Article Preview
TopIntroduction
The huge amount of data available online today is often composed of plain text field, for instances, clinical trial descriptions, adverse event reports or electronic health records. These texts often contain the real language (expressions and terms) used by the community. Although in the biomedical domain there exist hundred of terminologies and ontologies to describe such languages (Noy et al., 2009), those terminologies often miss concepts or possible alternative terms for those concepts. Our motivation is to improve the precision of automatic terms extraction process, the main reason for this, is that language evolves faster than our ability to formalize and catalog it. This is even more true for French in which the number of terms formalized in terminologies is significantly less important than in English.
NLP (natural language processing) tools and methods enable to enrich biomedical dictionaries from texts. Automatic Term Recognition (ATR) is an approach in language technology that involves the extraction of technical terms from domain-specific language corpora (Zhang et al., 2008). In addition, Automatic Keyword Extraction (AKE) is the process of extracting the most relevant words or phrases in a document. Keywords, which we define as a sequence of one or more words, provide a compact representation of a document’s content. Two popular AKE measures are Okapi BM25 and TFIDF, also called weighting measures. These two fields are summarized in Table 1.
Table 1. Differences between ATR and AKE
| Automatic Term Recognition (ATR) | Automatic Keyword Extraction (AKE) |
Input | one large corpus (i.e., not explicitly separated in documents) | single document within a dataset of documents |
Output | technical terms of a domain | keywords that describe the document |
Domain | very specific | none |
Exemples | C-value | TFIDF, Okapi |
In our work, we adopt as baseline measures an ATR method, C-value (Frantzi et al., 2000), and the best two AKE methods (Hussey et al., 2012). Indeed, the C-value, compared to other ATR methods, often gets best precision results and especially in biomedical studies (Knoth et al., 2009; Zhang et al., 2008; Zhang et al., 2004). Moreover, this measure is defined for multi-word term extraction but can be easily adapted for single-word term (presented later on) and it has never been applied to French text, which is appealing in our case. Okapi and TFIDF are the best AKE methods (Hussey et al., 2012). We propose to define new extraction methods by combining in different manners ATR and AKE measures, in order to rank the best candidate terms. Our experiment results underline the precision efficiency gain with the proposed methods. We give priority to precision in order to focus on extraction of new valid terms (precision) rather than on missed terms (recall), i.e., for a candidate term to be a valid biomedical term or not.