Lexical Event Ordering with an Edge-Factored Model

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1161–1171, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Lexical Event Ordering with an Edge-Factored Model Omri Abend, Shay B. Cohen and Mark Steedman School of Informatics, University of Edinburgh Edinburgh EH8 9AB, United Kingdom {oabend,scohen,steedman}@inf.ed.ac.uk Abstract Extensive lexical knowledge is necessary for temporal analysis and planning tasks. We ad- dress in this paper a lexical setting that al- lows for the straightforward incorporation of rich features and structural constraints. We ex- plore a lexical event ordering task, namely de- termining the likely temporal order of events based solely on the identity of their predi- cates and arguments. We propose an “edge- factored” model for the task that decomposes over the edges of the event graph. We learn it using the structured perceptron. As lexi- cal tasks require large amounts of text, we do not attempt manual annotation and instead use the textual order of events in a domain where this order is aligned with their temporal order, namely cooking recipes. 1 Introduction Temporal relations between events are often im- plicit, and inferring them relies on lexical and world knowledge about the likely order of events. For in- stance, to execute the instruction “fry the onion,” the hearer should probably obtain oil beforehand, even if not instructed so explicitly. Lexical knowledge about the likely order of events is therefore neces- sary for any semantic task that requires temporal rea- soning or planning, such as classifying temporal re- lations (Mani et al., 2006; Lapata and Lascarides, 2006; Yoshikawa et al., 2009; D’Souza and Ng, 2013; Mirza and Tonelli, 2014, inter alia), textual entailment (Dagan et al., 2013) or temporal infor- mation extraction (Ling and Weld, 2010). Lexical temporal knowledge is further important for model- ing grammatical phenomena such as tense and as- pect (Steedman, 2002). In this paper we address the task of lexical event ordering, namely predicting the ordering of events based only on the identity of the words compris- ing their predicates and arguments. Concretely, the task is to predict the order of an unordered set of predicate-argument structures. Predicting the likely order of event types is a step towards more in- tricate planning and reasoning scenarios (see §3), and is useful in itself for tasks such as concept- to-text generation (Reiter et al., 2000), or in val- idating the correctness of instruction sets. A re- lated idea can be found in modeling sentence coher- ence (Lapata, 2003; Barzilay and Lapata, 2008, in- ter alia), although here we focus on lexical relations between events, rather than coherence relations be- tween complete sentences. Compiling a resource of temporal tendencies be- tween events can hardly be done manually, given the number and wealth of phenomena that have to be accounted for. Temporally annotated corpora, often annotated according to TimeML principles (Puste- jovsky et al., 2003), are a useful resource for study- ing temporal relations. However, due to incurred costs, annotated …

Content truncated to 500 words