Turney (2001) has shown that computing the mutual information of a pair of words by using cooccurrence counts obtained via queries to the AltaVista search engine performs very effectively in a synonym detection task. Since manual synonym detection is a challenging task for terminologists, we investigate whether the AltaVista-based Mutual Information (AVMI) method can be applied to the task of finding pairs of synonyms in the lexicon of a specialized sub-language. In particular, we experiment with synonyms in the field of nautical terminology. Our results indicate that AVMI is very good at spotting synonym couples among pairs of unrelated terms (with precision close to 90% at 62.5% recall) and that it outperforms more standard methods based on contextual cosine similarity. However, AVMI is not able to distinguish between synonyms and other semantically related terms. Thus, AVMI can be used for synonym mining only if it is combined with techniques to filter out other semantic relations.
Using cooccurrence statistics and the web to discover synonyms in a technical language.
BARONI, MARCO;
2004
Abstract
Turney (2001) has shown that computing the mutual information of a pair of words by using cooccurrence counts obtained via queries to the AltaVista search engine performs very effectively in a synonym detection task. Since manual synonym detection is a challenging task for terminologists, we investigate whether the AltaVista-based Mutual Information (AVMI) method can be applied to the task of finding pairs of synonyms in the lexicon of a specialized sub-language. In particular, we experiment with synonyms in the field of nautical terminology. Our results indicate that AVMI is very good at spotting synonym couples among pairs of unrelated terms (with precision close to 90% at 62.5% recall) and that it outperforms more standard methods based on contextual cosine similarity. However, AVMI is not able to distinguish between synonyms and other semantically related terms. Thus, AVMI can be used for synonym mining only if it is combined with techniques to filter out other semantic relations.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
