Improving Neural Story Generation by Targeted Common Sense Grounding

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5988–5993, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5988 Improving Neural Story Generation by Targeted Common Sense Grounding Huanru Henry Mao, Bodhisattwa Prasad Majumder, Julian McAuley, Garrison W. Cottrell Department of Computer Science and Engineering UC San Diego {hhmao, bmajumde, jmcauley, gary}@eng.ucsd.edu Abstract Stories generated with neural language models have shown promise in grammatical and stylis- tic consistency. However, the generated sto- ries are still lacking in common sense reason- ing, e.g., they often contain sentences deprived of world knowledge. We propose a simple multi-task learning scheme to achieve quanti- tatively better common sense reasoning in lan- guage models by leveraging auxiliary training signals from datasets designed to provide com- mon sense grounding. When combined with our two-stage fine-tuning pipeline, our method achieves improved common sense reasoning and state-of-the-art perplexity on the Writing- Prompts (Fan et al., 2018) story generation dataset. 1 Introduction Story generation is the task of automatically pro- ducing compelling creative writing. Recent ad- vances in language modeling have yielded the- matic and stylistic coherence in story generation through large scale pretraining of Transformer models (Vaswani et al., 2017). The recent intro- duction of the General Pre-trained Transformer v2 (GPT2) (Radford et al., 2019)—a high-capacity Transformer trained on a large, diverse corpus of text crawled from the web (called WebText)—is capable of generating stylistically coherent text but commonly produces text with logical incon- sistencies. For example, in one sample the model writes: “It was a sunny, warm summer night”. Ob- viously, this writing is nonsense as it cannot be sunny at night. The lack of common sense rea- soning in such a strong language model suggests that minimizing next-token perplexity alone may be insufficient in producing models that can com- pose sensible stories. In this paper, we consider the challenge of com- mon sense reasoning (CSR) in language model- Figure 1: Our two-stage fine-tuning pipeline with aux- iliary multi-task learning. For perplexity ranking ex- amples, bolded text indicates the correct answer. For synthetic, the correct choice is written by a human, and the wrong choice is generated by a neural network. ing for story generation. Unlike other work in the CSR literature (Storks et al., 2019), which eval- uates CSR in isolation, we are specifically inter- ested in a generative model’s likelihood of pro- ducing text that exhibits common sense. Evaluat- ing common sense qualitatively in a model’s sam- ples is difficult, as it is subject to human bias and dependent on sampling procedure. We propose evaluating the common sense of a model automat- ically by ranking the model’s perplexity on spu- rious (plausible, but nonsense) text completions from SWAG (Zellers et al., 2018) and Story Cloze (Mostafazadeh et al., 2016) datasets, which are de- signed for common sense grounding. Our contributions are as follows: We propose a simple way to define better CSR in generative models, which leads to …

Content truncated to 500 words