Generating Reasonable and Diversified Story Ending Using Sequence to Sequence Model with Adversarial Training

Proceedings of the 27th International Conference on Computational Linguistics, pages 1033–1043 Santa Fe, New Mexico, USA, August 20-26, 2018. 1033 Generating Reasonable and Diversified Story Ending Using Sequence to Sequence Model with Adversarial Training Zhongyang Li, Xiao Ding and Ting Liu∗ Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {zyli,xding,tliu}@ir.hit.edu.cn Abstract Story generation is a challenging problem in artificial intelligence (AI) and has received a lot of interests in the natural language processing (NLP) community. Most previous work tried to solve this problem using Sequence to Sequence (Seq2Seq) model trained with Maximum Like- lihood Estimation (MLE). However, the pure MLE training objective much limits the power of Seq2Seq model in generating high-quality stories. In this paper, we propose using adversarial training augmented Seq2Seq model to generate reasonable and diversified story endings given a story context. Our model includes a generator that defines the policy of generating a story ending, and a discriminator that labels story endings as human-generated or machine-generated. Carefully designed human and automatic evaluation metrics demonstrate that our adversarial training augmented Seq2Seq model can generate more reasonable and diversified story endings compared to purely MLE-trained Seq2Seq model. Moreover, our model achieves better perfor- mance on the task of Story Cloze Test with an accuracy of 62.6% compared with state-of-the-art baseline methods. 1 Introduction The task of story generation was first proposed in the field of artificial intelligence (AI) (Schank and Abelson, 1977). Recently, benefiting from deep learning technology, story generation again received increasing interests in the natural language processing (NLP) community. In this paper, we propose using adversarial training augmented Seq2Seq model to generate story ending given a story context. For example, given the story context “Sara had lost her cat. She was so sad! She put up signs all over the neighborhood. Then a wonderful thing happened.”, our goal is to generate a possibly reasonable story ending “Somebody found her cat”. Much previous work in story generation or prediction focused on learning statistical models of event sequences from large-scale text corpora. Chambers and Jurafsky (2008) proposed unsupervised induc- tion of narrative event chains from raw newswire texts, with narrative cloze as the evaluation metric. However, they utilized a very impoverished representation of events as the form of (event, dependency). To overcome the drawback of this event representation, Pichotta and Mooney (2014) presented a script learning approach that employed events with multiple arguments. Pichotta and Mooney (2016a) showed that the LSTM-based event sequence model outperformed previous co-occurrence-based methods for event prediction. However, this line of work build their models based on discrete verbs and tokens, which is far from being a complete sentence or a story. There have been a number of recent work for story generation focusing on complete sentences. Kiros et al. (2015) described an approach for unsupervised learning of a generic, distributed sentence repre- sentations called skip-thought vectors. Such vectors can be used to predict neighboring sentences in the context. Pichotta and Mooney (2016b) used Seq2Seq framework directly operating on …

Content truncated to 500 words