SSIN: Sentence Semantic Interaction Network for Multi-Choice Reading Comprehension

Abstract

Multiple-choice reading comprehension is a challenging task in natural language processing, which aims to select the correct answer from a set of candidate options when given passage and question. Previous approaches usually focus only on word vector interactions and ignore the importance of sentence semantics for reading when modeling the relationship between passage and question. However, reading is a process that includes complex interactions of various knowledge such as vocabulary, syntax and semantics. Interactions based on word vectors alone do not effectively capture the relationship between passage and question. In this work, we propose the Sentence Semantic Interaction Network (SSIN), which models the relationship among passage, question, and answer options based on sentence semantics. The experimental results show that superior results are achieved on both the RACE and MCTest datasets, confirming that the interaction based on sentence semantic vectors can effectively improve the performance of the model reading comprehension.

Discriminative Question Answering via Cascade Prompt Learning and Sentence Level Attention Mechanism

“…As the document sets have been filtered in the document retrieval part, the YNQA part only needs to answer the questions with these retrieved documents. The methods of YNQA mainly include two categories: classificationbased models [47], [48] and prompt-learning-based models [19], [28], [49]- [51].…”

“…Zhang et al [47] proposed a dual co-matching network (DCMN), which models the relationship among passage, question, and answer options bidirectionally, and a classifier is used to select the most probable answer. Xu et al [48] proposed the sentence semantic interaction network (SSIN) to model the relationship among passage, question, and answer options based on sentence semantics, which can effectively improve the performance of the model reading comprehension. Moreover, with the pre-trained model proposed, more and more researchers use the BERT [23] or RoBERTa [52] model for the representation of the interaction between questions and documents, and use their [CLS] flag which is designed for classification to get the right answer.…”