ONTOLOGICAL TREE GENERATION FOR ENHANCED INFORMATION RETRIEVAL

International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
DOI : 10.5121/ijaia.2013.4405 41
ONTOLOGICAL TREE GENERATION FOR
ENHANCED INFORMATION RETRIEVAL
Anwaya Aras1
,Sakshi Paratap2
and Dr.Mangesh Bedekar3
1,2,3
Department of Computer Science,BITS-Pilani University,Goa,India
1
anwayaaras@gmail.com
2
sakshi.2392@gmail.com
3
bedekar@goa.bits-pilani.ac.in
ABSTRACT:
Information visualization seeks to leverage human visual processing to make sense of abstract information.
One particularly rich class of information structures ripe for visualization are those representable as
graphs (i.e. nodes and edges), including organization charts, website linkage, and computer networks. In
this paper we propose a methodology to extract information from big data and convert it into a human
comprehensible format of graphs to give the reader an objective overall idea of the document content. We
put forth the design and implementation details to mapping our data into the Open Directory Project or the
DMOZ tree and build a hierarchical ontological tree based on the extracted metadata.
KEYWORDS:
Ontology Information; visualization; Natural Language Processing; Information Retrieval; Key word
extraction; DMOZ; Computer Science Curriculum; Graph based knowledge representation
1. INTRODUCTION
Large related data, like directories, encyclopedias, books have massive amount of data which is
both extensive and important. While it is important to understand and grasp their content, the
process of reading such documents might get time consuming and cumbersome. Text in most
documents is highly redundant and sometimes irrelevant. Entire documents would need to be read
to understand what its content is. Much of such big data, specifically the one which is related can
easily visualized using a graphical hierarchical structure and every token representing a node and
edge linking related categories .To address the problems effectively, we are trying to use
techniques to pull out phrases that seem to characterize a document and then map them to achieve
an ontological tree based graph to represent the information.
For the purpose of experimentation, we have used the academic handouts of universities. For
initial analysis purpose we are focusing on undergraduate courses as they have a well structured
curriculum and range over a wider variety of subjects. Our system effectively generates their
respective DMOZ trees [2] give us a fair idea of the depth of the courses taught which are
represented by the handout. We then compare handouts of various courses and make notes of the
inferences and then correlate them to human drawn conclusions to verify the accurate working of
our system.
It is essential to organize the results into ontology, in particular a hierarchical ontology. Ontology
is an explicit formal specification of the terms and relations among terms in a domain. It can be

42
achieved by a systematic grouping of domain concepts (e.g., user interests) based on their
definitions, in machine-interpretable form [7].
In this paper we are proposing an approach for efficient classification of data into the large topic
ontology DMoz. The Open Directory Project, also known as DMOZ, is the largest human edited
directory of Internet sites. The DMOZ directory incorporated 590,000 categories and 6 million
quality website content organized into 15 levels. ODP uses a hierarchical ontology scheme for
organizing site listings. Listings on a similar topic are grouped into categories which then include
smaller categories. One effective technique for the display of such data is a focus+context
approach that uses lightweight modeling of user interest to inform the display of information.User
interest is modeled using a Degree-of-Interest (DOI) function, which assigns a single number
representing the estimated relative interest of the user to each node in the structure.These
numbers are used to appropriately layout and render the structure, for example by controlling
which nodes are visible and which are elided [9].
2. BRIEF METHODOLOGY
Our proposed plan consists of two parts namely document key word retrival and building of
ontological trees.[5] The first part uses techniques of NLP to extract only the important
keywords-unigrams,bigrams and trigrams from the large chunk of data. In the second stage, these
words are fed into the DMOZ directory and when a particular n gram phrase hits the tree, the
entire tree is generated.
2.1. Keyword Extraction
Algorithm:
Efficient extraction of related keywords in the form of n grams models was extremely imporatnt
as it directly affects the accuracy of the DMOZ tree. Although,work on summarization of large
amounts of data has been effectively done and substantial results have been achieved[5],
information extraction from related documents with a characteristic concept flow, requires a
different approach as shown by Rahman [11].
Initially, the data is fragmented and divided into multiple tokens. Then a Part Of Speech tagger is
then run on each sentence and the words are tagged accordingly. However, this tagging is not
accurate enough and hence we have used a novel technique for tagging based on a naturally
growing resource which has used concepts demonstrated by Taskaret. al(2012)[3].Once this
tagging is complete, all the important derivatives of speech are put into the output file.We then
have hard coded a probability value, to generate unigrams, bigrams and trigrams out of the
annotated data.
The following are all the steps that encompass the entire algorithm that runs behind our keyword
extraction module. Each of the techniques described below comprise of a separate module
running at the back end of our code.
* Stemming
We use stemming to narrow our overall word matrix so as to help with the lack of similar words
per text. We have used the standard Porter's stemming algorithm for the purpose. Overall, we see
mixed results for stemming.While stemming shows improvement it also negatively affects the
overall accuracy when combined with other aspects of descriptive texts.

43
* Stop Word Removal
As in much text, that there were many common, short function words, often prepositional terms
and other grammatical syntax fillers that were found specially in descriptive texts. We have not
yet attempted to build a domain-specific we used a standard set of Porter stem words.
Ex: the, is, at, who, which, on
Stop words removal turned out to drastically reduce the length of passage and descriptive texts
thereby giving huge accuracy boost.
* Named entity recognition
Entity identification is a very important part of summarization especially because it tells the
system extremely relevant data about the text data. Though NER systems are known to be
brittle,our system has been specifically designed for the big data domain and thus it rightly
extracts the necessary words depending upon the type and setting of the question.
* Punctuation Removal
Removing punctuation was another result of looking at our data and noticing that written texts
have a very large variance in phrasing and word choice. This is especially true for words that may
have multiple accepted forms, or words with punctuation in them. Also, because we parse the
item descriptors on spaces, any punctuations that are in the phrase are left in, including ellipses,
periods, exclamation points, and others. In addition, words that are concatenated are often used
differently. Punctuation removal was the also an effective feature normalization method used for
summarization.
* Lowercasing
While writing in English, texts tend to have the first word capitalized. In addition, different
writing styles will capitalize different words intentionally or otherwise, depending on their intent
interpretation of the word, of choice of capitalizing acronyms. This is generally not a useful
normalization for the system to understand as it deteriorates the performance of the POS tagger.
Ex. President, president, CD, cd, Windows, windows
2.2 Keyword synthesis
After application of the above techniques, the system generates a relevant set of words from the
document. However, tp precisely understand what the document is trying to convey, the
frequency of each word, specific occurrence with other words and the context of its usage is
equally important. Hence, once the raw word data set is available, we further analyze the words
and the information they convey. Three modules are used for the same and the words are further
divided as :
a.Unigramlist: These are the words which occur frequently in the document i.e occur more times
than some threshold value. This threshold value is decided on the basis of the total number of
words, the count of the word which has highest occurrence and the count of the word which has
lowest occurrence
b. Bigram list: These are the words which always occur together in the document. For example
“operating systems” or “computer networks”. Their pairing and occurring together is much
meaningful than the word occurring isolated and hence they carry more weight in word extraction

44
analysis. Two words get included in the bigram list if their count of occurring together or at an
offset of at most 3 words is more than a predetermined threshold value.
c. Trigram list: If three words are seen to occur together then they are included in the trigram
list. Occurrences of such words are rare, but if they occur they covey very subtle information and
help greatly in the DMOZ tree building
.2.3. Ontological tree generation
The problem of retrieving the complete category link for an interest and ranking the categories
based on their importance is resolved through the use of DMoz category hierarchy. Every
category in the DMoz dump consists of a listing and description of external pages associated with
that category. Our DMoz based approach works as follows: each interest is searched in the DMoz
RDF dump. The categories under which one or more of the external page descriptions contain the
concerned interest are selected.
Thus, when searching for an interest such as “programming” in the dump, we see that
“Computers” occurs frequently in the external page description of the category link. All such
category links under which the interest is found are extracted. To engineer the ontology, we use
only the top level categories. This avoids large scale duplication of the interest
instances.Furthermore, with DMoz, it becomes possible to retrieve the complete category link
associated with the interest. Thus, for example, the interest “programming” in DMoz belongs not
only to the category “Computers” but to its complete category link, which is Computers-
>Internet->Cloud Computing->Programming We parse the complete category link and every
term in the link becomes a node in the ontology. The interest is made a child of the lowest node in
the hierarchy..
As seen in the figure, all interests are accurately grouped under the respective categories to which
they belong.Simultaneously the depth of the node from the start point is attached at every level to
gauge the importance especially when the levels are more than 10.
Fig 1-A view of DMOZ tree output on the word programming without using any filter

45
The nodes can be then arranged in the required way to allow every user to personalize their space.
This can help users to relate certain data, ignore unnecessary information and get more
information in a shorter period.[10]
The various features like zooming, panning on mouse and keyboard actions enable only certain
data to be visible at a given time and increase readability. Simultaneously the application supports
4 different orientations of the data, top to bottom, bottom to up, right to left, left to right, making
it easier for the user to decide depending on the requirement.
The double-click function redirects to open another pop-up window (See figure)which has the
options either to delve deeper into the tree of the clicked topic or choose among the links
corresponding to its topic. Expanding the tree both hides the other irrelevant nodes not in the path
of the clicked data as well as displays its branches .(subtopics arising from the topic) Clicking on
any other button opens the corresponding link on the default web-browser. On the left hand side
are the topics and on the right hand side is a brief description about each topic.
Figure2: Popup window on clicking Oracle.
Fig3: Result on expanding the tree

46
Fig4: Result on clicking on the website link “Oracle FAQ”
3. IMPLEMENTATION
We have primarily coded in Python and Java to avail the benefits of both the languages. While
Python offers an extensive and a fully fledged library for language processing purposes, JavaJava
is fast, secure, reliable and with its underlying object-oriented principles provides an excellent
platform to work on.
The summarization and key word extraction part has been implemented in Python programming
language using NLTK library[1]. Since python contain an extensive library for language
processing, it stood out as the best language for development of robust modules for this
project.We have developed our own POS tagger,tokeniser and hand codded features for the
named entity recognition algorithm based on the dominant aspects of big data analysis.These
modules when run together successfully extract the relevant words of our dataand further go on to
generating the n gram words for our project
The requirements of our design of DMOZ trees are best supported by Java using
Prefuse[8].Prefuse is a set of software tools for creating rich interactive data visualizations. The
original prefuse toolkit provides a visualization framework for the Java programming language.
The prefuse flare toolkit also provides visualization and animation tools for ActionScript and the
Adobe Flash Player. The label renderer and JPanel components introduce interactivity in the
visualization.

47
4. INTEGRATED SYSTEM AND RESULTS
4.1. Handout Data extraction
The course handouts of Massachusetts institute of technology (MIT) and Birla Institute of
Technology and Science (BITS) have been used for analysis. Using this method all the
information of the course can be easily represented in graphical hierarchical form, thus making it
possible to get the sense of the topics, sub-topics and the complete course structure. By The
handouts used in consideration here are of the course Computer Networks , offered by the
Computer Science department of the colleges. Techniques have been used to perform
interestword sense disambiguation . For instance here, “Security” should refer to something in the
world of Computers only. Thus, this approach constructs a simple yet effective grouping of user
interests.
Fig5 -A view of DMOZ tree output on the handout of BITS.
Fig6 -A view of DMOZ tree output on the handout of MIT.

48
4.2 Results:
The tree gives complete information of the course structure, the linkage between categories,
classification, the volume and depth of each topic that will be covered.
It can be gauged if only the overview or details of particular topics would be covered in the
duration of the course. For instance in BITS, the node ethernet has no children, thus it is less
likely that the topic will be taught in much detail. Similarly we can even compare two topics and
safely say that protocols will be focused more than ethernet in BITS.
On formation of the tree we can easily extract useful comparison from the structure of courses:
• Here both MIT and BITS have topics till the level 6, thus providing equal information on
considering depth and detail level of the course taught.
• We can easily find out the broad topics being studied, which topics overlap and
differences. Here the topic Data communications is covered in BITS, including the
subtopic ethernet while MIT does not have it.
• If we go into the depth of the tree, we see that protocols is taught in much more depth in
MIT focusing on RFCs, NAT while in BITS the focus is on studying various types of
protocols without going into intricate details.
We similarly created a tree of all the subjects of interest to get a complete picture of the subjects
ordered by only important key topics, thus providing details on the whole course in one
graph.Zooming and panning provides easy navigation to any topic and its linked branches. For
details regarding a particular topic, we can double click the interest node and go into further depth
or web links relevant to the category.We aim to use highlighting and navigation techniques to
check overlapping information between subjects, check particular courses itself and thus provide
a clear picture.

49
Fig 7.Complete tree

50
Fig 8.Zooming in a particular area
5. FUTURE WORK AND CONCLUSION
The results provided by our application were satisfactory and gave precise information as well as
provided clarity in regard to the actual topics being covered and depth and those as resulting from
our application. The technique and methodology used provided upto 90% accuracy in most cases.
One of the future scope that stands for this project is increasing the number of relevant keywords
extracted by the system and thus increasing the recall value for key word extraction module.
Specifically, the system tends to reject acronyms as it identifies them as stray words and
overlooks their instances. Robust feature extractors for training the system using NER are needed
to be built to overcome this problem and get an even more efficient system.
The interface is being made more appealing by adding additional graphics and features to
increase interactivity. The project is being deployed as a web app to allow instant result display
on any file that is provided.
In future, similar procedure can be employed to other databases and directories like Wikipedia to
improve visualization and to provide exhaustive results as well as compare how the results are
mapped in each of the given dumps. More number of colleges as well as schools’ curriculum is
being compared .Analysis on how the information is related at every stage and usefulness and
relevance of subjects when compared to specialization in a particular field in that branch is being
done. The same Information Visualization Techniques and concepts can be used to create precise
memory maps and business analytics techniques can be formulated.

51
REFERENCES
[1] Natural language toolkit. URL http://www.nltk.org.
[2] Dmoz dump: http://dmozimporter.codeplex.com/
[3] S. Li, V. Graca and B.Taskar. 2012 .Wiki-ly Supervised Part-of- Speech Tagging. In Proc.EMNLP
2012 .
[4] C. cheng Lin and H. chun Yen, “On balloon drawings of rooted trees,” pp. 12–14, 2005.
[5] Laender, A. H. F., Ribeiro-Neto, B. A., da Silva,A. S., and Teixeira, J. S.2002. A briefsurvey of web
data extraction tools. SIGMOD Rec.31,2, 84
[6] Kushmerick, N.2002. Finite-state approaches to web information extraction. Proc. of 3rd Summer
Convention on Information Extraction.
[7] Thorsten JoachimsMandarHaridas and DoinaCaragea: Exploring Wikipedia and DMoz ,2009.
Knowledge Bases for Engineering a User Interests Hierarchy for Social Network Applications, USA.
[8] Jeffrey Heer,Stuart K. Card and James A. Landay: Prefuse:2004 A toolkit for interactive information
visualization
[9] Jeffrey Heer and Stuart K. Card: Efficient User Interest Estimation in Fisheye Views, ACM Human
Factors in Computing Systems (2003)
[10] Graphviz. http://www.research.att.com/sw/tools/graphviz/
[11] A. F. R. Rahman, H. Alam and R. Hartono. “Understanding the Flow of Content in Summarizing
HTML Documents”. In Int. Workshop on Document Layout Interpretation and its Applications,
DLIA01, Sep., 2001.
[12] Marko Grobelnik, Dunja Mladenić,2005.”Simple classification into large topic ontology of Web
documents” In Journal of Computing and Information Technology

ONTOLOGICAL TREE GENERATION FOR ENHANCED INFORMATION RETRIEVAL

More Related Content

What's hot (18)

Similar to ONTOLOGICAL TREE GENERATION FOR ENHANCED INFORMATION RETRIEVAL (20)

Recently uploaded (20)

ONTOLOGICAL TREE GENERATION FOR ENHANCED INFORMATION RETRIEVAL