Acronyms and abbreviations are widely used in clinical and other specialized texts. Understanding their meaning constitutes an important problem in the automatic extraction and mining of information from text. Moreover, an even harder problem is its sense disambiguation; that is, where a single acronym refers to many di erent meanings in different texts, a common occurrence in the clinical texts. In such cases, it is necessary to identify the correct corresponding sense for the acronym or abbreviation, which is often not directly speci ed in the text. Here we present an approach to identify acronyms and abbreviations for the BARR2 competition. We use cTAKES [7] as a framework to develop an approach to identify abbreviations and acronyms as part of a lookup entity recognition system and a word sense disambiguation classi er. The results of the BARR2 test set have shown a 79.13 F measure.
The problem of retrieving the meaning of acronyms and abbreviations in clinical text is related to that of entity identi cation, since it is necessary to know which entity an acronym or abbreviation expression (also called short forms) refers to in a text in order to accurately identify and extract target information.
The problem of automatically determining the meaning of short forms
in medical texts is both a critical as well as a di cult one. It is critical
because the performance of information retrieval and extraction tasks is
signi cantly degraded when acronym and abbreviation meanings are not
properly understood or interpreted. The problem is exacerbated in the
medical literature by the widespread use and frequent coinage of novel
short forms and new short form meanings. Furthermore, there is wide
variance in conventions within the medical communities on forming acronyms
from their "long forms". Acronyms and abbreviations were addressed as
a problem to solve in the biomedical domain a long time ago (e.g. [
Clinical terms may be noisy descriptions typed by healthcare professionals in the electronic health record system (EHR). Description terms contain clinical ndings, suspected diseases, among other categories of concepts. Descriptions are very short texts presenting high lexical variability containing synonymy, acronyms, abbreviations and typographical errors. Automatic mapping of description terms to normalized descriptions in an interface terminology is a hard task and it is based essentially on string similarity features. In this scenario, abbreviations and acronyms pose a special challenge for several reasons. The Joint Commission International1 requires that the use abbreviations must be controlled on patient materials and documents to ensure that patients and their families understand the information available in their records 2. Also, according to the SNOMED CT Editorial Guide, abbreviations are prohibited in fully speci ed names and synonyms, with speci ed exceptions.
The organization of BARR and BARR2 initiates a special e ort on this topic for Spanish language. Even if there are some compiled resources there are no public available databases for the clinical domain in Spanish language. There is great variability in the use of abbreviations and acronyms and many of them present high degree of ambiguity.
The problem of sense disambiguation is a crucial one in an information retrieval system. A common acronym such as AA has many di erent meanings, such as:3 { abdomen agudo { alcoholicos anonimos { amenaza de aborto { aminoacido { anemia aplasica { aorta abdominal { aorta ascendente { apendicitis aguda
The Hospital Italiano de Buenos Aires (HIBA) has an interface
Spanish vocabulary [
The Second Biomedical Abbreviation Recognition and Resolution (BARR2) track has the aim to promote the development and evaluation of clinical abbreviation identi cation systems. There are two sub-tracks and we chose (due to time limitations and scope focus) to participate in the Subtrack 2, the abbreviation resolution track. In this case the challenge is to identify the acronyms and abbreviations in the text and to provide the corresponding de nition or long form.
The BARR2 organization provided a training set consisting of 318 clinical cases that had been published in the clinical literature and the corresponding metadata for original record and the corresponding journal and publication date. A development set consisting of 146 clinical cases was also released, and nally for the challenge participating teams had to submit their predictions for the background set composed of 2879 clinical cases. The test set consisting of 220 clinical cases was released after the participating groups submitted their predictions for the background set. 3
We decided to use cTAKES as a framework to test acronym and
abbreviation resolution algorithms. The BARR2 source text has not been
previously tokenized, so di erent tokenization algorithms have an impact
on the system performance. We used cTAKES pipeline facility to test
different parameters. We slightly adapted cTAKES sentensi er, tokenizer,
and we used the universal POS tagger [
An initial assessment has shown that it was very di cult to add the data from Diccionario de Siglas Medicas. In particular, it was not easy to normalize those expansions that had the same meaning but had di erent long forms, i.e. synonym long forms. Therefore we decided to use only those data for which we had training sets for the classi er, HIBA abbreviations, and BARR2 training set. We split the data in training and test to Evaluate the classi er. Table 3 shows the data used.
Data HIBA BARR2 Total Training 222225 2558 224783 Test 79117 883 80000 Total 301342 3441 304783
Ambiguous 368 99 522
Table 1. Data Sets used for training the classi er
We performed a few tests on the training data and we found that it was very di cult to predict correct expansions. In particular there were some cases of spurious ambiguity: { virus de epstein-barr vs virus de epstein barr { tomograf a axial computadorizada vs tomograf a axial computarizada vs tomograf a axial computada
Therefore we took a very simple approach. We selected the most frequent expansions for those abbreviations in the BARR2 training set, in other words there were no ambiguous short forms from the training set. Those expansions that were not predicted by the classi er were discarded. This is what we called the fq3 model and we obtained a good baseline result using this simple strategy. We used also used the same data combined with the HIBA abbreviations and acronyms, and in this case we did not use any ltering. This is our fq3-HIBA model. Finally we used the same strategy using both the training and the development set. Our fq4 and fq4-HIBA models. 4
BARR2 organization provided the evaluation tool to be used in the training and development sets. The tool provides three measures, Ultra-strict, Strict and Flexible Evaluations. Ultra-strict evaluation requires that the exact same expansion string be predicted. Strict evaluation does not consider stop words nor word order, a list of stopwords was provided and uses lemmatized forms. The Flexible evaluation used a stemmer to compare predictions. We undestand that the Strict evaluation provides a closer comparison, given stopwords at the expansion usually do not provide semantic information, and lemmatized forms preserve meaning. Unfortunately we did not have a lemmatizer ready in the pipeline so we did not use lemmatized forms.
This work was prepared in a very short time, and the approach we used was very simple. It was apparent from the very beginning that the BARR2 and the HIBA data were very di erent, and it is re ected in the performance when HIBA data is used. One of the di culties we faced is the need of using a normalization function, based in string similarity, so as to map to a canonical string. We did not have time also to include a lemmatizer in the pipeline, which might improve a little the results.
Our strategy produced better Precision than Recall results, this can be seen as an e ect both of the preprocessing pipeline we used, and also on the ltering use of the classi er.