Several techniques have been explored in the literature to achieve domain adaptation in parsing. In principle fully unsupervised methods would be preferable, but the evidence so far is that none of them is effective, except for one special case of self-training used within one step of a reranking constituency parser. For the task of domain adaptation of dependency parsing to legal text, we hence chose to use a semi-supervised technique (i.e. active learning) which has consistently proved effective in other types of domain adaptation. We report on how we used active learning, i.e. selection criteria, parameters used, to perform domain adaptation in two languages: Italian and English. The results are quite positive on Italian and less on English. We discuss possible explanations for this discrepancy.
Active Learning for Domain Adaptation of Dependency Parsing on Legal Texts
ATTARDI, GIUSEPPE;SIMI, MARIA
2012-01-01
Abstract
Several techniques have been explored in the literature to achieve domain adaptation in parsing. In principle fully unsupervised methods would be preferable, but the evidence so far is that none of them is effective, except for one special case of self-training used within one step of a reranking constituency parser. For the task of domain adaptation of dependency parsing to legal text, we hence chose to use a semi-supervised technique (i.e. active learning) which has consistently proved effective in other types of domain adaptation. We report on how we used active learning, i.e. selection criteria, parameters used, to perform domain adaptation in two languages: Italian and English. The results are quite positive on Italian and less on English. We discuss possible explanations for this discrepancy.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.