History Modeling in Conversational Question Answering

Please enable javascript in your browser.

Hematian Hemati, Hamed | 2023

44 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 56390 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Beigy, Hamid
Abstract:
In this thesis, we propose a method for long document modeling to address the issue of handling lengthy documents. The results indicate that our method successfully improves the problem of independently modeling chunks that arise from long documents. Additionally, we present consistent training using history augmentation to enhance the representation of questions with a long history. Our experiments demonstrate that this proposed method increases the performance of the baseline model, particularly for questions with a long history. The long document modeling approach improves the baseline's performance by 2.1% based on the F1 criteria. Furthermore, the consistent training using history augmentation method improves the performance by an additional 0.8% based on the F1 criteria and it is shown that this method provides a considerable improvement for questions with a long history. Moreover, This thesis introduces a novel conversational question answering dataset in Farsi, and we present and evaluate several benchmarks for this dataset
Keywords:
Conversational Question Answering (CQA) ; History Modeling ; Conversational Artificial Intelligence ; Long Documents

No TOC