검색 상세

Efficient Paragraph Selector via Paragraph-level Relevance Combination

  • Seongjae Choi (최성재, 대학원 컴퓨터학과 소프트웨어전공)

초록/요약

Recently, the performance of question answering (QA) task has been significantly improved, surpassing human performance when involving a single document. However, complex machine reading comprehension requiring multi-hop reasoning across multiple documents still remains a challenge. In such multiple documents, a document may not be directly related to the given question, but may be related when information from the multiple documents is combined together. However, combining the documents into a single long document to find the answer, or finding the related documents one by one in order is inefficient and takes a lot of time. Moreover, unrelated documents can interfere in finding answers and lead to performance degradation. Based on this assumption, we propose an efficient paragraph selection method, in which we select documents by finding an appropriate combination of them to answer the questions. We evaluate our method on HotpotQA, a publicly available multi-hop reasoning dataset. Our method achieves state-of-the-art performances on the leaderboard within reasonable computing time, suitable for use in practical applications.

more

목차

1. Introduction 1
2. EPS 4
2.1 Input Representation 4
2.2 Paragraph Selector 6
2.3 Reading Comprehension Model 7
3. Experiment Results 10
3.1 Dataset 10
3.2 Implementation Details 11
3.3 Baselines 11
3.4 Quantitative Results 12
3.5 Qualitative Analysis 15
3.6 Ablation Study 15
4. Related Work 19
4.1 Reading Comprehension 19
4.2 Multi-hop QA 20
4.3 Sentence Selection 20
5. Conclusion 21

more
반출 Meta View 목록