透過電影中橋段隱喻偵測達成機器情境及抽象概念之理解

Situation and Abstract Concept Understanding by Trope Detection on Films

張宸晞(Chen-Hsi Chang)

指導教授 ： 徐宏民

國立臺灣大學/電機資訊學院/資訊網路與多媒體研究所/碩士(2020年)

https://doi.org/10.6342/NTU202003617

摘要

人類理解故事中的情境及抽象概念的能力是機器很難模仿的，例如：理解故事中某一個角色的意圖。理解故事中的情境及抽象概念需要對故事的敘述有相當的理解，而且通常需要考慮多個事件的發生、複雜的角色關係以及抽象的概念。為了使機器學習去理解故事中的情境及抽象概念，我們提出了一個非常具挑戰性的新問題--電影中橋段隱喻偵測。橋段隱喻(trope) 是一個說故事的工具，經常被拿來當作創作故事的材料。橋段隱喻所包含的內容很廣，從形容一個道德的概念到形容一系列的情境都有可能。儘管近年來自然語言處理的技術在內容嵌入上很成功，例如:BERT，但是要讓機器偵測橋段隱喻的表現達到人類的程度仍然非常的困難。我們提出了一個多層次理解的網路(Multi-Level Comprehension Network)，融合了需要偵測橋段的多種能力，並且設計了一個多步循環關係網路(Multi-Step Recurrent Relational Network)來推理角色之間的關係。我們提出的網路架構結合了各種的不同的理解能力，超越了BERT的表現。我們同時也提供了一個仔細的分析，為未來的研究鋪路。

關鍵字

機器學習 ； 自然語言處理 ； 電影 ； 橋段 ； 資料集

並列摘要

The human ability to understand situations and abstract concepts appearing in a story, such as understanding a character's intention, is inherently difficult for machines to mimic. It requires a sufficient understanding of the narratives presented, and often involves the consideration of multiple events, complex characters, and philosophical ideas. Here, we present a challenging new task, \textit{trope detection} on films, in an effort to create situation and abstract concept understanding for machines. Tropes are storytelling devices that are frequently used as ingredients in recipes for creative works. The meanings they represent can vary widely, from a moral concept to a series of circumstances. Despite the recent success of contextual embedding such as BERT, trope detection remains extremely challenging for machines to approach human level performance. We propose a Multi-Level Comprehension Network that incorporates different abilities required to detect the tropes and a Multi-Step Recurrent Relational Network to reason through relations among movie characters. Our proposed network outperforms BERT by aggregating multiple comprehension processes. We also provide a detailed analysis to pave ways for future researches.

並列關鍵字

Machine Learning ； Natural Language Processing ； Film ； Trope ； Dataset

參考文獻

[1] Y. Bengio. From system 1 deep learning to system 2 deep learning. Proceedings from NeuripS, 2019.

[2] K. Clark and C. D. Manning. Deep reinforcement learning for mention-ranking coreference models. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2256–2262. Association for Computational Linguistics, 2016.

[3] K. Clark and C. D. Manning. Improving coreference resolution by learning entity- level distributed representations. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 643– 653. Association for Computational Linguistics, 2016.

[4] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics, 2019.

[5] G. Gao, E. Choi, Y. Choi, and L. Zettlemoyer. Neural metaphor detection in context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 607–613. Association for Computational Linguistics, 2018.