=Paper=
{{Paper
|id=Vol-1178/CLEF2012wn-CLEFeHealth-MoenEt2012
|storemode=property
|title=Towards Retrieving and Ranking Clinical Recommendations with Cross-Lingual Random Indexing
|pdfUrl=https://ceur-ws.org/Vol-1178/CLEF2012wn-CLEFeHealth-MoenEt2012.pdf
|volume=Vol-1178
|dblpUrl=https://dblp.org/rec/conf/clef/MoenM12
}}
==Towards Retrieving and Ranking Clinical Recommendations with Cross-Lingual Random Indexing==
https://ceur-ws.org/Vol-1178/CLEF2012wn-CLEFeHealth-MoenEt2012.pdf
Towards Retrieving and Ranking Clinical
Recommendations with Cross-Lingual Random Indexing
Hans Moen, Erwin Marsi
Norwegian University of Science and Technology,
Department of Computer and Information Science,
7491 Trondheim, Norway
hans.moen@idi.ntnu.no, emarsi@idi.ntnu.no
Abstract. Clinicians have to deal with large amounts of textual data, searching
for and navigating information that satisfies their informational needs. Clinical
notes and Clinical Practice Guidelines (CPG) are textual resources that usually
contain free text. As language use in the medical domain is rather specialized,
generic information retrieval tools are suboptimal for such data. This calls for
specialized information retrieval systems that incorporate domain-specific
knowledge. Furthermore, information resources for relatively small languages
are limited, so physicians speaking those languages often have to resort to in-
formation sources in English or one of the other major languages spoken. In our
setting, given a patient record written in Norwegian, a physician is looking for
recommendations in CPGs written in English. The recommendations are to be
retrieved and ranked according to their relevance to (specified parts of) the pa-
tient’s medical history and possibly a user-supplied query. This means that the
query language and the resource language differ, requiring some form of Cross-
Lingual Information Retrieval. Most approaches rely on translating the input
query to the target language using bilingual dictionaries or machine translation
systems. This raises the familiar problems in machine translation such as lack of
lexical coverage and lexical translation ambiguity. We propose an alternative
method that avoids direct translation of queries by implicit encoding of cross-
lingual semantic relations in a vector space model. We modify the Random In-
dexing model in such way that distributional similarity between terms or docu-
ments can be measured across different languages. The key idea is to assign
identical index vectors to source language terms and their target language trans-
lations during construction of the models. This requires only a translation dic-
tionary and large source/target language text corpora, which do not need to be
aligned. The dictionary does not need to have full coverage as long as it in-
cludes the most frequent words. Additional target languages can be added with-
out having to re-train the existing models. So far we have implemented this
Cross-Lingual Random Indexing model and conducted some pilot experiments
on retrieval of semantically related terms in Norwegian and English, with en-
couraging results. Our contribution furthermore discusses the data required for
further training and evaluating such models, as well as the need for expert
knowledge in data labeling.
Keywords: Cross-Lingual; Health Data; Information Retrieval
1 Introduction
As the health sector is growing, clinicians are confronted with an ever-growing
amount of information stored in electronic format. In many countries, most of this
information is free text rather than structured data, and a large part of it is stored in
electronic health record (EHR) systems. The total amount of data stored in such sys-
tems throughout hospitals is large and continuously growing. As a result, clinicians
have to deal with large amounts of textual data, searching for information that satis-
fies their informational needs. As language use in the medical domain is rather spe-
cialized, generic information retrieval tools are suboptimal for the task. This calls for
specialized information retrieval systems that incorporate domain-specific knowledge.
In addition, resources for relatively small languages are limited so that physicians
speaking and writing in those languages often have to resort to information sources in
English or one of the other major languages spoken. This means that the query lan-
guage and the resource language often differ, requiring some form of cross-lingual
information retrieval (CLIR).1
Within this general context, our goal is to aid physicians in their daily work of
searching for and navigating relevant clinical practice guidelines (CPGs) and underly-
ing recommendations. The point of departure is a patient record written in Norwegian,
possibly combined with a free text query. A physician is looking for recommenda-
tions in CPGs written in English. The recommendations are to be retrieved and ranked
according to their relevance to (specified parts of) the patient’s medical history. This
abstract describes a possible approach to solving this problem using semantic lan-
guage models based on Random Indexing for statistical analysis of large amounts of
text. As this is work in progress, we focus on a description and discussion of the ap-
proach and required data, leaving experimental results to future work.
2 Approach
CLIR aims at identifying relevant documents in a language other than that of the que-
ry.1 Most approaches rely on translating the query to the target language using bilin-
gual dictionaries or machine translation systems. This raises the familiar problems in
machine translation such as lack of lexical coverage and lexical translation ambiguity.
We propose an alternative method that avoids direct translation of queries by implicit
encoding of cross-lingual semantic relations in a vector space model. This is motivat-
ed by the cross-lingual distributional similarity hypothesis:2 If in a text of one lan-
guage two words A and B co-occur more often than expected by chance, then in a text
of another language those words that are translations of A and B should also co-occur
more frequently than expected. For example, suppose a query-term in Norwegian
frequently co-occurs with certain neighboring terms as observed in a Norwegian text
corpus, then English terms that frequently co-occur with English translations of the
same Norwegian neighboring terms (according to a bilingual dictionary) in an English
text corpus are likely to have a high contextual similarity to the Norwegian query
term. Notice that we do not make the (stronger) assumption that the Norwegian query
term and the cross-lingually co-occurring English terms are proper translations of
each other, but merely that the two are semantically related, and that this semantic
relation can be exploited in CLIR.
In order to exploit cross-lingual distributional similarity for CLIR, we use the Ran-
dom Indexing (RI) model for distributional similarity.3 RI incrementally builds a vec-
tor space model of information units (terms in this case) in the following way. First, a
unique index vector (a sparse high-dimensional vector consisting of mainly zeros with
a couple of randomly chosen +1/-1 components) is assigned to each entry in the trans-
lation dictionary; the associated translations receive the same index vector. A term in
the source language will thus have the same index vector as its translations in the
target language. Next, a context vector is created for all terms in either language. This
is accomplished by sliding a context window over the text corpus. For each term in
the center of the window, the index vectors of the neighboring words within the win-
dow are added to its context vector. A term’s context vector therefore encodes the
terms it co-occurs with. Context vectors for larger text spans such as sentences, sec-
tions or documents are subsequently created by summing, and possible weighting (cf.
TF*IDF), the context vectors of all terms contained. Structured domain knowledge
can here be applied to further adjust the weightings of these vectors. This procedure
assumes only a translation dictionary and large source/target language text corpora
that does not need to be aligned in terms of content. The dictionary does not need to
have full coverage as long as it includes a substantial part of the most frequent words,
and words that are not in the dictionary will still be assigned context vectors. Addi-
tional target languages can be added in a similar way without having to re-train the
existing models. For retrieval, a query is converted to a query vector by summing the
context vector of all query terms in the source language. The query vector is then
directly matched against pre-computed document vectors in the target language(s)
using cosine similarity, resulting in a ranking of the most similar documents.
3 Discussion
So far we have implemented the outlined approach and conducted some pilot experi-
ments on retrieval of semantically related terms in Norwegian and English, with en-
couraging results. Our short-term goals are to evaluate on the CLEF data for CLIR,
followed by tests on real health data and for the specific task of searching for and
ranking recommendations in existing CPGs automatically. The later requires three
types of data for training:
1. a set of health records in the source language;
2. a set of CPGs written in the target language, containing recommendations relevant
w.r.t. the health records (preferably structured so that individual recommendations
are clearly separated, e.g. similar to the structure of the AT9 guidelines);
3. a bilingual dictionary covering the most frequent terms, preferably including do-
main-specific terminology.
Evaluation requires a gold standard consisting of pairs of health records, or a subset of
clinical notes from these, and an associated set of ranked recommendations originat-
ing from the CPGs. Since a health record contains notes with different timestamps
and written by physicians with different roles, it might be necessary to be selective
when it comes to which clinical notes to use as basis for the query. Due to the sensi-
tive content in health records, such data is difficult to acquire. Nor is it an easy task to
get access to the necessary expert knowledge required to develop such a gold stand-
ard. Development of these evaluation data for Norwegian-English is currently un-
der way.
4 Conclusion
There is a potential in aiding clinical work through retrieving and ranking CPGs using
CLIR. The method described here is a possible way of accomplishing such a task.
However, there is a need for establishing a suitable dataset for training and validating
this and similar approaches.
Acknowledgements
This work was partly funded by the EviCare project (http://www.evicare.no) and
by the European Community’s Seventh Framework Programme (FP7/2007–2013) under grant
agreement nr 248307 (PRESEMT).
References
1. Kishida K. Technical issues of cross-language information retrieval: a review. Information
Processing & Management. 2005;41(3):433-455.
2. Rapp R. Identifying word translations in non-parallel texts. Association for Computational
Linguistics: Proceedings of the 33rd annual meeting on Association for Computational
Linguistics. 1995:320-322.
3. Sahlgren M. An introduction to random indexing. Methods and Applications of Semantic
Indexing Workshop at the 7th International Conference on Terminology and Knowledge
Engineering, TKE 2005. 2005:1-9.