Computational linguistics

Incorporating Subject Areas into the Apertium
Machine Translation System

Jordi Duran, Lluís Villarejo, Mireia Farrús, Sergio Ortiz, and Gema Ramírez*

Abstract. The Universitat Oberta de Catalunya (Open University of Catalonia,
UOC), is a public university based in Barcelona. The UOC is characterised by
three main factors: (a) it is a virtual university based in an e-Learning model, (b) it
is based in a strongly Spanish-Catalan bilingual region, and (c) students come
from around the world, so that linguistic and cultural diversity is a crucial factor.
Within this context, it becomes essential to meet the UOC's linguistic needs
taking into account its particular characteristics. One of the tools created to this
end is the adaptation of Apertium, a free/open-source rule-based machine
translation platform, which can be found under http://apertium.uoc.edu/,
customised to the translation needs of the institution in order to offer the best
possible service to their user community.
In order to continue adapting and adding value to the existing tool for
generalisable large-scale applications, the UOC's translation system has recently
implemented a semantic filter based on subject fields aimed at improving the
translation quality and at better fitting the university needs. The paper will explain
all the steps of this adaptive process, as well as a demonstration of the resulting
tool: (a) the choice of the subject fields according to the university studies, (b) the
design and implementation of the dictionaries used to extract the required
information to filter and disambiguate homonym and polysemous terms, including
source code in the dictionaries, and (c) the design and implementation of the
corresponding web interface.

1 Introduction
During the last decades, machine translation (MT) tools have proved highly useful
in human translation work, since they save human translators a great deal of time

Jordi Duran · Lluís Villarejo · Mireia Farrús
Universitat Oberta de Catalunya, Barcelona, Spain
e-mail: lvillarejo@uoc.edu, mfarrusc@uoc.edu, jdurancal@uoc.edu

Sergio Ortiz · Gema Ramírez
Prompsit Language Engineering, Alacant, Spain
e-mail: sortiz@prompsit.com, gramirez@prompsit.com

A. Przepiórkowski et al. (Eds.): Computational Linguistics, SCI 458, pp. 281–292, 2013.
DOI: 10.1007/978-3-642-34399-5_15 © Springer-Verlag Berlin Heidelberg 2013

282 J. Duran et al.

by quickly providing initial draft translations (Craciunescu et al. 2004; Kirchhoff
et al. 2004; Villarejo et al. 2009a).
The Language Technologies group from the Office of Learning Technologies at
the Universitat Oberta de Catalunya (Open University of Catalonia, UOC) and the
university’s Language Service have recently been working on improvements to
the UOC’s Apertium-based translation (http://apertium.uoc.edu). Apertium is an
open-source MT platform that can be easily modified and adapted to users’ needs
(Forcada et al. 2009).
At the time of writing, the UOC has created an Apertium-based web application
incorporating specific terminology used by the university. It has also contributed
to the development of the system multiple-format support (Word, PowerPoint,
Excel, Writer, Calc, Impress, HTML files, compressed files, etc.) for document
translation service, which includes the option of creating and exporting translation
memories that allow for the customisation of translations (Villarejo et al. 2009a;
Villarejo et al. 2009b; Villarejo et al. 2010). The advantage of working with
Apertium is that all linguistic improvements made to the system are saved in a
common repository so that its entire community of users can benefit from them.
The Apertium application was placed at the disposal of the UOC’s internal
community (the university’s teaching staff and staff who manage and edit teaching
material) in December 2009, and was well received by the users. Soon we realised
that a user translating legal texts and another one translating computing texts
required from different translations. With a view to obtaining high quality
translations fully adapted to the university, the UOC subsequently set about to
implement the one sense per discourse assumption to improve the quality of
Apertium’s translations.
The one sense per discourse assumption states that if a polysemous word is
used more than once in a coherent discourse, it is very likely that all its
occurrences refer to the same sense. This assumption has been proved to be
effective to improve results in tasks like Word Sense Disambiguation (Gale et. al.
1992) and Statistical Machine Translation (Carpuat 2009). If we could detect that
some translation pairs (source and target word) are specifically used in some of
the university subject areas, we could build up a dictionary with them, ask the user
to manually introduce the subject area for the source text which is to be translated
and implement this way the one sense per discourse assumption. In this paper we
describe how we incorporated subject areas related to the university fields of
knowledge and the subjects taught into Apertium in order to improve the
translation quality. The inclusion of subject areas in the MT system, where they
act as semantic filters, allows for the disambiguation of homonyms and
polysemes. The translation process thus takes place on the basis of the topic of the
source text.
This chapter describes the work carried out in relation to incorporating subject
areas into Apertium. Sections 1 to 6 of the article deal with the planning of the
project, and sections 7 to 9 with its execution and conclusions, plus future lines of
work.

Incorporating Subject Areas into the Apertium Machine Translation System 283

2 Relevance of the Project
The new Apertium interface enables any user to translate documents on the basis
of the subject area corresponding to each source text. The function in question can
also be deactivated by selecting the ‘General vocabulary’ option instead of a
specific subject area.
Additionally, the subject area analysis process makes it possible to identify new
specific terminology and incorporate it into the MT system, as well as to produce
glossaries of terms to complement universities’ teaching material and academic
and administrative documentation.
As Apertium is an open-source platform, all linguistic improvements made to it
are stored in a freely accessible general repository. Any user can thus benefit from
the semantic filters implemented in this project. Given that the subject areas have
been established in relation to the UOC’s fields of knowledge, the improvements
made to Apertium in this project will be extremely useful in any sphere related to
the university.

3 Goals
The incorporation of subject areas into the Apertium machine translation system
within the context of the Universitat Oberta de Catalunya sought to accomplish
some objectives. In the first place, a specific analysis and selection of the most
useful subject areas for the UOC's community should be performed. As a
consequence, the subject areas will serve as filters for the MT system with a view
to improving its translated output. Secondly, the creation of a glossary of
polysemes and homonyms which can be disambiguated by assigning subject areas
to them will be fulfilled. And finally, the implementation of the subject filters in
the MT system's interface so that users can choose a subject area and generate a
much more accurate translation featuring the terminology used in the field in
question will be carried out.

4 Beneficiaries
First of all, it is of great importance for the success of the project to identify those
target people that will be potential users of the new machine translation. The main
groups set to benefit from the project are the following ones:

(1) Administrative staff looking to circulate versions of documents or
communiqués in various languages and who want to use MT to speed up their
work.
(2) Linguists who correct and translate universities’ teaching, academic and
administrative documentation.
(3) External language professionals who work with university language services.

284 J. Duran et al.

(4) Teaching staff seeking to use MT to prepare the teaching material with which
they provide students.
(5) Students who want to use MT to obtain rough translations to enable them to
understand material or carry out work in languages in which they lack
proficiency.
(6) In general, institutions, organisations, businesses, the media and members of
the public interested in using MT in their day-to-day activities.

5 Project Stages and Groups Involved
Language technology projects require collaboration between two types of experts,
namely linguists, with their mastery of language, and IT engineers specialising in
natural language processing, who provide access to the technology necessary for
the implementation of such projects. In that regard, the UOC boasts a
multidisciplinary team that combines the linguistic expertise of the staff of its
Language Service and the skills of the language technology specialists from its
Office of Learning Technologies.
Broadly speaking, the project consisted of three stages. Each stage and the
groups it involved are briefly outlined below.
Stage 1 included subject area analysis and description, as well as the
compilation of university material from which specific vocabulary could be
extracted. The activities involved in this stage were carried out by the Language
Service and the Office of Learning Technologies.
Stage 2 included producing a glossary of semantically ambiguous terms and
adding tags to dictionary entries to identify their subject area. The activities
involved in this stage were carried out by the Language Service together with the
Office of Learning Technologies and Prompsit, a company that had already
participated in technical work geared to MT in similar projects. In the case in
hand, Prompsit designed the data files for subject areas format and the integration
within Apertium and collaborated in the generation of linguistic data.
Stage 3 involved the implementation of subject areas in the new Apertium
interface. The Language Service and the Office of Learning Technologies
designed the interface, assessed it from a linguistic perspective and disseminated
it, while Prompsit carried out the actual implementation work as well as
subsequent technical testing.

6 General Tool Description
At present, the UOC has an internal translation service based on Apertium
(available at http://apertium.uoc.edu). Anyone wishing to do so may download the
same version of the system from a freely accessible general code repository
(http://sourceforge.net/projects/apertium/).


Fig. 1 Proposed new interface. In the example, word "conseller" in Catalan is translated
into "concejal" (councillor) in Spanish in the Law and Political Science domain but into
"consejero" (advisor) in the Bussiness Administration domain.

The MT system can be used to translate texts entered directly in the platform’s
text box, documents in multiple formats, web pages (retaining the full original
structure), HTML files and compressed files. It also allows for the creation and
use of TMX-format (Translation Memory eXchange) translation memories, which
help to generate much more accurate translations with a greater degree of
customisation.
The proposed new interface developed in this project features a drop-down list
of subject areas from which users can choose in line with the topic of the
translation to be performed (See Figure 1). The list of subject areas is based on a
general basic classification of eight areas established by the Catalan university

286 J. Duran et al.

system’s language services for previous language technology projects. The eight
areas in question are:

(1) Pure sciences
(2) Natural sciences
(3) Computer sciences
(4) Law, economics and politics
(5) Social sciences
(6) Art and humanities
(7) Construction and architecture
(8) Industry

The above areas served as a starting point for establishing sub-areas on the basis
of the UOC’s different study programmes and requirements.

7 Goal fulfilment
The degree to which the project fulfilled the goals established in its application for
funding is described below.
The purpose of the project was to improve the UOC’s Apertium existing
interface and adapt it to the university requirements. The main aspects of the
project were as follows:

(1) Incorporation into the MT system of subject areas related to the UOC’s fields
of knowledge and the subjects taught at the university.
(2) Configuration of the subject areas to act as semantic filters for the
disambiguation of homonyms and polysemes.
(3) Enhancement of translation quality through the application of the semantic
filters.
(4) Open access for the system’s community of users, who can apply and benefit
from the improvements made.
(5) Ease of use, even for those unfamiliar with MT.

8 Stages and Execution of the Project
As stated in section 5, the work was carried out in three different stages, which are
described below in detail.

Stage 1

a) Selecting subject areas

As mentioned previously, the classification established by the Catalan university
system’s language services (see Table 1) was used as a starting point for selecting


Table 1 Subject areas established by the Catalan university system’s language services
(2001)

Area Sub-areas
1 Pure sciences Chemistry, physics, mathematics, statistics
2 Natural sciences Earth science, life sciences, health sciences,
veterinary science, agriculture, stockbreeding,
fishing, environmental science
3 Computer science Computer science, telecommunications, electronics,
electrical engineering
4 Law, economics Law, public administration, defence, economics,
and politics home economics, administration, work and
employment, politics
5 Social sciences Sociology, anthropology, psychology, pedagogy,
communication, information, sports, games, tourism,
hotel industry, food
6 Art and humanities Geography, history, philology, philosophy, religion,
art, music
7 Construction and Construction, architecture
architecture
8 Industry Industrial engineering, various industries

subject areas to act as semantic filters. The classification was further developed to
include the subject areas contained in the UOC’s institutional repository and
analysed in this project (see Table 2).

b) Compiling material

Parallel documents (i.e. documents in Catalan with translations into Spanish or
English) corresponding to different subject areas (see Table 1) were used for the
purpose of identifying ambiguous terms. The documents in question were taken
from the volume of texts corrected and translated by the UOC’s Language Service
from July to December 2009.

Stage 2

a) Designing dictionaries

Apertium has a monolingual dictionary for each of its languages and a bilingual
dictionary for each of its language pairs. Each dictionary, be it monolingual or
bilingual, is a separate file that can be modified and used on the basis of the
language pair involved in a translation.
The same system of separate files was used for the purpose of assigning subject
areas to ambiguous terms. More specifically, two different files were used, one for

288 J. Duran et al.

Table 2 Subject areas analysed in this project

0. General
2. Natural sciences
2.1. Health sciences
3. Computer science
3.1.1. Computer science engineering
3.1.3. Technical engineering in computer systems
3.2. Multimedia
3.3. Free software
3.4. Communication technologies
4. Law, economics and politics
4.1. Economics and business
4.1.1. Business administration and management
4.1.2. Labour sciences
4.1.3. Business studies
4.1.5. Marketing and market research
4.2.1. Law
4.2.2. Political sciences and public management
4.2.3. City management and urban planning
5. Social sciences
5.1. Information and knowledge society
5.2.2. Teaching
5.2.3. Psychology
5.2.4. Educational psychology
5.2. Psychology and educational sciences
5.3. E-learning
5.4.1. Communication
5.4.2. Audiovisual communication
5.4.3. Information and documentation
5.5.1. Tourism
5.5.2. Travel programme
5.6. Food systems, society and culture
5.7. Humanitarian cooperation, peace and sustainability
6. Art and humanities
6.1. Humanities
6.1.1. Philosophy
6.2. Languages and cultures
6.3.1. Catalan philology
6.3.2. Translation and technologies
6.3.4. Arab and Islamic studies
6.3.5. Cultural management
8. Industry


specifying subject areas and the other for indicating the grammatical and semantic
information required for disambiguating terms:

● Subject area file. This file contains a list of subject areas, each with an
identifying code. For instance, the e-learning subject area, which has been
assigned the code 5.3, is represented as follows:

<d n="5.3">e-learning</d>

● Disambiguation file. This file indicates the different subject areas that can be
assigned to a particular term. The Spanish term mesa, for example, can be
assigned to the general corpus area (code 0), the law, economics and politics
area (code 4) or the humanities area (code 6.1). The corresponding Spanish-
Catalan dictionary entries are as follows:

<e d="0"><p><l>mesa<s n="n"/></l><r>taula<s n="n"/>.
<e d="4"><p><l>mesa<s n="n"/></l><r>mesa<s n="n"/>...
<e d="6.1"><p><l>mesa<s n="n"/></l><r>tabula<s n="n"/>...

The following rules must be observed where the disambiguation file is
concerned:

(1) At least the grammatical category must be indicated to take
advantage of the Apertium part-of-speech disambiguation
module based on hidden Markov models (Rabiner 1989), which
automatically selects the most probable part-of-speech for
homographs.
(2) The specific subject area must always be indicated.
(3) The terms that appear in the file must be incorporated into the
corresponding monolingual and bilingual dictionaries.
(4) If a single term has two translations within a given subject area,
an exclamation mark must be used to denote the translation to be
prioritised. The criterion used to prioritise a translation over
another is explained in the following section.
b) Extracting ambiguous terms

The first step in analysing ambiguous terms on the basis of the parallel corpus
consisted of extracting relevant terms for each subject area, including the general
area. Term relevance within an area was determined on the basis of its keyness,
i.e. its quality of being key in its context (Scott and Tribble 2006; Bondi & Scott
2010). If, proportionately speaking, a term appears with significantly greater
frequency in a text corresponding to a specific area than in an extensive general
corpus (a reference corpus), it is said to have keyness in the area in question. The
higher a term’s keyness value in an area, the more relevant it is within that area.

290 J. Duran et al.

The extraction of ambiguous terms was carried out in six steps, applying the
concept of keyness:

Step 1. Calculating keyness values for each area

The keyness values of the terms appearing in each specific area were calculated,
giving rise to a list (let’s name it List 1), which is an ordered list showing each
term’s relevance in the area in question.

Step 2. Looking up terms in different areas (I)

In each subject area, a search was performed for terms also appearing in at least
one other area and which thus had a certain probability of being ambiguous. This
gave rise to another list (let’s name it List 2) consisting of pairs of words and its
keyness value. (Examples of the eight words with highest keyness values in the e-
learning area in Catalan are: “ser (to be) 1513”; “haver (to have) 1437”;
“aprenentatge (learning) 1199”; “uoc: 918”; “anar (to go) 713”; “xarxa (net)
670”; curs (course) 659; ensenyament (teaching) 659”; etc..

Step 3. Filtering using the reference corpus

The thousand most relevant terms, according to the keyness value, in the general
reference corpus were extracted. Given that terms that are highly relevant in the
reference corpus are held to be of little relevance in specific areas, those extracted
from the reference corpus were deleted from List 2. Generic verbs (e.g. to be, to
have, to make, to do and to go) were also deleted. The result was List 3, consisting
of pairs of words and its keyness value. (Examples of the nine words with highest
keyness values in the e-learning area in Catalan are: “aprenentatge (learning)
1199”; “uoc 918”; “xarxa (net) 670”; “curs (course) 659”; “ensenyament
(teaching) 659”; “nou (new) 594”; “social (social) 572”; “obert (open) 561”;
“estudiant (student) 561”; etc.

Step 4. Looking up terms in different areas (II)

With the initial lists having been altered, a new search for terms appearing in two
or more areas was performed, giving rise to List 4.

Step 5. Filtering terms

Terms from languages other than that under analysis were deleted from List 4.
Terms other than verbs, nouns or adjectives were subsequently deleted. The result
was List 5.

Step 6. Checking for ambiguity

Having obtained a filtered list of terms and the subject areas in which they appear,
a search was performed for terms with more than one translation in other


languages and in other subject areas. Firstly, to that end, the texts of the Catalan-
Spanish and Catalan-English documents corresponding to each area were aligned
by using the Hunaling alignment tool (version 1.0). Next, the aligned sentences
containing the relevant terms were extracted, classified by area. That information
was used to manually determine which terms have different translations in
different subject areas. A final list of relevant terms by subject area having
different translations in different subject areas was obtained.

Stage 3

The list of terms with different translation by subject areas was incorporated into
the translation engine’s disambiguation files. Apertium's workflow was modified
to cope with this information during translation and, finally, the UOC's Apertium
user interface was modified to allow subject area selection. At the time of writing
(April, 2012), the apertium.uoc.edu translation service is awaiting the
corresponding update.

9 Conclusions and Future Lines of Work
By way of conclusion, it must be emphasised that the project’s original goals were
fulfilled. The UOC’s Office of Learning Technologies worked with the
university’s Language Service and Prompsit to ensure that the incorporation of
subject fields into the Apertium MT system was carried out with the utmost
linguistic and technological rigour.
Nevertheless, the UOC aims at constantly improve the needs of the academic
community, and to this end, new lines of work are planned as future developments
related to the Apertium MT system.
First of all, one the future objectives is to extend the range of potentially
ambiguous UOC terminology incorporated into the new interface featuring subject
areas. Second and related to the first objective, another goal is to extend the range
of subject areas that serve as semantic filters as the UOC adds further study
programmes to the courses it offers. Third, the UOC aims at systematising the
ambiguous terminology extraction process as more parallel documents become
available from the Language Service and to automatise polysemy disambiguation
by exploring the most recent work inside Apertium related with automatic lexical
selection (Tyers 2012). Finally, an evaluation of the usefulness of each term
identified in this project is also in the scope ot the future work and will be done
through the analysis of new documents corresponding to each subject area
covered.

Acknowledgements. The work in question was made possible by funding that the article’s
authors obtained from the Office of the Vice President for Research and Innovation of the
UOC, through the APLICA call for innovation projects geared to management made by the
university’s Open Innovation Office.

292 J. Duran et al.

References
Bondi, M., Scott, M.: Keyness in Texts. Benjamins, Amsterdam (2010)
Carpuat, M.: One sense per discourse. In: Proceedings of the NAACL HLT Workshop on
Semantic Evaluations: Recent Achievements and Future Directions, Boulder, Colorado,
pp. 19–27 (June 2009)
Craciunescu, O., Gerding-Salas, C., Stringer-O’Keeffe, S.: Machine Translation and
Computer-Assisted Translation: a New Way of Translating? The Translation Journal
8(3) (2005)
Forcada, M.L., Tyers, F.M., Ramírez-Sánchez, G.: The Free/Open-Source Machine
Translation Platform Apertium: Five Years on. In: Proceedings of the First International
Workshop on Free/Open-Source Rule-based Machine Translation FreeRBMT, Alacant,
Spain (2009)
Gale, W., Church, K., Yarowski, D.: One sense per discourse. In: HLT 1991: Proceedings
of the Workshop on Speech and Natural Language, Morristown, USA, pp. 233–237.
Association for Computational Linguistics (1992)
Kirchhoff, K., Turner, A.M., Axelrod, A., Saavedra, F.: Application of Statistical Machine
Translation to Public Health Information: a Feasibility Study. Journal of the American
Medical Informatics Association 18, 473–478 (2011), doi:10.1136/amiajnl-2011-000176
Rabiner, L.R.: A tutorial on hidden Markov models and selected applications in speech
recognition. Proceedings of the IEEE 77(2), 257–286 (1989)
Scott, M., Tribble, C.: Textual Patterns: keyword and corpus analysis in language
education. Benjamins, Amsterdam (2006)
Tyers, F.M., Sánchez-Martínez, F., Forcada, M.L.: Flexible finite-state lexical selection for
rule-based machine translation. Research paper accepted at the EAMT (to be released,
2012)
Villarejo, L., Cullen, D., Corral, A.: La integració de les tecnologies de la llengua en el flux
de treball del Servei Lingüístic de la UOC. Llengua i ús, Revista tècnica de Política
Lingüística 46 (2009a)
Villarejo, L., Ortiz, S., Ginestí, M.: Joint efforts to further develop and incorporate
Apertium into the document management flow at Universitat Oberta de Catalunya. In:
Proceedings of the First International Workshop on Free/Open-Source Rule-based
Machine Translation (2009b)
Villarejo, L., Farrús, M., Ortiz, S., Ramírez, G.: A web-based translation service at the
UOC based on Apertium. In: Proc. of the International Multiconference on Computer
Science and Information Technology, Wisla, Poland, pp. 525–520 (2010)

Computational linguistics

More Related Content

Viewers also liked (9)

Similar to Computational linguistics (20)

More from Springer (20)

Recently uploaded (20)

Computational linguistics