=Paper=
{{Paper
|id=Vol-1180/CLEF2014wn-eHealth-DrameEt2014
|storemode=property
|title=Query Expansion using External Resources for Improving Information Retrieval in the Biomedical Domain
|pdfUrl=https://ceur-ws.org/Vol-1180/CLEF2014wn-eHealth-DrameEt2014.pdf
|volume=Vol-1180
|dblpUrl=https://dblp.org/rec/conf/clef/DrameMD14
}}
==Query Expansion using External Resources for Improving Information Retrieval in the Biomedical Domain==
https://ceur-ws.org/Vol-1180/CLEF2014wn-eHealth-DrameEt2014.pdf
Query expansion using external resources for improving
information retrieval in the biomedical domain
Khadim Dramé, Fleur Mougin, Gayo Diallo
ERIAS, INSERM U897, ISPED, University of Bordeaux
146 rue Leo Saignat 33076, Bordeaux
firstname.lastname@u-bordeaux.fr
Abstract. This paper presents the first participation of the ERIAS team in task
3 of the ShARe/CLEF eHealth Evaluation Lab 2014. The goal of this task is to
evaluate the effectiveness of Information Retrieval systems to support patients
in accessing easily relevant information. We propose a method which exploits
external resources for improving information retrieval in the biomedical do-
main. The proposed approach is based on the well-known Vector Space Model
and it uses two extensions of this model to enhance its performance. Specifical-
ly, the MeSH thesaurus is used for query expansion with different configura-
tions. Experiments on a large collection of documents have shown that the use
of these external resources can improve performance in medical in-formation
retrieval.
Keywords: information retrieval, Lucene, Vector Space Model, n-gram extrac-
tion, query expansion, MeSH thesaurus.
1 Introduction
The role of an Information Retrieval (IR) system is to support users to access rele-
vant information corresponding to their needs. In the medical domain, accessing use-
ful information becomes increasingly important with the growing amount of available
information. To tackle this challenging issue, different approaches have been pro-
posed raising the challenge of assessing their performance. The ShARe/CLEF (Cross-
Language Evaluation Forum) eHealth Lab [1] is an evaluation campaign in the bio-
medical domain which aims at easing patients (and their relatives) to understand their
health-related information. Especially, the goal of the third task is to develop methods
which facilitate the access to valuable information to patients regarding their
health[2]. Indeed, the amount of biomedical information is growing rapidly with an
abundant production of digital collection of documents. Accessing to useful infor-
mation among this large amount of available data becomes a real challenge. To do so,
controlled vocabularies, such as the Medical Subject Heading (MeSH) thesaurus, are
widely used to improve the medical information retrieval (IR). In [3], the authors
proposed the use of the MeSH thesaurus for expanding user queries. Terms associated
189
with a MeSH descriptor are considered as synonyms and used therefore to expand
queries. They show, through their experiments, that queries expansion using MeSH
thesaurus improves significantly the performances of IR in the medical domain. A
similar queries expansion approach using the MeSH thesaurus is proposed in [4]. The
hierarchical structure of the MeSH descriptors was exploited to improve the retrieval
effectiveness. The popular MetaMap [5] tool is used to identify descriptors which are
then expanded with their children. They show that queries expansion using the MeSH
hierarchy yield significant improvements.
In this paper, we report the results achieved during our first participation in the
Task 3 of ShARe/CLEF eHealth Evaluation Lab 2014. We have investigated the ex-
ploitation of controlled vocabularies and especially the MeSH (Medical Subject
Heading) thesaurus in order to improve IR performance. First, we have explored the
use of terms’ synonyms to expand queries which contain vocabulary’s entries. We
have then investigated the use of semantic relations in MeSH for query expansion and
their combination with synonyms’ expansion. Evaluation on a large collection of over
one million documents shows the use of these re-sources may improve performances
in biomedical IR.
The remainder of the paper is organized as follows. First, we present our IR ap-
proach in Section 2. Then, experiments and results are described in Section 3 and
discussed in Section 4. We conclude and give perspectives to this work in Section 5.
2 Methods
In this section, we describe the different techniques that we have used to address
information access issue.
The proposed approach is based on the popular Vectorial Space Model (VSM) [6].
The principle of the VSM is to represent documents (and queries) by term vectors and
then to use the cosine similarity measure to determine relevant documents according
to a given query. In this work, instead of words, all unigrams and bigrams identified
within documents are used as descriptors to index them. In a preprocessing phase,
document contents are first tokenized and stop words are removed. Then, from these
preprocessed documents, all the unigrams and bigrams are extracted. The latter, with
their associated weights (term frequency – inverse document frequency), are used for
indexing the documents. Our baseline system is based on this model using only the
query title as the original query and does not use any external resources (Run 1). In-
deed, each query can include a title, a description and additional fields.
Secondly, we have explored the usefulness of controlled vocabularies to enhance
IR. Indeed, controlled vocabularies like MeSH thesaurus are widely used in medical
IR [3][4][7]. However, since the use of solely vocabulary’s terms for describing the
queries was proven insufficient for this task last year [8], we additionally exploited
the traditional keywords constituted by simple words. Thus, the n-grams model (our
baseline) is combined with MeSH terms for query expansion. First, original queries
190
are extended using its synonymous terms, descendants (using hierarchical relations)
and related MeSH terms (with the exploitation of the See Also relations in MeSH).
For query terms not contained in the MeSH thesaurus, we have taken into account
their synonyms retrieved in the Unified Medical Language System (UMLS) [7] Me-
tathesaurus®. This corresponds to Run 5 of our method. In Run 6, we have investi-
gated query expansion with only synonyms from the MeSH thesaurus or the UMLS,
like in the previous Run 5. The difference is, in Run 6, only term synonyms are used
to expand the queries while Run 5 combines term synonyms with related terms for
query expansion. Finally, we have exploited query descriptions to estimate how they
can improve the results. This last Run (7) is similar to the Run 6 but takes into ac-
count both title and query descriptions to build queries.
For identifying medical terms in queries, we have developed a simple method
which focuses on the most specific terms; we consider terms that are entries of the
vocabulary and that are not included in a longer entry in the query. So, for each query,
only its medical terms not contained in other ones of the query are extracted in order
to extend the query. After that, the original query is expanded by either terms’ syno-
nyms or other related terms or both. For example, for the query “Anoxic brain injury”
(qtest2014.4), only the most specific term “Anoxic brain injury” (thus ignoring the
more general terms brain, brain injury, injury) is extracted and its synonyms, such as
“Anoxic Encephalopathy” and “Anoxic Brain Damage”. The query “stroke and res-
piratory failure” (qtest2014.16) is constituted by the two specific terms “stroke” and
“respiratory failure”. The expanded query then includes synonyms like “cerebrovas-
cular accident”, “vascular accident, brain”, “kidney failure”, and “renal failure”. For
the query “aspiration pneumonia due to misplacement of dobhoff tube”
(qtest2014.11), the terms “aspiration pneumonia” and “dobhoff tube” are extracted
and the original query is extended by these terms’ synonyms and/or related terms,
such as “Acid Aspiration Syndrome” and “Respiratory Aspiration”.
3 Experiments and Results
A document collection over one million documents (web pages from medical web
sites) and 50 potential patient queries were provided by the organizers of
ShARe/CLEF eHealth Evaluation Lab [2]. Beforehand, five queries with their corres-
ponding relevance judgments were provided for the training. Each query includes a
title, a description and other additional fields. The Lucene open-source search engine
was used in our experiments for indexing and retrieving documents. For each query,
we have used our different configurations to retrieve the top 1000 relevant documents
The P@10 and NDCG@10 measures were primarily used to evaluate the submitted
systems. Thus, for each system, the top ten documents retrieved were considered in
the evaluation.
The results of our different runs are presented in Table 1. According to the evalua-
tion measures, all the runs using external resources got better results than the baseline.
191
Therefore, using the controlled vocabularies improves the retrieval performance as
expected. We also note that Run 6 which uses only query title as original query and
terms’ synonyms for query expansion obtained the best performance regarding P@10
and NDCG@10.
Compared with the global results in this task, our runs got reserved results accord-
ing to queries. For some queries, they yielded good and even the best results. For
example, our Run 5 got the best performance for the queries qtest2014.16 (P@10 =1),
qtest2014.44 (P@10 = 0.9), qtest2014.47 (P@10 = 1) and greatly exceeded the medi-
an (+0.4, +0.4, +0.5 respectively). For these queries, almost all terms are entries of the
MeSH thesaurus. In some cases, they just surpassed the median or got comparable
performance. For example, for the queries qtest2014.7 (+0.3), qtest2014.11 (+0.3),
Run 1 slightly surpassed the median and got the same P@10 as it for the queries
qtest2014.3, qtest2014.12 and qtest2014.18. For other queries, our method got poorer
performance. For example, for the queries qtest2014.4, all our runs yielded lower
performance than the median: -0.6 for Run 5 and Run 6 and -0.7 for Run 7; for Run 1,
the difference is less important (-0.1). This query is an entry of the vocabulary. So it is
extended by synonyms and related terms of the term “Anoxic brain injury”. Thus,
documents containing the more general term “brain injury”, which are considered as
relevant, are ranked after the ones containing query term synonyms or related terms.
The lowest P@10 compared with the median (-0.9) was yielded by Run 1 for query
qtest2014.6. Here, the query is an entry of the vocabulary and therefore using key-
words search, where the query is split into words, decreases the performances.
All our runs achieved a P@10 greater than or equal to 0.5 for half of the fifty que-
ries. For five queries, Run 1 achieved a P@10 of 1 while Run 5 and Run 6 achieved
this maximal precision in only four and two queries, respectively. But for five other
queries, Run 1 had a P@10 equal to 0 while both Run 5 and Run 6 achieved this bad
score for three queries. For queries qtest2014.37 and qtest2014.49, the top ten re-
trieved documents are irrelevant (P@10 = 0) for all our runs. The query qtest2014.37
is expressed with misspellings (“gynecolocigal” instead of “gynecological”).
Table 1: Performance of our different systems in Task 3 of CLEF eHealth 2014
Runs P@5 P@10 NDCG@5 NDCG@10
Run 1 0.5040 0.5080 0.4955 0.5023
Run 5 0.5440 0.5280 0.5470 0.5376
Run 6 0.5720 0.5460 0.5702 0.5574
Run 7 0.5960 0.5320 0.5905 0.5556
192
4 Discussion
The results presented in Table 1 show globally that query expansion using only
terms’ synonyms from MeSH improves the IR performance in our experiments (Run
6). Although the combination of synonyms and semantic relations for expanding
queries enhance the results compared with the baseline, they did not increase the per-
formance as expected. On the contrary, the use of related terms decreases the retrieval
performance. In addition, while it got the best performance according to P@5 and
NDCG@5, Run 7, which uses query description, provided worse results than Run 6
according to P@10 and NCDG@10. In other words, the use of query descriptions
decreases the system performance if the top ten documents are considered but remains
useful if we focus on the top 5 retrieved documents.
For several queries, Run 1 greatly exceeded the other runs even if it had lower per-
formance overall: qtest2014.4, qtest2014.7, qtest2014.13, qtest2014.18, qtest2014.39
are some examples. The performance drop of the query expansion methods is general-
ly caused by the negligence of some relevant terms not contained in the vocabulary
(for example, “secondary” in “surviving rates after a secondary myocardial infarctus”)
or the decomposition of the query in simple terms (“dizziness” and “hypotension” in
“dizziness and hypotension”) which does not take into account the context. On the
other hand, the results of Run 5 and Run 6 similarly vary but Run 6 often got a slight-
ly better performance. This may be due to noise resulting from the use of related
terms. For example the, from the query qtest2014.26, the term “gastrointestinal bleed”
is extracted and expanded by its synonym “Gastrointestinal Hemorrhage” in Run 6
and additionally its related terms like “Peptic Ulcer Hemorrhage”, “Endoscopic
Hemostasis” and so on in Run 5. Consequently, the Run 6 got a higher P@10 of 0.9
while Run 5 yielded a P@10 of 0.1 which is even lower than the baseline result
(P@10=0.4).
5 Conclusion and future work
In this work, we have investigated the usefulness of controlled vocabularies during
query extension. We have proposed a vector space based method with several query
expansion techniques and evaluated their impact in information retrieval effectiven-
ess. Overall, using these controlled vocabularies improves the retrieval performance
our experiments. However, the combination of terms’ synonyms and semantic relati-
ons for query expansion surprisingly yield worse results than using only synonyms.
On the other hand, although our query expansion methods outperform our baseline,
our results need to be improved compared with the global results of Task 3 partici-
pants. Overall, for our first participation to CLEF eHealth Lab, our method got promi-
sing results which we would like to improve.
193
In the future, to increase the performance of our retrieval method, we plan to ex-
plore other advanced retrieval models like the BM25 [9], which is used as the baseline
in this challenge and its combination with controlled vocabularies. We also believe
that the use of semantic similarity for query expansion may help to improve the in-
formation retrieval performance.
References
1. Goeuriot, L., Kelly, L., Li, W., Palotti, J., Pecina, P., Zuccon, G., Hanbury, A.,
Jones, G., Mueller, H.: ShARe/CLEF eHealth Evaluation Lab 2014, Task 3: User-
centred health information retrieval. Proceedings of CLEF 2014 (2014).
2. Kelly, L., Goeuriot, L., Suominen, H., Schrek, T., Leroy, G., Mowery, D.L., Ve-
lupillai, S., Chapman, W.W., Martinez, D., Zuccon, G., Palotti, J.: Overview of the
ShARe/CLEF eHealth Evaluation Lab 2014. Proceedings of CLEF 2014. Springer
(2014).
3. Díaz-Galiano, M.C., Martín-Valdivia, M.T., Ureña-López, L.A.: Query expansion
with a medical ontology to improve a multimodal information retrieval system.
Comput. Biol. Med. 39, 396–403 (2009).
4. Azcárate, M.C., Vázquez, J.M., López, M.M.: Improving image retrieval effec-
tiveness via query expansion using MeSH hierarchical structure. J. Am. Med. In-
form. Assoc. amiajnl–2012–000943 (2012).
5. Aronson, A.R.: Effective Mapping of Biomedical Text to the UMLS Metathesau-
rus: The MetaMap Program. (2001).
6. Salton, G., Wong, A., Yang, C.S.: A Vector Space Model for Automatic Indexing.
Commun ACM. 18, 613–620 (1975).
7. Jimeno-Yepes, A.J., Plaza, L., Mork, J.G., Aronson, A.R., Díaz, A.: MeSH index-
ing based on automatically generated summaries. BMC Bioinformatics. 14, 208
(2013).
8. Suominen, H., Salanterä, S., Velupillai, S., Chapman, W.W., Savova, G.K., El-
hadad, N., Pradhan, S., South, B.R., Mowery, D., Jones, G.J.F., Leveling, J.,
Kelly, L., Goeuriot, L., Martínez, D., Zuccon, G.: Overview of the ShARe/CLEF
eHealth Evaluation Lab 2013. In: Forner, P., Müller, H., Paredes, R., Rosso, P.,
and Stein, B. (eds.) Information Access Evaluation. Multilinguality, Multimodali-
ty, and Visualization - 4th International Conference of the CLEF Initiative, CLEF
2013, Valencia, Spain, September 23-26, 2013. Proceedings. pp. 212–231. Spring-
er (2013).
9. Robertson, S.E., Jones, K.S.: Simple, Proven Approaches to Text Retrieval.
(1997).
194