YNU-oxz at SemEval-2020 Task 5: Detecting Counterfactuals Based on Ordered Neurons LSTM and Hierarchical Attention Network

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 683–689 Barcelona, Spain (Online), December 12, 2020. 683 YNU-oxz at SemEval-2020 Task 5: Detecting counterfactuals based on Ordered Neurons LSTM and Hierarchical Attention Network Xiaozhi Ou, Shengyan Liu, Hongling Li∗ School of Information Science and Engineering Yunnan University, Yunnan, P.R. China Abstract This paper describes the system and results of our team’s participation in SemEval-2020 Task5: Modelling Causal Reasoning in Language: Detecting Counterfactuals, which aims to simulate counterfactual semantics and reasoning in natural language. This task contains two subtasks: Subtask1–Detecting counterfactual statements and Subtask2–Detecting antecedent and conse- quence. We only participated in Subtask1, aiming to determine whether a given sentence is counterfactual. In order to solve this task, we proposed a system based on Ordered Neurons LSTM (ON-LSTM) with Hierarchical Attention Network (HAN) and used Pooling operation for dimensionality reduction. Finally, we used the K-fold approach as the ensemble method. Our model achieved an F1 score of 0.7040 in Subtask1 (Ranked 16/27). 1 Introduction In recent years, in order to simulate counterfactual semantics and reasoning in natural languages, the introduction of counterfactuals in natural language understanding systems has received more and more research attention. According to the problem of counterfactual conditionals (Goodman, 1947), counter- factual statements can be transformed into opposing statements with true premises and results. Coun- terfactual statements describe events that did not occur or cannot occur, and the consequences that may occur after the event. More specifically, counterfactuals describe events that are contrary to facts and involve common sense, understanding, and reasoning. Therefore, whether the system can correctly iden- tify counterfactual natural language statements is very important. In SemEval-2020 Task 5: Modelling Causal Reasoning in Language–Detecting Counterfactuals (Yang et al., 2020). The organizer has de- signed two subtasks. Subtask1 requires determining whether a given sentence is counterfactual. Solving this problem is the basis for all downstream counterfactual related causal reasoning analysis using natural language. Subtask2 aims to locate the premises and results in of counterfactuals, and to point out that causal insights are inherent characteristics of counterfactuals. This is to further test the causal knowledge conveyed in counterfactual statements. In this competition, we only participated in Subtask1. We used a deep learning method to build a neural network model based on Ordered Neuron LSTM (ON-LSTM) (Shen et al., 2019) and Hierarchi- cal Attention Network (HAN) (Wang et al., 2019). The difference between Ordered Neurons LSTM (ON-LSTM) and original LSTM (Hochreiter and Schmidhuber, 1997) is that encodes the hierarchical structure of sentences into features to enhance the expression ability of LSTM (Wang et al., 2018). The Hierarchical Attention Network (HAN) model has two important characteristics. The first is layering, the word-level layer and the sentence-level layer, which conforms to the document structure; the second is the attention mechanism, which can give dynamic weight according to the content when weighting. In addition, we have added the Pooling operation, and in the Pooling layer, we used Average Pooling and ∗Corresponding author: honglingli66@126.com This work is licensed under a Creative Commons …

Content truncated to 500 words