There might be hundreds of arguments for a single debate topic. and it is usually difficult to read them all. Therefore, some of the existing studies have attempted to summarize and aggregate these arguments manually into about 14 types of content (key point) in order to make them easier to understand. Some researchers pair key points with arguments and use BERT to identify whether an argument and a key point are matching or not. However, the previous study averaged word vectors to compute sentence vectors, which resulted in the unbalance of important word information. This means that the important words should be weighted more than other words, but the previous study does not consider about that. To solve this problem, we implemented a sentence embedding model called Sentence-BERT, which is fine-tuned on NLI dataset with BERT. The results are better than the existing SOTA method, i.e. BERT. Furthermore, to compare each words between two sentences, we introduce MoverScore, which yield the best results.
抄録全体を表示