Towards a Mention-Pair Model for Coreference Resolution in Portuguese

Exploring Spanish Corpora for Portuguese Coreference Resolution

“…We follow the mention-pair model (described in Section II), as it is used in several recent state-of-the-art systems [15], and even more so in Portuguese [20], [21].…”

“…Regarding coreference resolution in the Portuguese language, state-of-the-art has lacked behind more resourced languages, but direct comparison is complex as evaluation is obviously performed in different corpora. To the best of our knowledge, the coreference resolution systems reporting best results in an unrestricted Portuguese dataset are the works of Fonseca et al [20] and of Rocha and Lopes Cardoso [21]. Both use hand engineered features and linear models for mentionpair classification with promising results.…”

Coreference Resolution: Toward End-to-End and Cross-Lingual Systems

“…Fonseca et al [69] explored random undersampling in coreference resolution, with encouraging results. Rocha and Lopes Cardoso [119] proposed heuristic-based strategies to undersample the originally unbalanced dataset of mention-pair learning instances. These training set creation strategies explore well-known properties of coreference resolution to generate more balanced distribution of labels while providing suitable learning instances for the mention-pair models.…”

“…Unfortunately, these techniques have their weaknesses, as oversampling tends to lead to overfitting, and undersampling may deprive the model of useful training instances [28,119].…”

“…Sapena et al [1] reported small but consistent improvements using WordNet to judge word similarity, resulting in improved recall of coreference links but with decreased precision, which the authors postulated to be due to the typical noise produced by WordNet. By exploring semantic-based features (based on WordNet and word embeddings similarity metrics), Rocha and Lopes Cardoso [119] improved the performance of the system compared to some baseline methods for the Portuguese language, mainly because these features allowed the system to solve some coreferences between noun phrases (e.g., semantically similar words, synonyms, hyperonyms, meronyms) that were not captured by lexical-, syntactical-, morphological-, or structural-based features.…”