources and tools, which makes it extra challenging to build
working Arabic systems.

2.1 Arabic Orthography
Arabic is written from right to left and its script is also used by other languages, including
Persian, Pashto, Urdu, etc. Arabic, like Hebrew, has a consonantal orthography. Due to
this fact, an Arabic word without vowel diacritics may be analyzed in many ways, and
sometimes it is very hard to resolve the meaning in such situations. We work with texts
that are, for the most part, unvocalized.
Arabic words may be written with tatweel (“elongation”, also called kasheeda), which is
used for highlighting words or simply for text justification. This is done by stretching
some of the letters, so the word will be spread over a bigger space than other words. For
example the word ‫ﺍﻻﻨﺴﺎﻥ‬ (AlAnsAn, “the humans”) may be written as ‫ن‬L‫ــ‬NOP‫ا‬ (note the
elongated third letter from the left).
There are several transliteration standards for Arabic. We use the Buckwalter
transliteration system [13], which was developed at Xerox by Tim Buckwalter in the
1990s. This transliteration system represents Arabic orthography strictly one-to-one:
Every Arabic letter or vocalization mark is assigned a single ASCII English sign.
2.2 Arabic Morphology and Syntax
Like many other Semitic languages, Arabic is highly inflected; words are derived from a
root and pattern, combined with prefixes, suffixes and circumfixes (see next chapter).
The root consists of 3-4 consonants, which are called radicals and the pattern is a
sequence of consonants and variables. Arabic words are created by assigning the root
radicals to the pattern variables. A root contains the seed meaning of the word, but the
pattern may change that meaning. Thus, combinations of the same root with different
patterns may have different meanings. For instance, the combination of the root ‫ﻙ‬
.
‫ﺕ‬
.
‫ﺏ‬

(k.t.b) and the pattern mXXX (here, X is a variable) results in the word ‫ﻤﻜﺘﺏ‬ (mktb,
“office”). Combining the same root with the pattern XXAX, results in the word ‫ﻜﺘﺎﺏ‬
(ktAb, “book”). Verbs are also generated in the same way. Actually, there are 10 different
patterns for verbs, but not all them are valid to combine with every root.
Sometimes, there are prefixes and/or suffixes attached to words. Those affixes may
modify several features of the word, including its number (singular, dual, plural or
collective), its gender (masculine, feminine or no gender), possession, definiteness, case
(nominative, accusative or genitive), tense (past, present or future) and more. Arabic
affixes have the feature of concatenating with each other according to predefined
linguistic rules, which increases the overall number of affixes [16].
Due to the complexity of Arabic morphology, sometimes words may be ambiguous. The
average overall morphological ambiguity in Arabic Penn Tree-Bank is 2.5 analyses per
word [17]. Resolution is usually left for higher levels of analysis.
Modern-standard-Arabic sentences usually stick to the Verb-Subject-Object (VSO)
structure. (English, by way of contrast, is SVO.) Within an Arabic sentence, there are
several required points of agreement between the verbs and the subjects. Adjectives and
nouns mostly need to agree on their gender, number and definiteness. An Arabic noun-
phrase places the noun before the adjectives, which is the opposite of English noun-
phrase syntax.

Chapter 3
Natural Language Processing

3. Chapter 3 - Natural Language Processing
As mentioned before, our translation process requires integration of several natural-
language-processing tools, in both languages. Those tools may be categorized into two
layers – morphology and syntax. This chapter explains briefly how those tools work, and
describes the way we have integrated them in our implementation.
3.1 Morphological Tools
Morphological tools operate on the word level, so the input to those tools is a single
word. First, let us describe some basic important definitions:
Morpheme – This is the smallest meaningful unit in a language. A word may be
composed of several morphemes. A morpheme may be a word by itself, but it can also be
a word affix. For example, the word “connected” is composed of the morphemes connect
and ed. The latter is not a standalone word.
Lemma - A lemma is the canonical form of a lexeme, which is the set of all surface forms
that represent the same word. For example, in English, “goes”, “go”, “went”, “going”,
etc., are all in the same lexeme, whose lemma is to go. A word may be derived from a
number of lemmas; each of them may have a different part-of-speech role in the
sentence/phrase. For instance, in English, the possible lemmas of the word “book” are the
verb to book (as in “to book a flight”) and the noun book. Finding the correct lemma of a
word within an Arabic sentence without diacritics (which is the situation in most of the
time), may be a difficult task. For example, the word ‫ﻭﻝﺩ‬ (wld) may analyzed as 
‫ﺩ‬‫ﹶ‬
‫ﻝ‬
‫ﻭ‬
(walada, “give birth to”) and therefore comes from the lemma walada, but it is also
analyzed as 
‫ﺩ‬‫ﹼ‬
‫ﻝ‬
‫ﻭ‬ (wala~da, “to generate”) and therefore comes from the lemma wala~da.

Both are verbs. It is worth mentioning that there are unvocalized Arabic texts that do
contains the ~ (shada) where needed. There are situations that it is enough to know the
part-of-speech role of a word in order to find its lemma, but sometimes it is simply not
sufficient.
Dictionary entries, in many languages, are usually based on lemmas. A lemma does not
have to be a real word in the language; there are languages in which the lemma is only a
code for the lexeme it is representing. In most English dictionaries, the lemma of a noun
is represented in the singular form, for example, the lemma of “houses” is house. The
lemma of a verb is in the first-person singular present-tense form, for example, the lemma
of “built” is build. Adjectives and adverbs are usually represented by the positive form of
the word (for example, the lemma of “unhappy” is happy), but sometimes, for instance in
WordNet, the lemma is the only existing form of the adjective/adverb (so the lemma of
“unhappy” remains unhappy). All other categories in English, such as prepositions,
conjunctions, interjection, pronoun, etc., are represented by the only existing form. In
Arabic, nouns are represented by the singular form of the noun and the lemma of a verb is
the third-person singular of the past tense.
Lemmatization – By that we means the process of automatically finding the lemma of the
words within a given text. It is also called context-sensitive lemmatization. A lemmatized
text is a text in which every word was replaced by its lemma. Arabic lemmatization tends
to be a difficult task [18] and, currently, there are no known Arabic lemmatizers (at least
for the author) that are available for public use.
Stem – A stem is the base form of a word. It is a combination of a root and derivational
morphemes that may be attached to the beginning of the word (prefixes), at the end

(suffixes) or somewhere in the middle (infixes). Circumfixes are discontinuous
morphemes; for instance, the English word “enlighten” is built from the root light and the
circumfix en___en.
Inflectional morphemes are not part of the stem. For instance, the root of the Arabic word
‫ﺒﻴﻭﺘﻬﻡ‬ (bywthm, “their homes”) is ‫ﺏ‬
.
‫ﻱ‬
.
‫ﺕ‬ and the stem is ‫ﺒﻴﻭﺕ‬ without the inflectional
morpheme ‫ﻫﻡ‬ (hm, “their“).
Given a word in its surface-form, stemming is the process of automatically revealing the
word's stem. In other words, stemming a word is actually the removal of all the
inflectional morphemes from the word's surface-form. Inflectional morphemes in English
words are, for the most the part, attached to the end of words, so stemming an English
word amounts to simply the removal of its inflectional suffixes. In Arabic, it is a much
more complicated task. As mentioned above, Arabic is a highly inflectional language, so
inflectional morphemes may occur as prefixes and/or suffixes.
There are several available English stemmers that produce satisfactory results (e.g., the
Porter stemmer [19]), but finding a satisfactory Arabic stemmer is much more difficult.
One available Arabic stemmer was built by Khoja and Garside [20], and was used by the
information retrieval system of the University of Massachusetts [21]. According to the
latter researchers, although this stemmer made many mistakes, it improved their system’s
performance immensely. Unfortunately, this stemmer is not available for public use.
3.1.1 Morphological Analysis
A morphological analysis technique is a computational process that analyzes natural
language words by considering their internal structures [16]. Given a word in its surface-
form, a morphological analyzer reproduces its morphemes. The output usually contains

information tags for the revealed morphemes, and their content may vary from one
analyzer to another. Basically, an information tag may contain the tense of the word, the
number, person and part-of-speech. For example, given the word “connective”, a
morphological analyzer will reproduce the morphemes connect [Verb] and ive, and the
first morpheme is also the stem of the word. The second morpheme, ive, modifies the first
morpheme, which is a verb, to an adjective.
In many cases, words may be analyzed in different ways and usually morphological
analyzers produce all those possible analyses. For instance, the Arabic word ‫ﻭﻝﺩﻱ‬ (wldy)
can be analyzed in several ways. One way is ‫ﻭﻝﺩ‬ + ‫ﻱ‬ (wld + y, “my son”), where the first
morpheme is the noun ‫ﻭﻝﺩ‬ (wld, “son”) and the second morpheme is a first-person
possession suffix. Another possible way is ‫ﻭ‬ + ‫ﻝﺩ‬
‫ﻯ‬ + ‫ﻱ‬ (w + ldY + y, “and with me”),
where the first morpheme is the conjunction ‫ﻭ‬ (w, “and”), the second morpheme is the
preposition ‫ﻝﺩ‬
‫ﻯ‬ (ldY, “with/by”) and the third morpheme ‫ﻱ‬ (y) is a first-person pronoun.
Arabic, like other Semitic languages, introduces a high proportional rate between a single
word and the number of its possible analyses, so, in most cases, resolving those
ambiguities is left for the higher processing layers, syntactic and even semantic analysis.
Rule-based and corpus-based approaches are the most common paradigms for building a
morphological analyzer. Rule-based morphological analyzers use large number of
predefined rules, which are represented by a finite-state machine. The Buckwalter Arabic
morphological analyzer [13] uses a large lexicon with 82158 entries representing 38600
lemmas (in version 1.0), 299 possible prefixes and 618 possible suffixes. It also uses a
predefined table (3500 entries) that contains morphological rules for the possible
combinations of those stems, prefixes and suffixes. Given an Arabic word, the analyzer

produces all possible morphological analyses of that word. Each analysis is composed of
the stem, lemma, comprised morphemes with their part-of-speech tags and English
glossaries.
Other analyzers for Arabic are the Beesley analyzer [22], as well as the analyzer of Berri,
Zidoum and Atif [23].
Corpus-based analyzers learn how to handle new inputs using a large tagged corpus, in
the case of a supervised analyzer, or a large untagged corpus, in the case of an
unsupervised analyzer. Morpho3, by RDI [24], and its shallow version Sebawai, by
Drweesh [25], are based on a hybrid paradigm. Those analyzers use template rules to find
the root of a new given word, and also a large corpus for resolving analysis ambiguities
by statistical methods. The Sebawai analyzer extracts the template rules automatically
from a table that was created by another morphological analyzer. The table contains
many Arabic words with their corresponding roots.
As mentioned earlier, our implementation uses the Buckwaler analyzer for Arabic
morphological analysis and WordNet 2.0 for lemmatizing English words.
3.2 Syntax Tools
Syntax tools deal with the grammatical structure of a given sentence. The most common
task on this level is (syntax) parsing. Given a sentence and a predefined grammar (in our
case, a natural-language grammar), parsing is the process of finding the grammatical
structure of the sentence, according to the given grammar. The structure is represented by
a grammar tree, which its internal nodes represent the phrases of the sentence and its
leaves represent the words with their part-of-speech [26].

Processing the syntax of a sentence helps in resolving some of the morphological
ambiguities of its words. Unfortunately, it also introduces another level of ambiguity --
syntactic ambiguity, which is, in most cases, harder to resolve.
As mentioned in the previous descriptions of other computational natural-language tools,
parsing an Arabic sentence is much more difficult than parsing an English one. The long
average length of an Arabic sentence and the complexity of Arabic syntax are two of the
reasons for that. Arabic sentences are much more ambiguous than English sentences [27].
Although recently there have been several interesting studies [28] in that field, currently
there is no known available (at least to the author) satisfactory Arabic parser for public
use.
Another important tool is the part-of-speech (POS) tagger. Part-of-speech tagging is the
process of marking sentence words with their part-of-speech. It is easier than parsing,
since it deals only with the words in the sentences and not with phrases. The tags are
taken from a tagset, which is a predefined tags list. Table 1 shows the well-known Penn
TreeBank tagset [29].

Tag Role Tag Role Tag Role
CC Coordinating
conjunction
NNS Noun plural TO To
CD Cardinal
number
NNP Proper Noun
singular
UH Interjection
DT Determiner NNPS Proper Noun
plural
VB Verb, base
form
EX Existential
there
PDT Predeterminer VBD Verb, past
tense
FW Foreign word POS Possessive
ending
VBG Verb,
gerund/present
participle
IN Preposition PRP Personal
pronoun
VBN Verb, past
participle
JJ Adjective PP$ Possessive
pronoun
VBP Verb, non-3s,
present
JJR Comparative
adjective
RB Adverb VBZ Verb, 3s,
present
JJS Superlative
adjective
RBR Comparative
adverb
WDT Wh-determiner
LS List item
marker
RBS Superlative
adverb
WP Wh-pronoun
MD Modal RP Particle WPZ Possessive Wh-
pronoun
NN Noun
singular
SYM Symbol (math,
scientific)
WRB Wh-adverb
Table 1 - The Penn TreeBank project tagset
For example, given the sentence, “Can I book one ticket for tomorrow?”, the following
are the tags of its words, using the Penn TreeBank tagset:
Can/MD I/PRP book/VB one/CD ticket/NN for/IN tomorrow/NN
A specific word may be analyzed with different part-of-speech tags according to its role
in the sentence. The word can for instance, may be tagged with MD (modal) as in the
previous sentence, but also as NN (noun) in other contexts.
Part-of-speech taggers, like other natural-language tools, have been developed based on a
rule-based paradigm or a corpus-based one. Rule-based taggers use a set of rules to
compute the tags of a new given sentence, while corpus-based taggers learn how to tag

new input from a large tagged corpus. Unsupervised taggers learn the necessary
information from an untagged corpus. Hybrid taggers also exist.
Brill’s tagger [30] is a part-of-speech tagger for English with a transformation-based
error-driven learning method. The entire process of the tagger is done in two main steps.
In the first step the tagger assigns the most appropriate tag to each word of the input
sentence and In the second step it uses a small set of rules, which were inferred
automatically from a large corpus, to transform the tag to the correct one. There is also an
Arabic version of Brill’s tagger developed by A. Freeman [31]. Freeman’s tagger uses a
tagset with 146 different tags, taken from the Brown corpus for English [32]. Another
Arabic part-of-speech tagger is APT (Arabic Part-of-Speech Tagger) [33,34], which is
based on a hybrid paradigm. That means it uses statistical computations and rule-based
technique to produce the required tags. The tagset contains 131 tags, taken from the BNC
corpus for English [35] and modified to match Arabic requirements. SVM-POS, which
has recently been published, is another tagger for Arabic [12]. The SVM-POS uses a
modified version of the Penn TreeBank tagset with 24 different tags and handles the
tagging problem as a classification problem. Given a number of features extracted from a
predefined linguistic context (such as number of tokens before the current word, number
of tokens after, part-of-speech tags of the preceding tokens, etc.), the task is to predict the
class of a token. The tagger uses the Support Vector Machine (SVM) to deal with the
classification issue.
Our translation system uses the Brill’s tagger for English and SVM-POS for Arabic.

3.2.1 Shallow Parsing
As mentioned in the previous section, parsing is a very complicated task. Moreover,
Arabic syntax is more ambiguous than English, which makes Arabic parsing even harder.
This is the main reason that many natural-language application developers prefer to use a
limited parser version – a shallow parser (also called a base-phrase chunker). A shallow
parser operates slightly above part-of-speech tagger and far below a full (tree) syntax
parser. Given a sentence with its part-of-speech analysis, the shallow parser finds only its
base-phrases, that is non-recursive and non-overlapping phrases. Each base-phrase
contains one head-word, which is the base of the phrase. For example, the head-word of
the phrase “the table” is table. A short description of English phrase structure can be
found in [36]. A base-phrase may be categorized as a Noun-Phrase (NP), Verb-Phrase
(VB), Adjective-Phrase (ADJP), Adverbial-Phrase (ADVP) or Preposition-Phrase (PP).
For instance, take the sentence: The house of the rising sun.
By part-of-speech tagging the sentence (using the tagset given in Table 1), we get:
The/DT house/NN of/IN the/DT rising/VBG sun/NN.
Shallow parsing the sentence will result in:
[ The/DT house/NN ] of/IN [ the/DT rising/VBG sun/NN ]
which means that The/DT house/NN and the/DT rising/VBG sun/NN are both base
noun-phrases. The word of was left as a standalone preposition. Relating it to the next
noun-phrase would have composed a complex recursive preposition-phrase, which is out
of the scope of the shallow parser task.
There are several available shallow parsers for English. We used the BaseNP parser [37].
BaseNP requires input sentences that were previously tagged by Brill’s tagger. Noun-

phrases are the only base-phrases that BaseNP identifies. Arabic shallow parsers are rare.
The most significant parser was developed by M.Diab, K.Hacioglu and D.Jurafsky [12],
which, like their part-of-speech tagger, handles this task as a classification problem and
uses SVM to solve it. Although it gives some impressive results, it is still far from being
perfect.

Chapter 4
Extracting Translation Examples

4. Chapter 4 – Extracting Translation Examples
In this chapter, we describe how we created our corpus of translation-examples.
4.1 Preparing the Corpus
The translation examples were extracted from a collection of parallel Arabic-English
documents that were taken from the United-Nations document inventory. The Linguistic
Data Consortium (LDC) aligned many parallel-documents form this inventory on the
sentence-level, and organized them as a large corpus. Unfortunately, at this point we
could not afford buying a license to use it or to any other aligned bilingual corpus, so we
decided to automatically align part of the UN parallel-documents and create our own
corpus. The UN inventory is available under the Official-Document-System (ODS) [38]
web interface and it contains official documents since 1993 until these days. The
documents are available for download in six UN official languages including English and
Arabic. We downloaded the documents as Microsoft Word files and took a quick look
over each of them to filter out those documents with complicated format structure
(bullets, tables, etc.), since it is harder to convert those documents to a pure ASCII file.
Parallel-documents that survived the initial filtering were manually preprocessed in the
following steps:
1. Removing all editors/translators notes from the documents.
2. Documents with more than one text column were transformed into a single
column documents.
3. Removing the special Tatweel character (“_”) from the Arabic documents (see
Section 2.1).

The preprocessed parallel-documents were converted to text files. Arabic documents
were converted to the Arabic ASCII “1256” encoding and English documents were
converted to the English ASCII “1252” encoding.
As was mentioned earlier, example-based systems use bilingual corpus, aligned at least
on the sentence or paragraph level. Since, in our case, manual alignment of the
documents is a major time-consuming (and not feasible) task, we have decided to
automatically align those documents on the paragraph level. It is worth to mention that
our work is not about developing an Arabic-English paragraph alignment algorithm so
the results are far from being perfect. We believe that using a better (we wish – manually)
aligned corpus, our translation algorithm will produce better results.
4.2 Parallel Text Alignment
Based on our final goal – creating a translation-example corpus – we could afford using a
non perfect alignment algorithm since we could take only paragraphs that were
successfully aligned and without any problems or suspicions. Paragraphs that are
translated into more than one paragraph in the other language (not a common situation)
were discarded. All parallel-paragraphs were taken as translation-examples and were
forward for further processing.
Given a preprocessed parallel-document, we used a modified version of DK-Vec
algorithm [39] to find as many parallel-text-anchors as possible and use them find match
for each Arabic paragraph. A parallel-text-anchor (also referred as anchor) is the
following quadruplet –
- Arabic word;
- location of the word in the Arabic version of the parallel-document;

English translation of the Arabic word; and
- location of the translation in the English version of the parallel-document.
Figure 4 demonstrates several anchors in a given parallel text. The colored words (also
connected with arrows) are only part of the anchors of the text.
YZ[]‫ا‬
56‫د‬89:‫ا‬
^_`a
<=>?
Ybc‫ا]^ا‬ YN[]
de ‫دة‬hibj]‫ا‬
@AB:‫ا‬
،
‫>رك‬E>FG
،Ybj[]‫ا‬ ‫م‬hm ،
7
‫ول‬p‫ا‬ qm^rs
/
^cht‫أآ‬
2005
YwLN]‫ا‬ ،
00
/
10
Sixth Committee
Summary record of the 4th meeting
Held at Headquarters, New York,
on 7 October 2005, at 10 a.m.
Figure 4 – Parallel-Text-Anchor examples
In the following sub-sections we are describing the entire process of extracting anchors in
a parallel-document and the way we use them to find parallel-paragraphs.
4.2.1 Finding Parallel Text Anchors
The DK-Vec algorithm searches for anchors in a parallel text based on the fact that the
positions of word pairs (Arabic word and its translation in the English version) are
distributed similarly throughout the Arabic and English parallel versions. Although we
have implemented the main steps of the DK-Vec algorithm, and since we are not aligning
the text on the word level, we have also added another condition for filtering out anchors,
so as to keep only the most reliable ones. DK-Vec was formerly used by Choueka, Conley
and Dagan [40] as part of a word-level alignment algorithm for the Hebrew-English pair.
They reported satisfied, although not perfect results.

In their implementation, as well as our version, searching for anchors is performed on the
lemma level. For that reason, in our implementation, each preprocessed parallel-
document was morphologically analyzed, to discover all the possible lemmas of each
word and all possible glossaries for the Arabic words. Arabic and English morphological
analysis is explained in Section 3.1.
Given a lemmatized parallel-document (actually, each word is represented by all its
possible lemmas), the algorithm creates a recency-vector [39] for each lemma. In our
version, a recency-vector of a lemma l contains the distances between all the possible
occurrences of l in the text. To be more specific, the i th position of the vector contains
the distance between the i th and the 1
−
i th possible occurrences of the lemma in the
text. The distance is simply the difference between the word indices, so one fact is that
the first position in every recency-vector contains the index of the first possible
occurrence of the lemma in the text. We refer to those occurrences as "possible" due to
the fact that for each word, we extracted all the possible lemmas of the word and each of
them was considered in the recency-vectors creation, but unfortunately only one lemma is
the correct lemma of the word so the occurrences that were calculated from the other
lemmas are erroneous. A real lemmatization of both text versions, which was left for
future work, would resolve it. Figure 5 shows some recency-vectors in a given English
text.

"…Having taken note of the report of the Global Environment Facility to the Conference of the
Parties,
Noting with appreciation the efforts of the Global Environment Facility to operationalize the
implementation phase of national adaptation programmes of action,
Noting with appreciation the progress made so far in the preparation of national adaptation
programmes of action,
Welcoming financial and technical assistance to support least developed country Parties to
integrate climate change issues into their development processes under Article 4, paragraph 1 (f)
of the Convention…"
Here are some examples for recency-vectors -
Lemma Recency Vector
develop 63, 10…
environment 10, 26…
Figure 5 - English recency-vector examples of a given English text, taken from our corpus
Once the recency-vectors had been captured in both versions of the text, the first step of
DK-Vec is to find a matched English recency-vector for each Arabic one. This step is
called matching. The second step is extracting anchors from each matched recency-
vectors pair.
4.2.2 DK-Vec Matching
In the matching step, the algorithm searches the best match for each Arabic recency-
vector. A best match for an Arabic recency-vector is the most similar English one, in the
sense of length and values. The basic assumption here is that if an Arabic word σ is
translated to an English word γ, their distribution in both texts is expected to be similar.
The first issue is to filter out poor English recency-vector candidates. Candidates that
match any of the following events were discarded:
1. Length event - the ratio between the frequencies of the Arabic lemma and the
candidate English lemma exceeds 2. That means that the size of one of the
recency-vectors is as twice as the other recency-vector. This event is suggested

originally by Fung and McKeown in their paper and also implemented by
Choueka, Conley, and Dagan.
2. Large distance event - the distance between the first occurrences of both lemmas
is equals or greater than half of the entire document text length (measured by
words count). This factor was also suggested by Fung and McKeown in their
paper.
3. Semantic distance event – this event does not comply with the original DK-Vec
approach. In our version, we are only keeping candidates that their lemmas are
related to each other, semantically. Actually, for each candidate, we also check
the semantic relation of the lemmas by lookup the English equivalent of the
Arabic lemma in a bilingual lexicon, expand both lemmas with synonyms from
WordNet and create two corresponding synonyms lists and then check whether
those lists are intersected. The lexicon is represented by the glossary entries
supplied by the Buckwalter morphological analyzer. For example, taking the
recency-vector that represents the lemma ‫ﺴﻼﻡ‬ (slAm, “peace”), lookup its English
equivalent in the lexicon returns the lemma peace. Expanding it with synonyms
form WordNet generates the following list: ataraxias, greet, greeting, heartsease,
military_greeting, pacification, peace_of_mind, peace_treaty, peacefulness,
pledge, present, public_security, recognize, repose, Sallam, salutation, salute,
serenity, toast, wassail. Now, every English vector that contains one of these
lemmas in its corresponding synonyms list, is a possible candidate. Currently, we
deal only with one word lemmas (for instance, in the last example, peace_treaty is
discarded). Applying this event, causes a dramatically reduction of candidates

number, but also improves significantly the reliability of the resulted anchors. As
mentioned above, since do not use DK-Vec for word-level alignment neither for
dictionary generation, we can afford such reduction.
For the remaining candidates of each Arabic recency-vector, we calculated a matching-
score in the same way as described by Fung and McKeown and also by Choueka, Conley,
and Dagan. Let s
v be the recency-vector of an Arabic lemma s
l and let t
v be the recency-
vector of an English lemma t
l . By definition, ]
[i
vs is the distance between the 1
−
i th
and i th occurrences of s
l in the Arabic version and ]
[ j
vt is the distance between the
1
−
j th and the j th occurrences of t
l in the English side. The matching-score should
reflect the similarity of the distances ]
[i
vs , ]
[ j
vt . In other words, the best English vector
candidate should have the smallest values of | ]
[i
vs - ]
[ j
vt |. Given a pair of recency-
vectors, we could find its matching-score by calculating the basic one-to-one Euclidean
distance of the positions of the vectors, but due to the linearity of that metric, sometimes
it compares distances that do not correspond. Fung and McKeown suggested using the
Time-warping method instead. This method solves this problem by recovering the
optimal alignment between the distances. It is called time-warping because it warps the
time axes of the two vectors in such a way that corresponding distances appear at the
same location. Figure 6 shows it graphically.

Figure 6 - Time Warping distance of two vectors
The calculation is done in a dynamic programming fashion, which is called Dynamic-
Time-Warping (DTW). On each iteration, we calculate the value ]
,
[ j
i
C , which
represents a partial matching-score of the candidate; actually ]
,
[ j
i
C is the matching-score
of positions i
,...,
2
,
1 and j
,...,
2
,
1 of s
v and t
v , respectively. Easily we conclude that
0
]
0
,
0
[ =
C since it represents the score of zero matches. For ]
[
0 s
v
size
i ≤
< and
]
[
0 t
v
size
j ≤
< , we compute ]
,
[ j
i
C using the following recursive formula:





−
−
−
−
+
−
=
]
1
,
[
]
,
1
[
]
1
,
1
[
min
]
[
]
[
]
,
[
j
i
C
j
i
C
j
i
C
j
v
i
v
j
i
C t
s
In other words ]
,
[ j
i
C is the minimum cumulative score of the positions i th and j th in
the vectors, among the adjacent positions. By looking on ]
,
[ j
i
C as a cell in a matrix with
s
v and t
v as its axes, the computation is illustrated in Figure 7. We see that each cell in
the path chooses its position according to which of the three options that it has to arrive
from, minimizes the cumulative score.

(i-1, j) (I, j)
(i-1, j-1) (I, j-1)
Figure 7 - Choosing the best option that minimizes the cumulative score in dynamic time warping
algorithm
The calculation is continued bottom-up until reaching ]]
[
],
[
[ t
s v
size
v
size
C which is the
final matching-score of the vectors s
v and t
v . The best match for an Arabic recency-
vector s
v is the English recency-vector t
v that achieves the maximum matching-score
]]
[
],
[
[ t
s v
size
v
size
C among all candidates.
4.2.3 Extracting Anchors
Finding the English recency-vector that matches a given Arabic vector does not mean
that all the positions in both vectors will be matched and taken as anchors. We discarded
positions that were matched by cases 2 or 3 (i.e., ]
,
1
[ j
i
C − or ]
1
,
[ −
j
i
C ). However, first
case positions (i.e., ]
1
,
1
[ −
− j
i
C ), were extracted and marked as anchors.
4.2.4 Extracting Corresponding Paragraphs
The anchors that were extracted in the previous step are used to find the corresponding
English paragraph for each Arabic paragraph. Given a parallel document, we divided the
Arabic and the English versions into paragraphs. Fortunately, this was an easy task since
the conversion of the documents from MS-Word into text places each paragraph in a
single line. In the English version there was a problem only with paragraphs that were
crossed a page transition in the MS-Word version, so we manually resolved such
situations.

For each Arabic paragraph we matched the English paragraph that maximizes the number
of common anchors. In case that more than one English paragraph has the same number
of common anchors with a specific Arabic paragraph, we chose the closer one with
respect to the beginning positions of the paragraphs in the document. A special word ratio
parameter was used to filter unreliable paragraphs pairs. Only paragraphs that the ratio
between the number of Arabic words and the number of English words is in the range 0.6
– 0.9 were used. The range was determined relying on the Language Proportion
Coefficient (LPC) suggested by Choueka, Conley, and Dagan for Hebrew-English pair.
Matched paragraph pairs were morphologically analyzed, part-of-speech tagged and
inserted to the corpus of translation-examples. A special look-up table that maps the
Arabic words to their corresponding English words in each parallel paragraph was also
created.

Chapter 5
Translation Algorithm

5. Chapter 5 - Translation Algorithm
This chapter describes the algorithm we used to develop our automated translation
system. We restricted the input to be a short Arabic sentence, taken from unseen United-
Nations documents published on the ODS system. The entire translation process is
performed in steps. First, the input sentence is preprocessed with various morphological
tasks. Then, in the matching step, the system searches our corpus for already translated
fragments of the input sentence. The third step is transfer, which is the generation of the
translation of each matched fragment and on the last step, recombination, the system
generates the complete English translation, pasting together the translated fragments. In
this chapter we will explain thoroughly each step.
5.1 Preprocessing and Data Preparation
Given an input Arabic sentence, first we analyze it with the Buckwalter morphological
analyzer. The analyzer transliterates and analyzes each word separately, to produces all
the possible morphological analyses of that word. Each analysis contains stem, lemma,
comprised morphemes with their part-of-speech tags and English glossaries. The input
sentence is also tagged using the SVM-POS tool, so we are considering only those
analyses with the correct part-of-speech tag. For instance, if a given word is tagged as
noun by the SVM-POS tool, we are considering only noun analyses of the Buckwalter
analyzer. SVM-POS uses a modified version of the Penn Treebank tagset (see Table 1)
and Buckwalter analyzer uses a different larger one, so we implemented a special
converter module that matches Buckwalter tags to/from the corresponding Penn Treebank
tags. The Buckwalter tagset is larger and more specific since it also tags word’s affixes.

Therefore, for each word, we considered only its main part-of-speech tag, the tag of the
stem.
The results of the preprocessing step are all the relevant analyses of each word organized
as XML stream as can be shown in Figure 8.
<word index="1" word="AlqDA'">
<analysis num="1" lemma="qaDA'_1" gloss="justice|judiciary" POS="DET|NOUN"/>
<analysis num="2" lemma="qaDA'_2" gloss="extermination" POS="DET|NOUN"/>
<analysis num="3" lemma="qaDA'_3" gloss="district|province" POS="DET|NOUN"/>
</word>
Figure 8 – Example for preprocessing results of the word ‫ء‬8TA:‫ا‬
The figure shows the results of a single word ‫ﺍﻝﻘﻀﺎﺀ‬ (AlqDA', <different translations are
given in the figure>) taken from an input sentence.
It is worth mentioning that a similar processing step is performed on each translation-
example that stored in our corpus. In the same way, our translation-examples stored with
the same XML data representing the relevant analyses of their Arabic words. The English
version of each translation-example is also analyzed and the relevant information is
stored exactly in the same format. The tools for the English part-of-speech tagging and
morphological analysis are described in Chapter 3.
5.2 Matching
In the matching step, we search our corpus for translation-examples that their Arabic
version match fragments of the preprocessed input sentence. A matched fragment must
contain at least two input sentence adjacent words. Congruent fragments are possible. For
instance, given the input sentence: ‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ ‫ﺭﺌﻴﺱ‬ ‫ﻤﻥ‬ ‫ﻤﺫﻜﺭﺓ‬ (“a memorandum by the
president of the Security Council”), we summarize the fragments that were found in
Table 2.

Fragment Translation of the fragment
‫ﻤﺫﻜﺭﺓ‬
‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ ‫ﺭﺌﻴﺱ‬ ‫ﻤﻥ‬ by the president of the security council
‫ﻤﻥ‬ ‫ﻤﺫﻜﺭﺓ‬
‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ ‫ﺭﺌﻴﺱ‬ a memorandum by
‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ ‫ﺭﺌﻴﺱ‬ ‫ﻤﻥ‬ ‫ﻤﺫﻜﺭﺓ‬ a memorandum by the president of the security
council
‫ﺭﺌﻴﺱ‬ ‫ﻤﻥ‬ ‫ﻤﺫﻜﺭﺓ‬
‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ a memorandum by the president
‫ﺭﺌﻴ‬ ‫ﻤﻥ‬ ‫ﻤﺫﻜﺭﺓ‬
‫ﻤﺠﻠﺱ‬ ‫ﺱ‬
‫ﺍﻷﻤﻥ‬ a memorandum by the president of council
‫ﺭﺌﻴ‬
‫ﺱ‬
‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ the president of the Security Council
‫ﺭﺌﻴﺱ‬ ‫ﻤﻥ‬ ‫ﻤﺫﻜﺭﺓ‬
‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ the security council
‫ﺭﺌﻴ‬
‫ﺱ‬
‫ﻤﺠﻠﺱ‬
‫ﺍﻷﻤﻥ‬ president of the council
‫ﻤﻥ‬
‫ﺭﺌﻴ‬
‫ﺱ‬
‫ﻤﺠﻠﺱ‬
‫ﺍﻷﻤﻥ‬ by the president of the council
‫ﻤﻥ‬
‫ﺭﺌﻴ‬
‫ﺱ‬
‫ﻤﺠﻠﺱ‬
‫ﺍﻷﻤﻥ‬ by president
Table 2 –Fragments that were found for the sentence: ‫ﺍﻷﻤ‬ ‫ﻤﺠﻠﺱ‬ ‫ﺭﺌﻴﺱ‬ ‫ﻤﻥ‬ ‫ﻤﺫﻜﺭﺓ‬
‫ﻥ‬
The same fragment can be found in more than one translation example, therefore a
special matching-score is given for each pair of fragment and its matched translation
example. The matching-score represents the matching quality of the fragment in the
specific translation example. Fragments are matched word by word so the score for a
fragment is calculated from the individual word matching-scores. Words are matched on
levels, with each level assigned a different score. The levels are text, stem, lemma, and
part-of-speech. Text (exact strings) and stem matches credit the words with the maximum
as possible, lemma match credits less and part-of-speech level credits the fragment with
the minimum amount. Table 3 summarizes the several word matching levels we used.

Match Level Description and Example
Text Match (Surface Form) Exact match of the words, The words are exactly the same.
Stem Match The words match in their stems but not in the surface
form. For example the following words match in their
stems:
1. ‫ﺍﻝﺩﺴ‬
‫ﺘﻭﺭﻴﺔ‬ (Aldstwryp, “the constitutionality“)
2. ‫ﺩﺴﺘﻭﺭﻴﺘﻲ‬ (dstwryty, “my constitutional“)
The stem of both words is ‫ﺩﺴﺘﻭﺭﻱ‬ (dusotuwriy)
Lemma Match Words share a lemma. For instance, the following words
match in their lemmas:
1. ‫ﻤﺎﺭﻕ‬ (mAriq, “apostate”)
2. ‫ﻤﺭﺍﻕ‬ (mur~Aq, “apostates”)
Note that the stems of these words are not equal.
Content Match This level is planned but not yet implemented. The idea is
that, for example, two location names would get a higher
score than two dissimilar proper nouns.
Part-Of-Speech Match The words match only in their part-of-speech tag. For
instance, both words are nouns. Actually, we also require
that both words will have the same tags for their affixes.
For example, if a word is tagged as a noun and it has the
definite article ‫ﺍل‬ (Al, “the”) attached at the beginning of
the word, the matched word must agree on both features –

to be a noun and to have the definite article attached at the
beginning.
Common Word Group Match This level is relevant only for common words and affixes,
taken from a predefined list. These words/affixes were
organized in groups that represent the same meaning.
Clearly, a word/affix may be a member of more than one
group.
Words/affixes that are members of the same group are also
matched on this level. For example the prefix ‫ﺏ‬ (b, “with”,
“by”, “in”) is in the same group of the preposition ‫ﻓﻲ‬ (fy,
“in”).
Table 3 –Match levels for words
5.2.1 Fragment Score Calculation
The matching-score for a fragment is calculated by the individual word matching-scores.
The matching-score for a single word match varies from 0, in case of completely
mismatch, to 1, in case of an exact match. The fragment score is simply the average of
the words matching-scores. Table 4 shows the scores for the previously presented match
levels.
Match Level Score
Text match 1
Stem match 0.9
Lemma match For lemma match we calculate a dynamic

score in the following way:
For every possible lemmas of the first
word, we create pairs with all the possible
lemmas of the second word. A matched
pair is a pair that both lemmas are equal.
The score is the rate between the number of
matched pairs and the total number of
pairs.
Content match 0.8
Part-of-speech match 0.3
Common word group match 1
Table 4 –Matching level scores
Text and stem match have almost the same score for now since currently, we have not
handled the translation modification in such situation. There are of course complicated
situations when both words have the same stems but not the same lemmas, for example,
the words ~t‫آ‬ (katab, “wrote”) and ~t‫آ‬ (kutub, “books”). Such cases are not yet handled
since we have not worked with a context sensitive Arabic lemmatizer and so cannot
derive the correct lemma of an Arabic word. Still, the combination of the Buckwalter
morphological analyzer and the SVM-POS tagger allows us to reduce the number of
possible lemmas for every Arabic word so as to reduce the amount of ambiguity.
Actually, by lemma match we mean that the words were matched on any one of their
possible lemmas. The matching-score in such case is the ratio between the number of
equal lemmas, and the total number of lemma pairs (one of each word). Further

investigation as well as developing and working with a context sensitive Arabic
lemmatizer is needed in order to better handle all such situations.
For instance, let us take again the previous input sentence ‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ ‫ﺭﺌﻴﺱ‬ ‫ﻤﻥ‬ ‫ﻤﺫﻜﺭﺓ‬ (m*krp
mn r<ys mjls AlAmn, “a memorandum by the president of the Security Council”) and
assume that our system founds a match for the pattern ‫ﺍﻷﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ (mjls AlAmn, “the
Security Council”) in a translation example that contains the text:
…‫ﻭﻴﺠﺭﻱ‬ ‫ﺍﻝﻠﺠﻨﺔ‬ ‫ﺃﻋﻀﺎﺀ‬ ‫ﺍﻝﻭﺯﺭﺍﺀ‬ ‫ﻤﺠﻠﺱ‬ ‫ﻭﻴﻌﻴﻥ‬ …
The translation example pattern that matches the source pattern is ‫ﺍﻝﻭﺯﺭﺍﺀ‬ ‫ﻤﺠﻠﺱ‬ (mjls
AlwzrA', “government”). The total fragment score is calculated and presented in Table 5.
Tokens Match Level Score
‫ﻤﺠﻠﺱ‬, ‫ﻤﺠﻠﺱ‬ Text Match 1
‫ﺍﻷﻤﻥ‬, ‫ﺍﻝﻭﺯﺭﺍﺀ‬ Part-Of-Speech Match
(noun + determiner)
0.3
Table 5 –Word level matching scores for the fragment ‫ﺍﻝﻭﺯﺭﺍﺀ‬ ‫ﻤﺠﻠﺱ‬ (mjls AlwzrA', “government”)
So the fragment score is simply the average (1 + 0.3)/ 2 = 0.65
Fragments with a score below some predefined threshold are discarded, since passing
low-score fragments to the next step dramatically increases total running time. Note that a
larger corpus, with the concomitant increase in the number of potential fragments, would
require raising the threshold. Reported results in this paper are based on a corpus
containing 13,500 translation examples with threshold value of 0.5.
5.2.2 Fragments Storing
Fragments are stored in the following structure:
source pattern – fragment’s Arabic text, taken from the input sentence;
example pattern – fragment’s Arabic text, taken from the matched translation example;
example – the matched translation example;

matching score – the matching-score of the fragment to the translation example.
Fragments with the same example pattern are collected and stored in a higher-level
structure called general-fragment. Note that a general-fragment with only one fragment is
also possible.
5.3 Transfer
The input for the transfer step is all the general-fragments that were found in the
matching step, and the output of that step is the translations of those general-fragments.
The translation of a general-fragment is the best generated translation among the
fragments of the general-fragment. Translating a fragment is done by extracting its
translation from the English version of the translation-example and performing some
modifications on it. We can summarize the translation process by the following two
steps:
• Extraction of the translation of the fragment’s example-pattern by aligning the
translation-example on the word level, using a bilingual lexicon.
• Fixing the extracted translation so that it will be the translation of the fragment’s
source-pattern.
5.3.1 Step 1 – Translation Extraction
The first step is to extract the translation of the fragment’s example-pattern from the
English version of the translation-example. Actually, as mentioned above, for each
Arabic word in the pattern, we look up its English equivalents in the lexicon and expand
those with synonyms from WordNet. Then we search the English version of the
translation example for all instances on the lemma level. Finally, we take the shortest
English segment that contains the maximum number of aligned words. Since finding all

instances of every Arabic word in the English version of the translation example is a huge
time consumer task, we decided to perform it offline and store the results also in the
corpus. Usually a word in some Arabic example pattern has several English equivalents,
which makes the translation extraction process complicated and error prone. For this
reason, we also restrict the ratio between the number of Arabic words in the example
pattern and the number of English words in the extracted translation, bounded them by a
function of the ratio between the total number of words in the Arabic and English
versions of the translation example. Figure 9 demonstrates the translation extraction
process. The following sections explain each step thoroughly.
Arabic version of the
translation example
dZit]‫ا‬ ‫ون‬Lbt]‫وا‬ Ym‫ر‬LrtP‫ا‬ ‫ت‬Laƒ„]‫ا‬
‫ق‬hi† ‫‡ƒان‬a de
‫ن‬LNOˆ‫ا‬
English version of the
translation example
ADVISORY SERVICES AND TECHNICAL
COOPERATION IN THE FIELD OF HUMAN RIGHTS
Alignment table
(created offline using
the Buckwalter
glossary entries and
WordNet)
SERVICES ‫ت‬Laƒ„]‫ا‬
ADVISORY Ym‫ر‬LrtP‫ا‬
COOPERATION ‫ون‬Lbt]‫وا‬
TECHNICAL dZit]‫ا‬
IN de
FIELD ‫‡ƒان‬a
RIGHTS ‫ق‬hi†
HUMAN ‫ن‬LNOˆ‫ا‬
example pattern ‫ن‬LNOˆ‫ا‬ ‫ق‬hi† ‫‡ƒان‬a
Translation extraction First, we mark the alignments of the relevant example

process pattern words, using the alignment table:
ADVISORY SERVICES AND TECHNICAL COOPERATION
IN THE FIELD OF HUMAN RIGHTS
Now, we take the shortest segment with maximum
number of word alignments:
ADVISORY SERVICES AND TECHNICAL COOPERATION
IN THE FIELD OF HUMAN RIGHTS
At this point of the research, we check if a determiner
exists one place before the beginning of the extracted
segment. If the answer on this question is true, we add
it to the beginning of the extracted segment. Later, as
future work, we should analyze the Arabic pattern to
find if it requires the determiner. The same approach is
taken for some prepositions from a predefined list.
Finally we extract the translation THE FIELD OF
HUMAN RIGHTS.
Figure 9 –Translation extraction demonstration
This is of course only a simple demonstration. As mentioned, more complicated one
would contain more than one alignment for each Arabic word, as will be explain shortly.

5.3.1.1 First Task – Finding English Equivalents
This task is performed offline, at the corpus creation time. Each translation example is
stored together with a look-up table that maps between every Arabic word in the
translation example and all its English equivalents in the English version of the example.
An English equivalent may be a word or a phrase that is a possible translation of the
Arabic word. An Arabic word may have more than one English equivalent in the English
version of the translation example. We use the extracted relevant Buckwalter glossaries
for each Arabic word and expand these with synonyms from WordNet. Actually, a
glossary entry might be composed of more than one word, so we find the synonyms for
each one of them, except of common words identified from a predefined list.
Now, after creating the synonym lists, next we search the English version of the
translation example for all instances on the lemma level. Of course the English
equivalents may be found more than once in the English version of the translation
example, that is the reason for taking the shortest English segment with maximum
number of word alignments as will be explain in the following step.
There are several complicated situations in finding the English equivalents. We are now
describing some of them:
1. The Arabic word is part of a phrase that is translated to a single word in the
English version. Such situations are partly handled by capturing some basic
Arabic language structures. For instance, let us take the Arabic example pattern
‫ﺭﺴﻤﻲ‬ ‫ﻏﻴﺭ‬ (gyr rsmy, “not formal”), which -- in many cases -- we might find it as
“informal” rather than “not formal”. Neither the synonym list of the word ‫ﺭﺴﻤﻲ‬
(rsmy, “formal”) nor the list of the word ‫ﻏﻴﺭ‬ (gyr, “not”) contain the word

“informal“1
. Handling such situation is done by a manually defined rule that is
triggered whenever the system encounter the word ‫ﻏﻴﺭ‬ (gyr, “not”). The system
checks the following word, and -- instead of building a synonym list -- builds an
antonym list, using WordNet. In this example, the word “informal” appear as an
antonym of the word “formal” in WordNet. There are more complicated
structures, which are not handled yet, but capturing and writing rules for such
situations seems quite feasible.
2. The translation of an Arabic word is expressed as a short phrase in the English
version, for example verb and additional particle, as in the pattern ‫ﺍﻝﺘﻌﺎﻭﻥ‬ ‫ﻴﺘﺨﺫ‬ (ytx*
AltEAwn), which was translated in one of our translation examples to “the
cooperation should take on”. We see that the word ‫ﻴﺘﺨﺫ‬ (ytx*, “take”) is translated
to “take on”. This is partially solved using the large number of English phrases
and particles list for every verb contained by WordNet. We defined a window
around each extracted English equivalent in the translation example. The left
boundary of the window is placed before the location of the equivalent and the
right boundary is placed after it. The size of the window is the total words taken
from the left border to the right border, ignoring the equivalent text itself, in a
way that the number of words on both sides is equal. For example, consider the
sentence: “promoting participation of developing countries in new and dynamic
sectors of world trade“. Assume that the extracted equivalent text is the word
“countries”. A six-word window is the text: “participation of developing
countries in new and”. We check the existence of English phrases that contain the
1
Recall that we search the English version on the lemma level, and since we lemmatize the English text using WordNet,
the lemma of the word “informal” is also “informal”.

word “countries” inside the window. First we check the phrase that is composed
of the word “countries” concatenated with the next word. Then, we concatenate
the next word and check again and so on until reaching the right boundary of the
window. The same process is done backward by checking “countries” with the
previous word and so on until reaching the left boundary of the window.
Eventually, we extract the larger phrase that was found. Large windows are better
for dealing with long phrases but also increase running time dramatically.
Empirically, we are working with four-word windows.
3. There are situations were an Arabic phrase is translated to an English phrase, but
the individual translation of each Arabic word is different. For instance, the
phrase ‫ﺍﻻﺴﺘﻌﻤﺎﺭ‬ ‫ﺇﻨﻬﺎﺀ‬ (<nhA' AlAstEmAr) usually is translated to “decolonization”
and not to the individual translation of each word: ‫ﺇﻨﻬﺎﺀ‬ (<nhA', “termination”) and
‫ﺍﻻﺴﺘﻌﻤﺎﺭ‬ (AlAstEmAr, “the colonialism”). In addition, there are phrases that are
always translated to the same English phrase, or maybe to a finite set of phrases,
for example, the phrase ‫ﺃﺠل‬ ‫ﻤﻥ‬ (mn >jl) may translated to “in order to” or to
“because of”. Further investigation is needed for handling such cases. For
instance, the latter cases may be handled using a look-up table that map Arabic
phrases to their corresponding English phrases.
4. As mentioned earlier, sometimes we have problems with determiners or other
functional words that should be placed at the beginning of the extracted phrase.
This happens especially in noun-phrases. For example, consider the pattern from
Figure 10, ‫ﺍﻹﻨﺴﺎﻥ‬ ‫ﺤﻘﻭﻕ‬ ‫ﻤﻴﺩﺍﻥ‬ (“the field of human rights”). The determiner ‫ﺍل‬ (Al,
“the”) is on the word ‫ﺍﻹﻨﺴﺎﻥ‬ (AlAnsAn, “the human”) but in English the determiner

is on “the field“. Currently, when we find a text with a preceding determiner, we
always extract the text with it. Unfortunately, other functional words are not
handled yet. In the future, we are planning to analyze Arabic noun-phrases to find
whether they determined (or related by other functional words) and extract the
text respectively.
5.3.1.2 Second Task – Fragment Alignment
The second task is to extract the translation of the fragment using the equivalent texts we
have just found. Basically, this is done in the following two steps:
1. Extraction of all possible sequences using the corresponding English equivalents.
A sequence is an English segment, extracted from the English version of the
translation example and contains (all/part of) the equivalents of the Arabic pattern
words.
2. Choosing the shortest English segment that contains maximum number of
equivalents, as the translation of the Arabic pattern.
5.3.2 Step 2 – Fixing the Translation
After extracting the translation of the pattern, sometimes there is a need to modify it.
Remember that the match of an example pattern to the input pattern is not always exact,
so sometimes more work needs doing, to get a real translation. In this section we will
discuss several modification issues.
5.3.2.1 Dealing with Words that were Matched on the Part-Of-Speech Level
Recall that the match of a corpus fragment to the input fragment can be inexact: words
may be matched on several levels. Exactly matched words are assumed to have the same
translation, but stem or lemma matched words may require modifications (mostly

inflection and prepositions issues) to the extracted translation. These issues were left for
future work. Words matched on the part-of-speech level require complete change of
meaning. For example, in Figure 10, we take the input fragment ‫ﺍﻻﻤﻥ‬ ‫ﻤﺠﻠﺱ‬ (mjls AlAmn,
“the Security Council”), matched to the fragment ‫ﺍﻻﻤﻥ‬ ‫ﻤﺴﺅﻭﻝﻴﺔ‬ (ms&wlyp AlAmn, “the
security responsibility”) in some translation example. The words ‫ﻤﺠﻠﺱ‬ (mjls, “council”)
and ‫ﻤﺴﺅﻭﻝﻴﺔ‬ (ms&wlya, “responsibility”) are matched on the part-of-speech level (both are
nouns). Assume that the extracted translation from the translation example is “the
security responsibility”, which is actually a translation of ‫ﺍﻻﻤﻥ‬ ‫ﻤﺴﺅﻭﻝﻴﺔ‬ (ms&wlyp AlAmn,
“the security responsibility”) and is not the translation of the input pattern at all. But, by
replacing the word “responsibility” from the translation example with the translation of
‫ﻤﺠﻠﺱ‬ (mjls, “council”) from the lexicon, we get the correct phrase: “the security council”.
The lexicon is imitated using the glossaries extracted from the Buckwalter morphological
analyzer and expanded with WordNet synonyms as was explained above.
Figure 10 –– Fixing part-of-speech level matches

The lexicon is imitated using the glossaries extracted from the Buckwalter morphological
analyzer and expanded with synonyms from WordNet as was explained in Section
5.3.1.1.
5.3.2.2 Dealing With Unnecessary Words in the Middle
Sometimes the extracted translation contains extra unnecessary words in the middle.
Those words appear mostly because of the different syntax of a noun-phrase in both
languages. For example, given the translation-example - ‫ﺍﻻﻗﻠﻴﻤﻲ‬ ‫ﺍﻻﻤﻥ‬ ‫ﻤﻭﻀﻭﻉ‬ (mwDwE
AlAmn AlAqlymy) and its translation: “the subject of regional security”, and extracting the
translation of the pattern ‫ﺍﻻﻤﻥ‬ ‫ﻤﻭﻀﻭﻉ‬ (mwDwE AlAmn) we obtain: “the subject of
regional security” (since it is the shortest segment that contains maximum word
alignments). Clearly, the word “regional” is unnecessary in the translation because it is
the translation of the word ‫ﺍﻻﻗﻠﻴﻤﻲ‬ (AlAqlymy, “the regional”), which does not exist in the
pattern. So by removing the word from the translation we are getting the correct
translation of the pattern. The word “regional” appears in the extracted translation due to
the fact that Arabic adjectives come after the nouns, which is the opposite of English
syntax. The noun-phrase ‫ﺍﻻﻗﻠﻴﻤﻲ‬ ‫ﺍﻻﻤﻥ‬ (AlAmn AlAqlymy, “the regional security”) is
translated so that the translation of the word ‫ﺍﻻﻗﻠﻴﻤﻲ‬ (AlAqlymy, “regional”) appears first
and then the translation of the word ‫ﺍﻻﻤﻥ‬ (AlAmn, “security“). Identifying such situation is
done by searching the translation of the word “regional“ in the Arabic text that comes
immediately after the pattern in the translation example. Currently we search only three
words after the pattern because of time-complexity reasons, but basically in the future we
will further investigate the use of an Arabic chunker to find the boundaries of that noun-
phrase and search the word within that phrase.

Removing the unnecessary words from the extracted translation must preserve the correct
English syntax of the remaining translation, which in some cases seems to be a difficult
task. For that purpose, we compiled several rules to deal with different situations. The
rules are based on the syntax structure of the English extracted translation to identify
cases that need special care. First, we chunk the translation to discover its basic noun-
phrases, using the BaseNP [37] chunker. To do that, we first apply Brill’s part-of-speech
tagger [41] to the translation. By looking on the chunked English text, we check the
effects of removing the unnecessary word. In the previous example, removing the word
“regional” from the English text: “the subject of regional security” may be done without
any further modifications, since by tagging and chunking the segment, we get (phrases in
brackets are noun-phrases):
[ the/DT subject/NN ] of/IN [ regional/JJ security/NN],
and “regional” is simply an adjective within a noun-phrase which still has the same head.
Prepositions and other function-words (in this case the preposition “of“) that relate to the
phrase are still necessary, so we keep them.
Consider another example: ‫ﺍﻝﻤﻨﺎﻅﻤﺔ‬ ‫ﻋﻥ‬ ‫ﺍﻝﻤﺭﺍﻗﺒﺔ‬ (AlmrAqbp En AlmnAZmp, “the observation
on the organization”) with the following extracted translation: “observer for the world
tourism organization”. The words “world” and “tourism” should be removed because
they are not a translation of any of the Arabic pattern words. So again, by chunking the
text we get:
[ Observer/NNP ] for/IN [ the/DT World/NNP Tourism/NNP Organization/NN],
both words “world“ and “tourism“ are noun-modifiers within a noun-phrase. Since we
work with base (non-recursive) noun-phrases, which their head word is located last in the

phrase, we identify that the noun-phrase still exists after the removal of the unnecessary
words so we keep the definite article “the” and the preposition “for”. By removing both
words we almost get the correct translation: “observer for the organization”. The pattern
“observer for“ is not correct, but this happened as a result of a lemma-level matching of
the word ‫ﺍﻝﻤﺭﺍﻗﺒﺔ‬ (AlmrAqbp, “the observation”) and the word ‫ﻤﺭﺍﻕ‬ (mrAqb, “observer”).
There are situations when there is a need to remove the head of the noun-phrase. For
instance, consider the fragment: ‫ﺍﻝﺩﻭﻝﻲ‬ ‫ﻭﺍﻝﻘﺎﻨﻭﻥ‬ ‫ﺍﻻﻨﺴﺎﻥ‬ (AlAnsAn wAlqAnwn Aldwly, “the
humans and the international law“). Suppose that the extracted translation is: “to human
rights and international law”. Note that the word “rights” is unnecessary here so by
chunking the translation we get:
to/TO [ human/JJ rights,/NN ] and/CC [ international/JJ law/NN].
As we promised, the word “rights” is the head of a noun-phrase. That situation happens
because of the fact that the word “human” is an adjective in the extracted translation, but
it is also a noun in the required translation. Although the word “human“ should be left as
is, currently no work has been done to find out what to do with prepositions and other
function words that related that noun-phrase. This and more complicated situations are
left for future work.
5.3.3 Choosing a General Fragment’s Translation
As mentioned above, a general-fragment may contain several fragments sharing the same
Arabic example pattern. Among the extracted translations of the fragments, which are all
translations of the same Arabic pattern, we choose the translation that covers the
maximum number of Arabic words to represent the general-fragment. A translation-
score is calculated for the chosen translation, which is actually the ratio between the

number of covered words and the total number of words in the Arabic pattern. For
instance, the translation-score of a fragment with the Arabic pattern ‫ﺍﻝﺨﺎﺹ‬ ‫ﺍﻻﺠﺘﻤﺎﻉ‬ ‫ﻭﺨﻼل‬
(wxlAl AlAjtmAE AlxAS, “and during the special meeting”) and a translation: “and during
the meeting“ is 0.666, since the text “and during” covers the word ‫ﻭﺨﻼل‬ (wxlAl), the text
“the meeting” covers the word ‫ﺍﻻﺠﺘﻤﺎﻉ‬ (AlAjtmAE), but the word ‫ﺍﻝﺨﺎﺹ‬ (AlxAS) left
without any matched English translation, so the score is the ratio between the number of
covered words, which is 2, and the total number of pattern words, which is 3.
5.4 Recombination
In the recombination task, we paste together the extracted translations to form a complete
translation of the input sentence. The recombination task is generally composed of two
subtasks. The first is finding the N-best combinations of the extracted translations that
cover the entire input sentence and the second is smoothing out the recombined
translations into a fully grammatical English sentence. Currently we handle only the first
subtask, while the second one left for future work.
5.4.1 The Recombination Algorithm
The input for the recombination step is all the extracted translations of the general-
fragments from previous steps. Remember that a general-fragment actually covers a
specific fragment of the input sentence, but there may be more then one general-fragment
that covers the same one, maybe even with different lengths. The main goal of the
recombination algorithm is to choose the best sequence of general-fragments that
together combine a full cover of the input sentence. Our recombination algorithm actually
provides the N best full cover sequences.

The algorithm uses the total-score of the general-fragments to recombine the required
sequences. The total-score of a general-fragment is calculated by multiplying the
previously introduced matching-score of the general-fragment (see Section 5.2.1) and the
translation-score (see Section 5.3.3) of the chosen fragment’s translation.
The main strategy of our recombination algorithm is greedy: choosing the best general-
fragment for every position in the recombined sentence. In other words, we have a
number of general-fragments, were each one may have a different length and starting-
index (the position of the first covered word within the input sentence). It starts with
index i=0, which is the beginning of the input sentence. Now it takes the higher ranked
general-fragment among all the general-fragments that start at the same index and also
have the same length (in their Arabic pattern they cover). Then it searches the next best
general-fragment among all general-fragments that start exactly where the last one ends
or is congruent with a single word (the last covered word of the previous general-
fragment is the first covered word of the next general-fragment), and have the same
length. This process is last until it covers the entire input sentence. Actually, on each step
of the algorithm, it chooses several best general-fragments, each one among a different
length group.
For example, consider the following input sentence: ‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ (wxlAl
Alftrp Alm$mwlp bAltqryr, “and during the reporting period”). Table 6 presents all the
general-fragments that were found for the pattern ‫ﻭﺨﻼل‬
‫ﺍﻝﻔﺘﺭﺓ‬ (wxlAl Alftrp, “and during
the period“):
Matched Arabic Pattern Extracted Translation Total-Score
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﻓﻲ‬ Time 0.325
‫ﻓﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ During 0.5

‫ﺍﻝﻤﻨﺎﻗﺸﺔ‬ ‫ﻭﺨﻼل‬ During 0.325
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ During the period 1.0
‫ﻓﺘﺭﺍﺕ‬ ‫ﻭﺨﻼل‬ The time 0.5
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺃﺜﻨﺎﺀ‬ Period and during 0.65
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ And throughout the
period
1.0
Table 6 –– All general-fragments that were found for the pattern ‫ﻭ‬
‫`_ل‬
‫@ة‬bc:‫ا‬ (waxlAl Alftrp, “and
during the period”)
Among all of those general-fragments, which are all of the same length, the
recombination algorithm arbitrarily chooses one2
of the higher ranked ones ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬
(xlAl Alftrp, “and throughout the period”). Now, on the next steps it keeps choosing the
best general-fragments among those that start either with index 1 or index 2. Table 7
summarizes those steps and presents all the chosen general-fragments from all length
groups.
Start
Index
General Fragment
Cover
General Fragment
Translation
General
Fragment
Score
0 ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ and throughout the period 1.0
0 ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬
and during the reporting
period
0.75
0 ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ during the past phase 0.533
1 ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ included a period 0.65
1 ‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ the reporting period 0.66
2 ‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ report 0.5
Table 7 –– The chosen general-fragments from all collections that were found for the pattern
‫ﻭﺨﻼل‬
‫ﺍﻝﻔﺘﺭﺓ‬ (waxlAl Alftrp, “and during the period”)
Clearly, the first row is the general-fragment that was chosen in the first step from those
of length 2 and start-index 0, and the second row presents another chosen general-
fragment with the same start-index but from a different length group.
2
Another option is to iterate separately for each one of the higher ranked general-fragments, but it seems to cause a
major increase in running-time. Anyway, further investigation of the entire recombination step was left for future work.

Among the chosen general-fragments, the recombination algorithm generates all the
possible sequences. Each sequence is called translation. The translations are ranked using
a final-translation-score that is calculated by multiplying the total-score of the comprised
general-fragments. Table 8 shows the five best translations of the previous example. The
character “|” marks the beginning positions of a new general-fragment.
Generated Sequence (Translation) Translation
Score
and during the reporting period 0.75
and throughout the period | the reporting period 0.66
and throughout the period | report 0.5
and throughout the period | included a period |
report
0.325
during the past phase | report 0.266
Table 8 –– Five best translations and translation-scores for the input sentence:
@E@Ab:8d 5:>BeB:‫ا‬ ‫@ة‬bc:‫ا‬ ‫و`_ل‬ (wxlAl Alftrp Alm$mwlp bAltqryr, “and during the reporting period”),
created by our system
The best among the N best translations is the one with the highest rank. The final-
translation-score is a multiplication of the comprised general-fragments total-scores since
it reduces the score of translations that are composed of several general-fragments. Recall
that for now, we are not smoothing out the translations, so their quality reduces as the
number of comprising general-fragments increases.
5.5 Example
In this section we summarize the different steps in producing N best translations for the
sentence ‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ (wxlAl Alftrp Alm$mwlp bAltqryr, “and during the
reporting period“). As you realized, this may be a complicated task, so we present only
the important relevant steps.
Input sentence: ‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬
Translation: “and during the reporting period”

5.5.1 Preprocessing and Data Preparation
In this step, the sentence is analyzed by several tools, as was explained above, producing
the following information on each word:
<word index="0" word="wxlAl">
<analysis num="1" lemma="xilAla_1" gloss="during|through" POS="CONJ|PREP"/>
</word>
<word index="1" word="Alftrp">
<analysis num="1" lemma="fatorap_1" gloss="phase/time period/interval"
POS="DET|NOUN|NSUFF_FEM_SG"/>
</word>
<word index="2" word="Alm$mwlp">
<analysis num="1" lemma="ma$omuwl_1" gloss="included/contained/implied"
POS="DET|ADJ|NSUFF_FEM_SG"/>
</word>
<word index="3" word="bAltqryr">
<analysis num="1" lemma="taqoriyr_1" gloss="decision/determination"
POS="PREP|DET|NOUN"/>
<analysis num="2" lemma="taqoriyr_2" gloss="report/account"
POS="PREP|DET|NOUN"/>
</word>
5.5.2 Matching
In the matching step, the system searches for all the relevant Arabic fragments in the
corpus. The matching is done in levels (exact, lemma, stem and part-of-speech) as was
explained above. Table 9 shows some of the fragments that were found by the system for
the given input sentence. The left column contains the input sentence pattern; the middle
column contains the matched translation example patterns and the right column provides
their matching-score.
Source Pattern Example Pattern Match Score
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻤﺴﺎﺌل‬ 0.65
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻤﻨﻁﻘﺔ‬ 0.65
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬
‫ﺍﻝﺯﻤﻨﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65

‫ﺍﻝﺘﺤﻀﻴﺭﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﻼﺤﻘﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﻤﻘﺒﻠﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﺤﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 1.0
‫ﺍﻝﻤﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﻤﻤﺘﺩﺓ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﻤﺫﻫﻠﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻻﻨﺘﺨﺎﺒﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻻﻨﺘﻘﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﺘﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﺴﺎﺒﻘﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻻﺴﺘﻌﻤﺎﺭﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻷﺨﻴﺭﺓ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﺤﺭﺠﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺍﻝﻘﺼﻴﺭﺓ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.65
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﻓﺎﻝﻔﺘﺭﺓ‬ 1.0
‫ﺍﻝﻔﺘﺭﺓ‬
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ 1.0
‫ﺒﺎﻻﺴﺘﻌﺭﺍﺽ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.76
‫ﺒﺎﻻﺘﻔﺎﻗﻴﺔ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻷﻁﻔﺎل‬ 0.53
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﻝﻠﻔﺘﺭﺓ‬ 1.0
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﻓﻲ‬ 0.65
‫ﻓﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ 1.0
‫ﺍﻝﻤﺅﺘﻤﺭ‬ ‫ﻭﺨﻼل‬ 0.65
‫ﺍﻝﻤﻨﺎﻗﺸﺔ‬ ‫ﻭﺨﻼل‬ 0.65
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ 1.0
‫ﻓﺘﺭﺍﺕ‬ ‫ﻭﺨﻼل‬ 1.0
‫ﻭﺃﺜﻨﺎ‬
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺀ‬ 0.65
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ 1.0
‫ﺒﺎﻻﺘﻔﺎﻗﻴﺔ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ 0.65
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ 1.0
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬
‫ﺒﺎﻻﺴﺘﻌﺭﺍﺽ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ 0.65
‫ﻝﻠﺩﻭﺭﺓ‬ ‫ﺍﻷﻭل‬ ‫ﺍﻝﺠﺯﺀ‬ ‫ﻭﺨﻼل‬ 0.475
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﻓﻲ‬ 0.825
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﻓﻲ‬
‫ﺒﺎﻻﺴﺘﻌ‬
‫ﺭﺍﺽ‬
0.65
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﻁﻭﺍل‬
0.825

‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻓﺨﻼل‬ 1.0
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ 1.0
‫ﺍﻝﺭﺍﺒﻌﺔ‬ ‫ﺍﻝﺠﻭﻝﺔ‬ ‫ﻭﺨﻼل‬ 0.53
‫ﺍﻝﻤﺎﻀﻲ‬ ‫ﺍﻝﻌﺎﻡ‬ ‫ﻭﺨﻼل‬ 0.53
‫ﺍﻝﻌﺩﻴﺩﺓ‬ ‫ﺍﻷﻋﻭﺍﻡ‬ ‫ﻭﺨﻼل‬ 0.53
‫ﺍﻝﻘﻠﻴﻠﺔ‬ ‫ﺍﻷﺸﻬﺭ‬ ‫ﻭﺨﻼل‬ 0.53
‫ﻭ‬
‫ﺍﻝﺨﺎﺹ‬ ‫ﺍﻻﺠﺘﻤﺎﻉ‬ ‫ﺨﻼل‬ 0.53
‫ﺍﻝﻤﺎﻀﻲ‬ ‫ﺍﻝﺼﻴﻑ‬ ‫ﻭﺨﻼل‬ 0.53
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ 1.0
‫ﺍﻝﺘﺤﻀﻴﺭﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ 0.76
‫ﺍﻝﻤﻤﺘﺩﺓ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ 0.76
‫ﺍﻻﻨﺘﻘﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ 0.76
‫ﺍﻝﺘﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ 0.76
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ 1.0
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬
‫ﺍﻷﺨﻴﺭﺓ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ 0.76
Table 9 – Fragments that were found by the system for the input sentence ‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬
(wxlAl Alftrp Alm$mwlp bAltqryr, “and during the reporting period”)
Some of the patterns were found in more than one translation example. Those were
collected together to a general-fragment.
5.5.3 Transfer
In the transfer step we find the translation for each general-fragment. The translation is
created by first extracting the translation of the pattern from the matched translation
examples and then fixing the translation so it will actually be a translation of the input
pattern. For instance, let us take the general-fragment with the example-pattern ‫ﺍﻝﻔﺘﺭﺓ‬
‫ﺍﻻﻨﺘﺨﺎﺒﻴﺔ‬ (Alftrp AlAntxAbyp, “the electoral period”) that was matched to the input pattern
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ (Alftrp Alm$mwlp, “the included period”). First by extracting its translation
from the translation examples we get: “electoral period”. The second step is to fix the
translation so that it will be the translation of the input pattern. We notice that the word
‫ﺍﻻﻨﺘﺨﺎﺒﻴﺔ‬ (AlAntxAbyp, “electoral”) from the example pattern was matched on the part-of-

speech level to the word ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ (Alm$mwlp, “included”) so its translation should be
replaced with the translation of the word ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ (Alm$mwlp, “included”) using the
Buckwalter analyzer as explained above. As a result, the final translation of the general-
fragment is: “included period”. Table 10 summarizes the translations of all the general-
fragments that were found in the previous matching step.
Input
Pattern
Example-
Pattern
Translation Translation-
Score
Total-
Score
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻤﺴﺎﺌل‬ 0.0 0.0
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻤﻨﻁﻘﺔ‬ included in the area 0.8 0.52
‫ﺍﻝﺯﻤﻨﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ included 1.0 0.65
‫ﺍﻝﺘﺤﻀﻴﺭﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ 0.0 0.0
‫ﺍﻝﻼﺤﻘﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ with the phase out 0.5 0.325
‫ﺍﻝﻤﻘﺒﻠﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ in the period to come 1.0 0.65
‫ﺍﻝﺤﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ to formulate 0.5 0.325
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ carrying 0.5 0.5
‫ﺍﻝﻔﺘﺭ‬
‫ﺍﻝﻤﺎﻝﻴﺔ‬ ‫ﺓ‬ 0.0 0.0
‫ﺍﻝﻤﻤﺘﺩﺓ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ the period 0.5 0.325
‫ﺍﻝﻤﺫﻫﻠﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ included and
terrifying period
1.0 0.65
‫ﺍﻻﻨﺘﺨﺎﺒﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ included period 1.0 0.65
‫ﺍﻻﻨﺘﻘﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ the transitional
period
1.0 0.65
‫ﺍﻝﺘﺎﻝﻴﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ period of 12 months
included
1.0 0.65
‫ﺍﻝﺴﺎﺒﻘﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ period 0.5 0.325
‫ﺍﻻﺴﺘﻌﻤﺎﺭﻴﺔ‬
period 0.5 0.325
‫ﺍﻷﺨﻴﺭﺓ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ included 0.5 0.325
‫ﺍﻝﺤﺭﺠﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ included period 1.0 0.65
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬
‫ﺍﻝﻘﺼﻴﺭﺓ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ included a period 1.0 0.65
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﻓﺎﻝﻔﺘﺭﺓ‬
report 0.33 0.33
the reporting period 0.66 0.66
‫ﺒﺎﻻﺴﺘﻌﺭﺍﺽ‬
the period under
decision
0.66 0.51
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻷﻁﻔﺎل‬ 0.0 0.0

‫ﺒﺎﻻﺘﻔﺎﻗﻴﺔ‬
‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﻝﻠﻔﺘﺭﺓ‬
report 0.33 0.33
‫ﺍﻝﻔﺘ‬ ‫ﻭﻓﻲ‬
‫ﺭﺓ‬ time 0.5 0.325
‫ﻓﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ during 0.5 0.5
‫ﺍﻝﻤﺅﺘﻤﺭ‬ ‫ﻭﺨﻼل‬ 0.0 0.0
‫ﺍﻝﻤﻨﺎﻗﺸﺔ‬ ‫ﻭﺨﻼل‬ during 0.5 0.325
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ during the period 1.0 1.0
‫ﻓﺘﺭﺍﺕ‬ ‫ﻭﺨﻼل‬ the time 0.5 0.5
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺃﺜﻨﺎﺀ‬ period and during 1.0 0.65
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬ and throughout the
period
1.0 1.0
‫ﺒﺎﻻﺘﻔﺎﻗﻴﺔ‬
0.0 0.0
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ report 0.5 0.5
decision 0.5 0.325
‫ﺍﻝﺠﺯﺀ‬ ‫ﻭﺨﻼل‬
‫ﻝﻠﺩﻭﺭﺓ‬ ‫ﺍﻷﻭل‬
0.0 0.0
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﻓﻲ‬
the reporting period 0.5 0.4125
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﻓﻲ‬
the period under
decision
0.5 0.325
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﻁﻭﺍل‬
0.0 0.0
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻓﺨﻼل‬
0.0 0.0
‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﺨﻼل‬
and during the
reporting period
0.75 0.75
‫ﺍﻝﺠﻭﻝﺔ‬ ‫ﻭﺨﻼل‬
‫ﺍﻝﺭﺍﺒﻌﺔ‬
0.0 0.0
‫ﺍﻝﻌﺎﻡ‬ ‫ﻭﺨﻼل‬
‫ﺍﻝﻤﺎﻀﻲ‬
during the past phase 1.0 0.53
‫ﺍﻷﻋﻭﺍﻡ‬ ‫ﻭﺨﻼل‬
‫ﺍﻝﻌﺩﻴﺩﺓ‬
phase 0.33 0.177
‫ﺍﻷﺸﻬﺭ‬ ‫ﻭﺨﻼل‬
‫ﺍﻝﻘﻠﻴﻠﺔ‬
0.0 0.0
‫ﺍ‬ ‫ﻭﺨﻼل‬
‫ﻝﻔﺘﺭﺓ‬
‫ﺍﻻﺠﺘﻤﺎﻉ‬ ‫ﻭﺨﻼل‬ during the meeting 0.66 0.35

‫ﺍﻝﺨﺎﺹ‬
‫ﺍﻝﺼﻴﻑ‬ ‫ﻭﺨﻼل‬
‫ﺍﻝﻤﺎﻀﻲ‬
during the past 0.66 0.35
0.0 0.0
‫ﺍﻝﺘﺤﻀﻴﺭﻴﺔ‬
0.0 0.0
‫ﺍﻝﻤﻤﺘﺩﺓ‬
included range of
participants at this
stage
0.66 0.51
‫ﺍﻻﻨﺘﻘﺎﻝﻴﺔ‬
and during the
transition period
0.66 0.51
‫ﺍﻝﺘﺎﻝﻴﺔ‬
period of 12 months
included
0.66 0.51
0.0 0.0
‫ﺍﻷﺨﻴﺭﺓ‬
included 0.33 0.25
Table 10 –System translations for the general-fragments that were found for the input sentence:
‫ﺒﺎﻝ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬
‫ﺘﻘﺭﻴﺭ‬ (wxlAl Alftrp Alm$mwlp bAltqryr, “and during the reporting period”)
The input pattern and example pattern columns are the same as the previous table. The
translation column presents the translations that were created for each general-fragment.
The translation-score and the total-score are also given.
5.5.4 Recombination
In this step the system pastes together the translations of the general-fragments in order to
produce the final translation of the input sentence. As mentioned above, currently we are
not smoothing out the recombined translation to form a coherent grammatical English
sentence. The recombination step is described in Section 5.4.
Table 11 presents the final N best translations with their scores.
Translations Score
and during the reporting period 0.75
and throughout the period | the reporting period 0.66
and throughout the period | report 0.5

and throughout the period | included a period | report 0.325
during the past phase | report 0.266
Table 11 – Final five best translations and total scores for the input sentence:
‫ﺒﺎﻝﺘﻘﺭﻴﺭ‬ ‫ﺍﻝﻤﺸﻤﻭﻝﺔ‬ ‫ﺍﻝﻔﺘﺭﺓ‬ ‫ﻭﺨﻼل‬ (wxlAl Alftrp Alm$mwlp bAltqryr, “and during the reporting period”)

Chapter 6
Experimental Results and Conclusions

6. Experimental Results and Conclusions
In this chapter we present some experimental results, future work and conclusions.
6.1 Results
Experiments were conducted on a corpus containing 13,500 translation examples. The
following results are based on 400 Arabic short sentences (5.5 words per sentence on
average) that were taken from unseen documents of the United-Nations inventory. The
ten best results were evaluated by some of the common automatic criteria for machine
translation evaluation, although our system does not perform the entire translation
process (discarding the important smooth out task, as explained). All evaluations
computed case-insensitive (lower-case), with no punctuation marks and substituting
hyphens '-' with blank space.
BLEU score [42]: Comparing a mechanical system’s translation to human-produced
reference translations, the BLEU score uses unigram, bigram, trigram and fourgram co-
occurrence precision in combination with a brevity penalty for short sentences. We
evaluated our results using only two translation references. We expect a significant
improvement of the results as the number of references will increase. The fact that our
system may recombine translated fragments that intersect at a single word and the lack of
the smoothing-out step, means that sometimes translations are not fluent enough, so in
addition to the common 4-gram BLEU score, we also include 1,2,3-grams BLEU score.
NIST score [43]: This score is similar to BLEU. It also uses one- to four-gram co-
occurrence precision, but it takes the arithmetic mean of the n-gram counts. Both BLEU
and NIST measure accuracy, so higher scores are better.

METEOR score [44]: This score is based on unigram co-occurrence precision. Here,
unigrams can be matched based on their surface forms, stemmed forms and also their
meanings taken from WordNet.
Table 12 presents the results for the experiments.
BLEU
1-gram 2-gram 3-gram 4-gram
NIST METEOR
Best translation
chosen by the
system
0.4421 0.3183 0.2375 0.1849 4.1792 0.4851
Best translation
chosen by a human
referee
0.5229 0.3960 0.3100 0.2488 5.1281 0.5363
Table 12 – Evaluation of 400 Arabic UN sentences (5.5 words per sentence in average)
The first row of the table presents evaluations of the system’s highest ranked result for
each input sentence. The second row shows the same evaluation but for the best result
that was chosen by a human referee from the ten best results reported by the system for
each input sentence. In most cases, the best result that was chosen by the referee had a
close (or even the same) translation-score as the system’s best result.
Just for example, Table 13 shows the 2005 report of the annual NIST evaluation event for
machine translation systems that translates Arabic text to English [45]. Participating
systems were required to translate 100 Arabic articles into English. Each system was
evaluated by the BLEU score comparing it to four sets of high-quality independently
generated human translation.

Site BLEU (4-gram) Score
GOOGLE 0.5131
ISI 0.4657
IBM 0.4646
UMD 0.4497
JHU-CU 0.4348
EDINBURGH 0.3970
SYSTRAN 0.1079
MITRE 0.0772
FSC 0.0037
Table 13 –The report from 2005 of the annual NIST evaluations for machine translation systems that
translates Arabic text to English
Of course, at this point it is impossible to compare those scores to ours, since, besides the
fact that our system is not completed yet, we also evaluated the system on a different set
of sentences.
As mentioned, currently, we did not consider running time as an important factor, so
translating a sentence may sometimes take an unreasonable amount of time. This happens
especially due to the fact that during the matching step, we pass many fragment
candidates to the transfer step for translation. (Actually, in our 400-sentence experiment,
the average number of fragments given to the transfer step per input sentence is about
1900.) Transfer is a time-consuming step, since it uses several algorithms for different
cases, and some of them involve calling external time consuming tools, such as WordNet
and the base-phrase chunker. Surely, more work needs to be done to dramatically reduce
the amount of time that is required for a simple translation.
6.2 Future Work
The first planned issue is to enrich and enlarge the existing corpus and check how it will
affect system’s results. We are also checking the possibility of buying an existing
sentence-aligned bilingual corpus.

We will further investigate the matching and transfer steps in order to find a way to
reduce the number of fragment candidates for the transfer step, so that the entire running
time will be significantly reduced.
On the Arabic side, we will investigate many complicated situations that have not yet
handled; on the English side, the most important issue is the development of the missing
last recombination step.
When we finish implementing all the components, especially the recombination step, we
will be able to participate in the NIST annual evaluation event, and compare our results
with other existing Arabic systems.
6.3 Conclusions
We believe that we have demonstrated the potential of the example-based approach for
Arabic, with only minimum investment in Arabic syntactical and linguistic issues.
We found that matching fragments on the level of lemma and stem, and also part-of-
speech, enabled the system to better exploit the small number of examples in the corpus
we used.
As mentioned, more work is needed to enlarge and enrich the corpus, as well as rules to
deal with various problematic situations that are not yet handled. This all appears quite
feasible.
Finally, we do not claim that the example-based method is sufficient to handle the
complete translation process. It seems that, for Arabic, it should work together with some
kind of a rule-based engine as part of a multi-engine system, so as to better handle more
complicated situations.

References
[1] Y. Bar Hillel, “The Present Status of Automatic Translation of Languages”. Advances
in Computers 1, pp. 91-163, 1960.
[2] B. Vauquois, “A Survey of Formal Grammars and Algorithms for Recognition and
Translation in Machine Translation”. FIP Congress-68, Edinburgh, pp. 254-260, 1968.
[3] B. J. Dorr, P. W. Jordan and J. W. Benoit, “A Survey of Current Paradigms in
Machine Translation “. Advances in Computers, Vol. 49, M. Zelkowitz (Ed), Academic
Press, London, pp. 1-68, 1999.
[4] M. Nagao, “A Framework of Mechanical Translation between Japanese and English
by Analogy Principle”. In A.Elithorn and R.Banerji, editor, Artificial and Human
Intelligence. North-Holland, 1984.
[5] S. Nirenburg, S. Beale and C. Domashnev, “A Full-Text Experiment in Example-
Based Machine Translation”. In Proceedings of the International Conference on New
Methods in Language Processing, NeMLap Manchester, UK pp. 78-87, 1994.
[6] S. Sato and M. Nagao, “Toward Memory-Based Translation”. In the 13th
International
Conference on Computational Linguistics (COLING ’90), Vol. 3, pp. 247-252, 1990.
[7] R. Brown, “Example-Based Machine Translation in the Pangloss System”. In
COLING ’96, pp. 169-174, Copenhagen, 1996.
[8] K. Imamura, H. Okuma, T. Watanabe and E. Sumita, “Example-Based Machine
Translation Based on Syntactic Transfer with Statistical Models”. In the 20th
International
Conference on Computational Linguistics (COLING ’04), Vol. 1, pp. 99-105, 2004.

[9] E. Sumita and H. Iida, “Experiments and Prospects of Example-Based Machine
Translation”. In Proceedings of the ACL 29th
Annual Meeting, pp. 185-192, Berkeley,
CA, 1991.
[10] C. Brockett, T. Aikawa, A. Aue, A. Menezes, C. Quirk and H. Suzuki, “English-
Japanese Example-Based Machine Translation Using Abstract Linguistic
Representations”. In Proceedings of the Workshop on Machine Translation in Asia,
Taiwan, 2002.
[11] H. L. Somers, “Review Article: Example-based Machine Translation”. Machine
Translation 14, pp. 113-157, 1999.
[12] M. Diab, K. Hacioglu and D. Jurafsky, “Automatic Tagging of Arabic Text: From
Raw Text to Base Phrase Chunks”, The National Science Foundation, USA, 2004.
[13] T. Buckwalter, “Buckwalter Arabic Morphological Analyzer Version 1.0“.
Linguistic Data Consortium, Philadelphia, 2002. URL http://www.ldc.upenn.edu/Catalog/
CatalogEntry.jsp?catalogId=LDC2002L49 (viewed on 21/11/2006)
[14] J. Kremers, “The Arabic Noun Phrase”, LOT (Dutch Graduate School of
Linguistics), Utrecht University, The Netherlands, 2003. URL http://www.lotpublications
.nl/index3.html (viewed on 27/11/06).
[15] A. Freeman, “Andrew Freeman's Perspectives on Arabic Diaglossia”, 1996. URL
http://www-personal.umich.edu/~andyf/digl_96.htm (viewed on 26/11/06)
[16] I. A. Al-Sughaiyer and I. A. Al-Kharashi, “Arabic Morphological Analysis
Techniques: A Comprehensive Survey”. In JASIST (Journal of the American Society for
Information Science and Technology) Vol. 55, Num. 3, pp. 189-213, 2004.

[17] N. Habash, “Introduction to Arabic Natural Language Processing”. Tutorial in the
ACL 43th
annual meeting, University of Michigan – Ann Arbor, 2005.
[18] J. Dichy, “On Lemmatization in Arabic, A Formal Definition of the Arabic Entries
of Multilingual Lexical Databases”. In Proceedings of the ACL 39th
Annual Meeting.
Workshop on Arabic Language Processing; Status and Prospect. Toulouse, pp. 23-30,
2001.
[19] M. F. Porter, “An Algorithm for Suffix Stripping”. In Readings in Information
Retrieval, K. Sparck Jones and P. Willett, Eds. Morgan Kaufmann Multimedia
Information And Systems Series. Morgan Kaufmann Publishers, San Francisco, CA, pp.
313-316, 1997.
[20] S. Khoja and R. Garside, “Stemming Arabic Text”. Computing Department,
Lancaster University, Lancaster, U.K, 1999.
[21] L. S. Larkey and M. E. Connell, “Arabic Information Retrieval at UMass in TREC-
10”. In Proceedings of the TREC 2001. Gaithersburg, Maryland, pp. 562-570, 2001.
[22] K. Beesley, “Finite-State Morphological Analysis and Generation of Arabic at
Xerox Research: Status and Plans in 2001”. Xerox Research Centre Europe, 2001. URL
http://www.xrce.xerox.com/Publications/Attachments/2001-009/finite-state.pdf (viewed
on 21/11/2006)
[23] J. Berri, H. Zidoum and Y. Atif, “Web-Based Arabic Morphological Analyzer”, In
Gelbukh, A (Ed.): CICLing 2001, LNGS 2004, pp. 216-225. Springer-Verlag Berlin
Heidelberg, 2001.

[24] A. M. Attia, “A Large-Scale Computational Processor of the Arabic Morphology,
and Applications.” A Master Thesis, Faculty of Engineering, Cairo University, Cairo,
Egypt, 2000.
[25] K. Darwish ,”Building a Shallow Arabic Morphological Analyzer in One Day“. In
Proceedings of the ACL 40th
Annual Meeting. Workshop on Computational Approaches
to Semitic Languages, Philadelphia, PA, 2002.
[26] J. Allen, “Natural Language Understanding”, Benjamin Cummings, 2nd Edition,
Chp. 3, pp. 41-74, 1994.
[27] K. Daimi, “Identifying Syntactic Ambiguities in Single-Parse Arabic Sentence”.
Journal of Computers and the Humanities 35, pp. 333–349, 2001.
[28] E. Othman, K. Shaalan and A. Rafea, “A Chart Parser for Analyzing Modern
Standard Arabic Sentence”. In Proceedings of the MT Summit IX Workshop on Machine
Translation for Semitic Languages: Issues and Approaches, New Orleans, Louisiana,
USA, September, 2003.
[29] The Penn Treebank Project , Computer and Information Science, Penn University.
URL http://www.cis.upenn.edu/~treebank/ (viewed on 27/11/06)
[30] E. Brill, “Some Advances in Transformation-Based Part-of-Speech Tagging”. In
Proceedings of the 12th
National Conference on Artificial Intelligence, Seattle, pp. 722-
727, 1994.
[31] A. Freeman, “Brill's POS Tagger and a Morphology Parser for Arabic”. In
Proceedings of the ACL 39th
Annual Meeting. Workshop on Arabic Language
Processing; Status and Prospect. Toulouse, France, 2001.
[32] The Brown Corpus, URL http://icame.uib.no/brown/bcm.html (viewed on 27/11/06)

[33] S. Khoja. “APT: Arabic Part-Of-Speech Tagger”. In Proceedings of the Student
Workshop at the Second Meeting of the North American Chapter of the Association for
Computational Linguistics (NAACL2001), Carnegie Mellon University, Pennsylvania,
2001.
[34] S. Khoja, R. Garside and G. Knowles, “A Tagset for the Morphosyntactic Tagging
for Arabic”. In Proceedings of the Corpus Linguistics. Lancaster University (UK), Vol.
13 - Special Issue, 341, 2001.
[35] British National Corpus (BNC), URL http://www.natcorp.ox.ac.uk (viewed on
27/11/06)
[36] J. Allen, “Natural Language Understanding”, Benjamin Cummings, 2nd Edition,
Chp. 2, pp.25-36, 1994.
[37] L. Ramshaw and M. Marcus, “Text Chunking Using Transformation-Based
Learning". In Proceedings of the 3rd
ACL Workshop on Very Large Corpora, MIT, June,
1995.
[38] United Nations Official Document System (ODS), URL http://www.ods.un.org
(viewed on 29/11/06)
[39] P. Fung and K. McKeown, “Aligning Noisy Corpora Across Language Groups:
Word Pair Feature Matching by Dynamic Time Warping”. In Proceedings of the 1st
Conference of the Association for Machine Translation in the Americas (AMTA-94), pp.
81-88, Columbia, Maryland, USA, 1994.
[40] Y. Choueka, E. S. Conley and I. Dagan, “A Comprehensive Bilingual Word
Alignment System. Application to Disparate Languages: Hebrew and English”. In J.

Veronis, “Parallel Text Processing: Alignment and Use of Translation Corpora”, Kluwer
Academic, 2000.
[41] E. Brill, “A Simple Rule-Based Part of Speech Tagger”. In Proceedings of the
DARPA, Speech and Natural Language Workshop. pp. 112-116. Morgan Kauffman. San
Mateo, California, 1992.
[42] K. Papineni, S. Roukos, T. Ward and W. J. Zhu, “Bleu: a Method for Automatic
Evaluation of Machine Translation”. In Proceedings of the ACL 40th
Annual Meeting, pp.
311-318, Philadelphia, PA, July, 2002.
[43] G. Doddington, “Automatic Evaluation of Machine Translation Quality Using n-
gram Co-occurrence Statistics”. In Proceedings of Human Language Technology, pp.
128-132, San Diego, California, March, 2002.
[44] S. Banerjee and A. Lavie, “Meteor: An Automatic Metric for MT Evaluation with
Improved Correlation with Human Judgments”. In Proceedings of the ACL 43th
Annual
Meeting. Workshop on Intrinsic and Extrinsic Evaluation Measures for MT and/or
Summarization, pp. 65-72, Ann Arbor, MI, June, 2005.
[45] NIST Evaluations Report for 2005. URL http://www.nist.gov/speech/tests/mt/mt05
eval_official_results_release_20050801_v3.html (viewed on 26/12/06)

Appendix
Here are some input/output examples using our system. For every example we present
two different human-made reference translations. System results are shown with their
total scores. Fragment translations are separated by ‘|’.
Input: ‫ﺍﻝﺸﺤﻥ‬ ‫ﺴﻔﻴﻨﺔ‬ ‫ﻋﻥ‬ ‫ﺍﻝﺴﻠﻁﺎﺕ‬ ‫ﺃﻓﺭﺠﺕ‬
Reference Translation 1: the cargo was released
Reference Translation 2: the authorities have released the cargo ship
System Results:
Result Total Score
authorities have also release | from ships | 0.39
authorities have also release | some of whom have
responsibility for competition ship |
0.2773
authorities have also release | authorities about the
development |
0.2773
authorities have also release | authorities about | from
ships |
0.2535
authorities have also release | authorities about | some of
whom have responsibility for competition ship
0.1802
Input: ‫ﺃﻭﺭﻭﺒﺎ‬ ‫ﻓﻲ‬ ‫ﻭﺍﻝﺘﻌﺎﻭﻥ‬
Reference Translation 1: and cooperation in europe
Reference Translation 2: and the cooperation in europe
System Results:
Result Total Score
cooperation | in europe | 0.65
cooperation in a europe | 0.511
Input: ‫ﻤﺸﺘﺭﻜﺔ‬ ‫ﺠﻤﺎﻋﻴﺔ‬ ‫ﻤﺴﺅﻭﻝﻴﺔ‬
Reference Translation 1: common and shared responsibility
Reference Translation 2: shared and common responsibility
System Results:
Result Total Score
collective responsibility | social common market | 0.65
social common responsibility | 0.533

Input: ‫ﺇﻗﻠﻴﻤﻴﺔ‬ ‫ﻤﻨﻅﻤﺎﺕ‬ ‫ﺴﻴﺎﻕ‬ ‫ﻓﻲ‬
Reference Translation 1: within regional organizations
Reference Translation 2: in the context of regional organizations
System Results:
Result Total Score
the context in | regional organizer | 0.4225
the context in | context regional organizer | 0.3466
organizations or context | regional organizer | 0.3322
regional organizer | 0.325
the context in | organizations or context | regional
organizer |
0.2746
Input: ‫ﺍﻝﺫﺭﻱ‬ ‫ﺍﻹﺸﻌﺎﻉ‬ ‫ﺒﺂﺜﺎﺭ‬ ‫ﺍﻝﻤﻌﻨﻴﺔ‬
Reference Translation 1: on the effects of atomic radiation
Reference Translation 2: that is concerned with the effects of the atomic radiation
System Results:
Result Total Score
effects of atomic radiation | 0.75
regarding the bamako convention on the ban | of atomic
radiation |
0.65
regarding the bamako convention on the ban | effects of
atomic radiation |
0.65
radiation s groups in india effects | of atomic radiation | 0.355
Input: ‫ﺘﻀﻊ‬
‫ﺍﻋﺘﺒﺎﺭﻫﺎ‬ ‫ﻓﻲ‬
Reference Translation 1: bearing in mind
Reference Translation 2: taking into consideration
System Results:
Result Total Score
to take into consideration | 0.511
by | viewed | 0.5

Machine Transalation.pdf

More Related Content

Similar to Machine Transalation.pdf (20)

Recently uploaded (20)

Machine Transalation.pdf