The automatic processing of medical language represents a clue for computational linguists due to intrinsic feature of these sub-codes: its lexicon comprises a vast number of terms that appear infrequently in texts. In addition, the presence of many sub-domains that can coincide in a single text complicates the collection of the training set for a supervised classification task. This paper will tackle the problem of unsupervised classification of medical scientific papers based on a hybrid Multiword Expression Discovery. We apply a morpho-semantic approach to extract medical domain terms and their semantic tags in addition to the classic MWEs discovery strategies. The collected MWEs will be used to vectorize texts and generate a network of similarities among corpus documents. With this approach, we try to solve both problems caused by the medical domain features. The presence of a vast lexicon of low-frequency terms is dealt with by extracting many semantic tags with a small dictionary; the issues of co-occurring sub-domains are solved by generating clusters of similarity values instead of a rigid classification.
Unsupervised Classification of Medical Documents Through Hybrid MWEs Discovery
Maisto, Alessandro
2021-01-01
Abstract
The automatic processing of medical language represents a clue for computational linguists due to intrinsic feature of these sub-codes: its lexicon comprises a vast number of terms that appear infrequently in texts. In addition, the presence of many sub-domains that can coincide in a single text complicates the collection of the training set for a supervised classification task. This paper will tackle the problem of unsupervised classification of medical scientific papers based on a hybrid Multiword Expression Discovery. We apply a morpho-semantic approach to extract medical domain terms and their semantic tags in addition to the classic MWEs discovery strategies. The collected MWEs will be used to vectorize texts and generate a network of similarities among corpus documents. With this approach, we try to solve both problems caused by the medical domain features. The presence of a vast lexicon of low-frequency terms is dealt with by extracting many semantic tags with a small dictionary; the issues of co-occurring sub-domains are solved by generating clusters of similarity values instead of a rigid classification.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.