數理推演與對話中的自然語言理解之研究

Investigating Language Understanding in Arithmetic Reasoning and Conversation Modeling

江廷睿(Ting-Rui Chiang)

指導教授 ： 陳縕儂

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202000671

摘要

本論文主要嘗試研究目前深度學習技術對於自然語言理解的能力。所研究的理解能力主要包含兩者：第一是模型對於數理問題理解與推演的能力，第二則是模型理解對話的能力。以上兩者能力都是人類所具有的能力，但卻是還沒有被自然語言處理界使用深度學習深入調查的問題。

為了研究深度學習模型對於數理問題理解與推演的能力，本論文的第一部份著重在數學應用問題的解題任務之上。受到人類解決數學應用問題的方式的啟發，本篇論文設計了一個讓機器可以根據符號的語意產生算式的架構。結果證明，使用符號的語意果然可以增強機器的推演能力。

本論文的第二部份則著重於調查前人所提出的對話問答模型對於對話理解的能力。這部份主要關注兩個著名的的對話問題資料集 QuAC 以及 CoQA 。本篇論文在這個部份提出了一系列的實驗以檢驗模型理解的能力，並發現了一些潛在的問題。期望這些貢獻能夠有助於未來在這個方向的研究。

透過這兩方面的研究，本篇論文深入地探討了現有模型在語言理解能力上的不足。在數學應用問題的解題任務方面，仍然需要有更完善的方式檢驗模型泛化的能力。而在對話理解方面，如何設計出一個具有理解對話能力的模型也是一個未解的問題。這些都是未來的研究能夠繼續探討的方向。

關鍵字

語言理解 ； 對話問答 ； 數學應用問題解題

並列摘要

This work mainly attempt to investigate the natural language understanding capability of current deep learning models. The main understanding capabilities include two: first, the capability of language understanding and arithmetic reasoning capability for math word problems; second, the capability of conversation understanding. The above capabilities are possessed by human, but have not been well explored in the natural language processing filed with deep learning.

To investigate the arithmetic reasoning capability of deep learning models, the first part of this work focuses on the task of math work problems solving. Motivated by the solving process of human, this work proposes a framework that allows the model to generate math expressions by manipulating the symbols based on their semantics. The results show its effectiveness of improving the arithmetic reasoning capability.

The second part of this work investigates the conversation understanding capability of previous proposed models. Two renown datasets QuAC and CoQA are focused here. This part proposed a series of experiments that can serve as a tool to diagnose the conversation understanding capability of models, discovering some potential hazards.

By investigation in this two aspects, this work scrutinizes the incapabilities of current models. For the task of math word problems solving, some more efforts are still required to validate the generalizability of current models. For the task of conversation understanding, the way to design a model that understands conversations remains an unsolved problem. All of these are prospective future research directions.

並列關鍵字

language understanding ； conversational question answering ； math word problem solving

參考文獻

[1] S. Mandal and S. K. Naskar, “Solving arithmetic mathematical word problems: A review and recent advancements,” in Information Technology and Applied Mathe- matics, pp. 95–114, Springer, 2019.

[2] P. Rajpurkar, J. Zhang, K. Lopyrev, and P. Liang, “Squad: 100,000+ questions for machine comprehension of text,” in Proceedings of the 2016 Conference on Empir- ical Methods in Natural Language Processing, pp. 2383–2392, 2016.

[3] T. Nguyen, M. Rosenberg, X. Song, J. Gao, S. Tiwary, R. Majumder, and L. Deng, “Ms marco: A human generated machine reading comprehension dataset,” arXiv preprint arXiv:1611.09268, 2016.

[4] P. Rajpurkar, R. Jia, and P. Liang, “Know what you don’t know: Unanswerable questions for squad,” in Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 784–789, 2018.

[5] E. Choi, H. He, M. Iyyer, M. Yatskar, W.-t. Yih, Y. Choi, P. Liang, and L. Zettle- moyer, “Quac: Question answering in context,” in Proceedings of the 2018 Confer- ence on Empirical Methods in Natural Language Processing, pp. 2174–2184, 2018.