Anotação de papéis semânticos para o português por Conditional Random Fields

Abstract

Semantic Role Labeling (SRL) can be described as a mean to achieve different purposes. Several subfields inside Natural Language Processing (NLP) benefit from semantic tags for their own goals. Reported in the literature over several centuries, the SRL task regained its popularity since 2000, when the first automatic annotated system was written. Large part of the literature is about SRL for the English language. Moreover, many papers evaluate each constituent of the sentence separately, and do not benefit from the sequential nature of words in which the task is included. The latest SRL works tend to decentralize the initial approach and reuse methodologies applied for the English language in their own languages, such as Spanish, Chinese, French, Swedish and Portuguese. Some methods were proposed for Portuguese, however, they failed to reach the level of quality obtained for the English language, and nonetheless, only one work was capable of annotating semantic roles from raw text. Thus, this work proposes an alternative system for semantically annotate portuguese text without embedded information, using a sequential model called Conditional Random Fields.