Exploiting WordNet Synset and Hypernym Representations for Answer Selection

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 106–115 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 106 Exploiting WordNet Synset and Hypernym Representations for Answer Selection Weikang Li and Yunfang Wu∗ MOE Key Lab of Computational Linguistics, School of EECS, Peking University {wavejkd, wuyf}@pku.edu.cn Abstract Answer selection (AS) is an important sub- task of document-based question answering (DQA). In this task, the candidate answers come from the same document, and each an- swer sentence is semantically related to the given question, which makes it more challeng- ing to select the true answer. WordNet pro- vides powerful knowledge about concepts and their semantic relations, so we employ Word- Net to enrich the abilities of paraphrasing and reasoning of the network-based question an- swering model. Specifically, we exploit the synset and hypernym concepts to enrich the word representation and incorporate the sim- ilarity scores of two concepts that share the synset or hypernym relations into the attention mechanism. The proposed WordNet-enhanced hierarchical model (WEHM) consists of four modules, including WordNet-enhanced word representation, sentence encoding, WordNet- enhanced attention mechanism, and hierar- chical document encoding. Extensive exper- iments on the public WikiQA and SelQA datasets demonstrate that our proposed model significantly improves the baseline system and outperforms all existing state-of-the-art meth- ods by a large margin. 1 Introduction Answer selection (AS) is a challenging subtask of document-based question answering (DQA) in nat- ural language processing (NLP). The AS task is to select a whole answer sentence from the document and can be regarded as a ranking problem, which is different from the machine reading comprehen- sion (MRC) task on the SQuAD and MS-MARCO datasets. Compared with a single word or phrase, returning the full sentence often adds more value as the user can easily verify the correctness without reading a lengthy document (Yih et al., 2013). In ∗ Corresponding author. this paper, we focus on the AS task of DQA. Table 1 gives a real example of this task. Lots of fruits on answer selection have been achieved via deep learning models, including con- volutional neural network (CNN) (Yang et al., 2015), recurrent neural network (RNN) (Tan et al., 2015), attention-way (Wang et al., 2016) and gen- erative adversarial networks (GAN) (Wang et al., 2017a). Recently proposed models often consist of an embedding layer, an encoding layer, an interac- tion layer, and an answer layer (Weissenborn et al., 2017; Wang et al., 2017b; Hewlett et al., 2017). Different from other question answering like community-based question answering, the candi- date answers of DQA come from the same docu- ment, and each candidate answer is semantically related to the question. From the example in Table 1, we can see that almost every candidate answer contains the information related to the word “food” and “afghan” in the given question. As a result, it is difficult for the existing network-based models to choose the right answer, since …

Content truncated to 500 words