Improving Coreference Resolution with Semantic Knowledge

Abstract

In this paper we present a Basque coref-erence resolution system enriched with semantic knowledge. An error analysis carried out revealed the deficiencies that the system had in resolving coreference cases in which semantic or world knowledge is needed. We attempt to improve the deficiencies using two semantic knowledge sources, specifically Wikipedia and Word-Net.

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.…”

“…On the Portuguese Corref-PT dataset, 923,566 mention-pairs were generated, 45,659 of which were positive learning instances and 877,907 were negative learning instances, corresponding to a 4.9%/95.1% split. This class imbalance problem has been extensively studied in literature, and is usually tackled by using random undersampling of the majority class [20], [31]. We chose to perform our own study using one of the proposed architectures, aiming to identify which undersampling percentage is able to maximize performance on specific coreference metrics.…”

“…Additionally, due to using a recent corpus for Portuguese coreference resolution [26], we report the first coreference-specific results on the Corref-PT corpus. Direct comparison of our system with the works of Fonseca et al [20] or Rocha and Lopes Cardoso [21] is not possible, as the performance of these systems on model-independent metrics is not reported, besides using a different corpus.…”

“…Aiming to improve model performance, we use the Arch2 architecture, described in Section IV-D, to study which undersampling percentage is best suited for this task. Despite Fonseca et al [20] using one-to-one undersampling (equal sampling of both classes), and measuring satisfactory results with this method [31], performance was reported on sampled mention-pairs, not model-independent coreference metrics. We train the model on the AnCora corpus, and report performance on macro-F1 of the mention-pairs test set (with original sampling) and the CoNLL metric.…”