Pheonix at SemEval-2020 Task 5: Masking the Labels Lubricates Models for Sequence Labeling

Abstract

This paper presents the deep-learning model that is submitted to the SemEval-2020 Task 5 competition: "Detecting Counterfactuals". We participated in both Subtask1 and Subtask2. The model proposed in this paper ranked 2nd in Subtask2: "Detecting antecedent and consequence". Our model approaches the task as a sequence labeling. The architecture is built on top of BERT; and a multi-head attention layer with label masking is used to benefit from the mutual information between nearby labels. Also, for prediction, a multi-stage algorithm is used in which the model finalize some predictions with higher certainty in each step and use them in the following. Our results show that masking the labels not only is an efficient regularization method but also improves the accuracy of the model compared with other alternatives like CRF. Label masking can be used as a regularization method in sequence labeling. Also, it improves the performance of the model by learning the specific patterns in the target variable.

SemEval-2020 Task 5: Counterfactual Recognition

“…To predict the start and ending positions of antecedents and consequents, the model utilizes an ensemble of RoBERTa models and extend it in the same manner as how BERT was extended for the SQuAD dataset (Rajpurkar et al, 2016). The second-place system pouria babvey uses a sequence labelling approach: the authors develop the model on top of BERT with a multi-head attention layer and label masking to capture mutual information between nearby labels (Babvey et al, 2020). Label masking, in which only part of the labels is fed during training and the rest have to be predicted, has shown to be particularly effective for improving accuracy, which can be seen as a form of regularization.…”