뉴럴 및 심볼릭 방법을 통합한 하이브리드 단락 검색

다국어 초록 (Multilingual Abstract)

Neural retrieval has the advantage of resolving the mismatch problem of symbolic retrieval, but the efficiency of symbolic retrieval is still high in a specific question and paragraph environment, so both methodologies are being studied in parallel. In this paper, we proposed a hybrid retrieval that integrates the two search methods in the Korean text. The features used in the hybrid search are a term frequency-based symbolic feature vector and a neural feature vector represented by language model based encoder. We calculated the relevance score of the question-passage by combining the two feature vectors. As a result of the experiment, the BR@Top1 improved by 6.35% and 4.41% compared to the single retrieval method in the wikipedia and office document collections, and also improved 1.00% and 0.73% via BR@Top50.

참고문헌 (Reference)

1 J. Ramos, "Using TF-IDF to Determine Word Relevance in Document Queries" 133-142, 2003

2 Massimo Nicosia, "Semantic Linking in Convolutional Neural Networks for Answer Sentence Selection" Association for Computational Linguistics 1070-1076, 2018

3 R. Nogueira, "PASSAGE RE-RANKING WITH BERT"

4 Zhenzhen Li, "LSTM-based Deep Learning Models for Answer Ranking" IEEE 90-97, 2016

5 "KorQuad1.0"

6 H. Padigela, "Investigating the Successes and Failures of BERT for Passage Re-Ranking"

7 J. Ma, "Hybrid First-stage Retrieval Models for Biomedical Literature" 2020

8 D. Rau, "How Different are Pre-trained Transformers for Text Ranking?"

9 "Faiss"

10 "ETRI AIOpen"

11 V. Karpukhin, "Dense Passage Retrieval for Open-Domain Question Answering" 6769-6781, 2020

12 J. Devlin, "Bert: Pre-training of Deep Bidirectional Transformers for Language Understanding" 4171-4186, 2019

13 Billel Aklouche, "BM25 Beyond Query-Document Similarity" Springer International Publishing 65-79, 2019

14 A. Vaswani, "Attention is all you need" 6000-6010, 2017