Abstract
Attention mechanism is now being widely used in machine translation, image segmentation, and many other neural network-based applications. In the attention mechanism, there is an index called “attention score” which could reflect the relevance degrees between the output tokens with the input tokens. In this paper, we proposed a method using the attention score in extracting the question-related sentences from the articles of reading comprehension tests. We investigated and compared the accuracies of the extracted sentences when in word-level embedding, sentence-level embedding, and using the traditional cosine similarity. Through the evaluation of the randomly selected results of the extracted sentences, the method using the attention mechanism with sentence-level embedding obtained the accuracy of 88.3%, which was 80.9% when using word-level embedding and 83.5% when using the general cosine similarity. Meanwhile, the sentence-level embedding method also obtained the highest precision and recall in the three methods. The results suggest that using the attention mechanism with sentence-level embedding could extract the question-related sentences more accurately than traditional cosine similarity.
Information
タイトル：Extraction of Question-related Sentences for Reading Comprehension Tests via Attention Mechanism
著者：単駿杰，西原陽子，山西良典，前田亮
概要：Attention mechanism is now being widely used in machine translation, image segmentation, and many other neural network-based applications. In the attention mechanism, there is an index called “attention score” which could reflect the relevance degrees between the output tokens with the input tokens. In this paper, we proposed a method using the attention score in extracting the question-related sentences from the articles of reading comprehension tests. We investigated and compared the accuracies of the extracted sentences when in word-level embedding, sentence-level embedding, and using the traditional cosine similarity. Through the evaluation of the randomly selected results of the extracted sentences, the method using the attention mechanism with sentence-level embedding obtained the accuracy of 88.3%, which was 80.9% when using word-level embedding and 83.5% when using the general cosine similarity. Meanwhile, the sentence-level embedding method also obtained the highest precision and recall in the three methods. The results suggest that using the attention mechanism with sentence-level embedding could extract the question-related sentences more accurately than traditional cosine similarity.
書籍情報：2020 International Conference on Technologies and Applications of Artificial Intelligence (TAAI). IEEE.
発表種別：国際会議論文
発表日：2020年12月03日