=Paper=
{{Paper
|id=Vol-1168/CLEF2002wn-iCLEF-LlopisEt2002
|storemode=property
|title=iCLEF at Universities of Alicante and Jaen
|pdfUrl=https://ceur-ws.org/Vol-1168/CLEF2002wn-iCLEF-LlopisEt2002.pdf
|volume=Vol-1168
|dblpUrl=https://dblp.org/rec/conf/clef/LlopisFGDS02
}}
==iCLEF at Universities of Alicante and Jaen==
https://ceur-ws.org/Vol-1168/CLEF2002wn-iCLEF-LlopisEt2002.pdf
iCLEF at Universities of Alicante and Jaen
Fernando Llopis, Antonio Ferrández, y Jose L. Vicedo Manuel C. Diaz y Fernando Martínez
Departamento de Lenguajes y Sistemas Informáticos Departamento de Ciencias de la Computación
University of Alicante University of Jaen
Alicante, Spain Jaen , Spain
{llopis, antonio, vicedo}@dlsi.ua.es {mcdiaz, dofer} @ ujaen.es
Abstract
In this paper, the obtained results on iCLEF-2002 are presented. This is the first time that we try to face up the
iCLEF task, and we have used a Passage Retrieval approach, specifically our previously developed system called
IR-n. This system previously divides the document in fragments or passages, and after that, the similarity of each
passage with the query is measured. Finally, the document that contains the most similar passage is returned as
the most relevant document. In the interactive document selection task, we have experimented with this system
in the following way, the system only shows the most relevant passage of each document instead of the entire
document. In this work, the results obtained with this system are presented, where the relevant passages
presented to the user have been automatically translated into Spanish by means of Systran. Moreover, the results
are compared with Z-Prise system, which is based on the reading of the entire document.
1. Introduction
The focus of this paper is the interactive document selection task. The main objective of this task is to design a
system to facilitate users to find relevant documents about their information needs. The classical Information
Retrieval (IR) systems use the whole document in order to determine the relevance of the document with
reference to a query. The main problem of this kind of systems is that they can return entire relevant documents,
but they cannot locate the most relevant piece of text in the document. For example, a document about the
“Biography of Felipe II” is relevant for the query “the town were Felipe II was born”, but only a part of this
document is relevant for the information required. In this way, when a user has to determine if a document is
relevant or not, he/she has to probably read the entire document.
A new IR proposal that tries to overcome this problem is called Passage Retrieval (PR). The PR systems divide
the document into pieces of text that are called passages. After that, the similarity measure is obtained for each
passage, and finally, the document will obtain the similarity value of its most relevant passage.
The IR system used in this paper, called IR-n, employs the PR strategy too. The IR-n system has been used as IR
system in CLEF 2001 [5], and as a module in a Question Answering (QA) system in TREC-10 [8], where it
reduces the amount of text in which the QA system works.
In this paper, the results obtained with the IR-n system for iCLEF task are presented, that is to say, when the user
determines if a document is relevant or not by means of reading only the most relevant passages returned by this
system. These relevant passages are automatically translated into Spanish by Systran. This paper is structured in
the following way. Firstly, an introduction of PR systems is presented. Secondly, the architecture of IR-n system
and the experiments for tuning the IR-n system for the document selection task are described. Thirdly, the results
are explained and compared with those obtained with the Z-Prise system, which is based on the IR approach that
uses the entire document to determine the relevance to a query. Finally, the conclusions obtained with this work
and future works are presented.
2. The state of the art in Passage Retrieval
Previous works [4] show that PR systems can improve the precision of IR systems between a 20 and 50%. The
PR systems can be classified according to the way that they define the passages in a document. A general
classification usually quoted by researchers is that proposed in [1], where the PR systems are divided into those
based on the discourse, those based on semantic models and those based on a window model. The first one uses
the structural properties of the documents, such as sentences, paragraphs or HTML marks in order to define the
passages. The second one divides each document in semantic pieces, according to the different topics in the
document. The last one uses windows with a fixed size to form the passages. Moreover, we can find another
taxonomy of window models in [4], where it is distinguished between those that use the structure of the
document in the moment to define the passages, and those that do not use this kind of information.
On the one hand, it looks coherent that discourse-based models are more effective since they are using the
structure of the document itself. However, the greater problem of them is that the results could depend on the
writing style of the document author. Moreover, this kind of models produces a very heterogeneous set of
passages, with reference to the size of each passage. On the other hand, window models has the main advantage
that they are simpler to accomplish, since the passages have a previously known size, whereas the remaining
models have to bear in mind the variable size of each passage. However, they have the problem that when the
passage starts in whatever word of the sentence, these passages could not be adequate in order to be presented to
the user as the most relevant passage, since they are not logic and coherent fragments of the document.
3. Architecture of the IR-n system
The IR-n system [5] is based on a window model that uses the structure of the document in the moment to define
the passages. The main characteristics of this system are the following:
1. A document is divided into passages, which are formed by a fixed number of sentences. This is because
a sentence usually represents an idea in the document, whereas the paragraphs can be used just for
giving a visual structure to the document. Moreover, the sentences are logic and complete units of
information, whereas those window models that start on whatever word in the document can return
incoherent fragments of text.
2. The number of sentences that form a passage can be separately determined for each set of documents.
Previous experiments for the documents of Los Angeles Times show that the best results are obtained
with passages of nine sentences.
3. The system uses windows with overlapped pieces of text in order to fine-tune the results. For example,
with passages of nine sentences, the first passage is formed by sentences from 1 to 9, the second one
from 2 to 10, and so on. We have used these overlapped passages because we have obtained better
results in the experiments presented in [6], than using other kinds of passages (e.g. those with no
overlapping, or with other degrees of overlapping). The overlapping process increments the running
time, but this increment is not very high, since the first passage starts in the first sentence where a key
word of the query appears, and the last passage in the last sentence where a key word appears.
4. We are using the cosine measure but with no normalization with reference to the size of the passage,
because the passages are quite homogeneous (the same number of sentences with a similar number of
words).
4. Experiments
Some experiments have been carried out with two main aims: firstly, with the aim of configuring the way of
presenting the documents to the user; secondly, with the aim of facilitating the reading process of the documents.
Firstly, an HTML interface was created as it is shown in Figure 1.
Figure 1. HTML interface for presenting the documents to the user.
In the HTML interface, information useful for iCLEF was collected, such as the number of the question and the
name of the user. In iCLEF’2002, the most relevant passage formed by nine sentences was presented to the user.
After the first experiments, several elements were added to the interface in order to facilitate the decision of the
user about the relevance of each document:
1. The visualization of the first line of the title of the document in capital letters.
2. The following sentence after the title is the previous sentence to the most relevant passage. This is
because in the IR-n system, the most relevant passage always starts in the first sentence where a key
word of the query appears. In this way, some required information in the query could appear close to
the first sentence of the passage. If we include this previous sentence, then the comprehension of the
passage is highly increased.
3. The sentences are shown in different lines in the interface. This fact facilitates the comprehension of the
passage, and it was quite easy to carry out since IR-n performs an indexation on sentences.
5. Results
The experiments were carried out by university students with a medium level of understanding of English
(although it is not important since the passages are automatically translated into Spanish). The IR-n system
results have been compared with those obtained with Z-Prise [9]. Only the first 25 most relevant documents have
been used for each query, which could explains the low recall results. Finally, we would like to notice that only
title and description of each query have been used. The obtained results have been presented in Table 1.
Average F_alpha
Z-Prise 0.2166
IR-n 0.3248
Table 1. The obtained results
Given that only 25 relevant documents were retrieved for each topic, it is interesting to study the precision
results. In Table 2, the precision results are presented for each topic, for both the Z-Prise and our system, IR-n.
Firstly, it is remarkable the low results obtained for one topic, because any relevant document was retrieved by
any system. Secondly, we should remark the high precision in two of the three remaining topics, which was
obtained with just the most relevant passage.
Average Precision IR-n BASE
Topic 1 0,4601 0,6371
Topic 2 0,8098 0,5925
Topic 3 0 0
Topic 4 0,7643 0,3748
Average 0,5085 0,4011
Table 2. The obtained results
6. Conclusions and future works
In this paper, we have described an experiment that study the ability of users to judge relevance of documents, in
which the users can only read the most relevant passages of these documents. The results have been quite good,
because the users take short time to judge the relevance since they have to read short pieces of text. However,
these short pieces of text contains the most relevant information about the information required, therefore the
precision results have been high, even higher than those obtained by means of reading the entire document.
Anyway, there are some points to notice, once the individual results and the opinions of the users have been
analysed:
• Firstly, the users find a great anxiety when they do not find the relevant document in the list of relevant
passages. This has occurred in one of the queries in which only one relevant document appeared in the
25 documents presented by our system. In this case, the users judged as relevant some non-relevant
documents that were not been selected in other cases.
• Moreover, the users find the automatic translations into Spanish quite unreadable most of the times
(more than it was expected).
• We think that the results have been influenced by presenting just the title and description of the queries,
which have supposed some doubts about the relevance of the passages.
• It has been difficult to find users to carry out the experiments, which explains the reduction of the
number of documents to study (only 25) for each query. This has highly decreased the recall of the IR-n
system, although we are quite happy with the obtained results, since the users have found a high
percentage of relevant documents in not much time (an average between 8 and 9 minutes per query).
This is because the piece of text that has to be read is only formed by nine sentences.
As future works, firstly, we pretend to improve the automatic translations. Systran has been used in order to
translate the passages presented to the user. Given that the automatic translation of the Los Angeles Times
collection was imperfect, and even it was sometimes unreadable, we will try to present the results to the user in a
more structured way. This task will be carried out by means of retrieving information from a collection similar to
the Los Angeles Times, specifically, the EFE news of the same year, which is available in Spanish.
Secondly, we have to improve the interactivity with the system. In the present work, we present to the user the
first 25 most relevant documents. However, we think that the results would have been improved if we have
expanded the query, according to the relevance that the user detects about these documents. In CLEF’2002, we
have already used this type of techniques of query expansion with a high improvement.
7. Acknowledgements
This work has been supported by the Spanish Government (CICYT) with grant TIC2000-0664-C02-02.
8. References
[1] Callan, J. Passage-Level Evidence in Document Retrieval. In Proceedings of the 17th Annual ACM SIGIR
Conference on Research and Development in Information Retrieval, Dublin, Ireland 1994, pp. 302-310.
[2] López-Ostenero, F., Gonzalo, J., Peñas, A. and Verdejo, F. Noun Phrase Translations for Cross-Language
Document Selection. Working Notes for the CLEF 2001. To appear in Lecture Notes in Computer Science.
Darmstdt, Germany 2001, pp 231-242.
[3] Jourlin, P., Johnson, S.E., Spärck Jones, K. and Woodland, P.C. General Query Expansion Techniques for
Spoken Document Retrieval. In Proceedings of the ESCA Workshop on Extracting Information from Spoken
Audio Cambridge, UK. 1999.
[4] KaszKiel, M. and Zobel, J. Effective Ranking with Arbitrary Passages. Journal of the American Society for
Information Science, Vol 52, No. 4, February 2001, pp 344-364.
[5] Llopis, F. and Vicedo, J. IR-n system, a passage retrieval system at CLEF 2001 Working Notes for the CLEF
2001. To appear in Lecture Notes in Computer Science. Darmstdt, Germany 2001, pp 115-120.
[6] Llopis, F., Ferrández, A. and Vicedo, J. Text Segmentation for efficient Information Retrieval Third
International Conference on Intelligent Text Processing and Computational Linguistics. Lecture Notes in
Computer Science. Mexico 2002.
[7] Wang, J and Oard, D. iCLEF 2001 at Maryland. Working Notes for the CLEF 2001. To appear in Lecture
Notes in Computer Science. Darmstdt, Germany 2001, pp 219-230
[8] Vicedo, J., Ferrandez, A. and Llopis, F. University of Alicante at TREC-10. In Tenth Text REtrieval
Conference, Gaithersburg, USA. 2001
[9] ZPrise developed by Darrin Dimmick (NIST) Available on demand at
http://www.itl.nist.gov/iaui/849.02/works/papers/zp2/zp2.html