ArT: All-round Thinker for Unsupervised Commonsense Question Answering

Proceedings of the 29th International Conference on Computational Linguistics, pages 1490–1501 October 12–17, 2022. 1490 ArT: All-round Thinker for Unsupervised Commonsense Question Answering Jiawei Wang1,2, Hai Zhao1,2∗ 1 Department of Computer Science and Engineering, Shanghai Jiao Tong University 2 Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong University wjw_sjt@sjtu.edu.cn, zhaohai@cs.sjtu.edu.cn Abstract Without labeled question-answer pairs for nec- essary training, unsupervised commonsense question-answering (QA) appears to be ex- tremely challenging due to its indispensable unique prerequisite on commonsense source like knowledge bases (KBs), which are usu- ally highly resource consuming in construc- tion. Recently pre-trained language models (PLMs) show effectiveness as an alternative for commonsense clues when they play a role of knowledge generator. However, existing work either relies on large-scale in-domain or out-of- domain labeled data, or fails to generate knowl- edge of high quality in a general way. Moti- vated by human thinking experience, we pro- pose an approach of All-round Thinker (ArT) by fully taking association during knowledge generating. In detail, our model first focuses on key parts in the given context, and then gener- ates highly related knowledge on such a basis in an association way like human thinking. Be- sides, for causal reasoning, a reverse thinking mechanism is especially added to further en- hance bidirectional inferring between cause and effect. ArT is totally unsupervised and KBs- free. We evaluate it on three commonsense QA benchmarks: COPA, SocialIQA and SCT. On all scales of PLM backbones, ArT shows its brilliant performance and outperforms previ- ous advanced unsupervised models. Our code is available at https://github.com/WangJW424/ commonsenseQA-ArT. 1 Introduction Commonsense question-answering (QA) has been a more challenging natural language understand- ing (NLU) task than conventional QA tasks, for it requires extra commonsense knowledge, which cannot be directly acquired from the given context, to make an appropriate answer (Niu et al., 2021). ∗ Corresponding author. This paper was partially sup- ported by Key Projects of National Natural Science Founda- tion of China (U1836222 and 61733011). Notes: 1) Shadow is an area without light. 2) Grass is geen. 3) Sun is heat and light. 4) Cutting means to remove. Context: My body cast a shadow over the grass. Question: What is the cause of this? Options: A. The sun was rising. B. The grass was cut. Knowledge: 1) Light is obscured to produce shadows. 2) The sun emits light. Figure 1: An commonsense QA exmaple in COPA. Text with blue background is the raw example and the gold answer is ticked. Text with green background is the needed commonsense knowledge. Text with orange background is notes generated by our ArT. Keyphrases are marked in red. Figure 1 gives a typical example. To select the correct cause of the fact “My body cast a shadow over the grass”, models should know that “Light is obscured to produce shadows” and “The sun emits light”. As human beings, we can immedi- ately make this association and judgment because above knowledge is deeply branded in our mind due …

Content truncated to 500 words