Integrating Order Information and Event Relation for Script Event Prediction

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 57–67 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Integrating Order Information and Event Relation for Script Event Prediction∗ Zhongqing Wang†,‡, Yue Zhang‡ and Ching-Yun Chang‡ †Soochow University, Suzhou, China ‡Singapore University of Technology and Design, Singapore wangzq.antony@gmail.com, yue zhang@sutd.edu.sg, chang.frannie@gmail.com Abstract There has been a recent line of work au- tomatically learning scripts from unstruc- tured texts, by modeling narrative even- t chains. While the dominant approach group events using event pair relations, L- STMs have been used to encode full chain- s of narrative events. The latter has the advantage of learning long-range tempo- ral orders1, yet the former is more adap- tive to partial orders. We propose a neu- ral model that leverages the advantages of both methods, by using LSTM hid- den states as features for event pair mod- elling. A dynamic memory network is u- tilized to automatically induce weights on existing events for inferring a subsequen- t event. Standard evaluation shows that our method significantly outperforms both methods above, giving the best results re- ported so far. 1 Introduction Frequently recurring sequences of events in pro- totypical scenarios, such as visiting a restauran- t and driving to work, are a useful source of world knowledge. Two examples are shown in Figure 1, which are different variations of the “restaurant visiting” scenario, where events are partially or- dered and can be flexible. Such knowledge is useful for natural language understanding because texts typically do not include event details when mentioning a scenario. For example, the reader is expected to infer that the narrator could have been ∗This work has been done when the first author worked at SUTD. 1The term “temporal order” is used throughout this work to indicate the narrative order in texts, following Chambers and Jurafsky (2008). Strictly speaking, the event order we extract is the narrative order. X walk to restaurant X be seated X order food Y serve food X eat food X make payment X leave restaurant X walk to restaurant X be seated X eat food X make payment X leave restaurant X wait in line X read menu X take food X order food (a) (b) Figure 1: Event sequences for restaurant visiting. driving or cycling given the text “I got flat tire”. Another typical use of event chain knowledge is to help infer what is likely to happen next given a previous event sequence in a scenario. We inves- tigate the modeling of stereotypical event chains, which is remotely similar to language modeling, but with events being more sparse and flexibly or- dered than words. Our work follows a recent line of NLP re- search on script learning. Stereotypical knowl- edge about partially-ordered events, together with their participant roles such as “customer”, “wait- er”, and “table”, is conventionally referred to as scripts (Schank et al., 1977). NLP algorithms have been investigated for automatically inducing scripts from unstructured texts (Mooney and De- Jong, …

Content truncated to 500 words