ATT1: Temporal Annotation Using Big Windows and Rich Syntactic and Semantic Features

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 20–24, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics ATT1: Temporal Annotation Using Big Windows and Rich Syntactic and Semantic Features Hyuckchul Jung and Amanda Stent AT&T Labs - Research 180 Park Ave Florham Park, NJ 07932, USA hjung, stent@research.att.com Abstract In this paper we present the results of exper- iments comparing (a) rich syntactic and se- mantic feature sets and (b) big context win- dows, for the TempEval time expression and event segmentation and classification tasks. We show that it is possible for models using only lexical features to approach the perfor- mance of models using rich syntactic and se- mantic feature sets. 1 Introduction TempEval-3 Temporal Annotation Task (UzZaman et al., 2012) has three subtasks: A Time expression extraction and classification - extract time expressions from input text, and de- termine the type and normalised value for each extracted time expression. B Event extraction and classification - extract event mentions from input text, and determine the class, tense and aspect features for each extracted event. C Temporal link identification - identify and cate- gorise temporal links between events in the same or consecutive sentences, events and time expres- sions in the same sentence, and events and the document creation time of the input text. Here we report results for the first two tasks. Previous TempEval competitions have shown that rich syntactic and semantic feature sets can lead to good performance on event and time expression ex- traction and classification tasks (e.g. (Llorens et al., Type Files EVENT TIMEX AQUAINT gold 73 4431 579 TimeBank gold 183 6698 1243 TE3-Silver silver 2452 81329 12739 Table 1: Frequency of event and time expressions in the text portions of the TempEval-3 data sets 2010; UzZaman and Allen, 2010)). In this work, we show that with large windows of context, it is pos- sible for models using only lexical features to ap- proach the performance of models using rich syn- tactic and semantic feature sets. 2 Data Using the gold and silver data distributed by the TempEval-3 task organizers (see Table 1), we pro- cessed each input file with the Stanford CoreNLP (Stanford Natural Language Processing Group, 2012) and SENNA (Collobert et al., 2011) open- source NLP tools. From the Stanford CoreNLP tools we obtained a tokenization of the input text, the lemma and part of speech (POS) tag for each token, and dependency and constituency parses for each sentence. From SENNA, we obtained a seman- tic role labelling for each sentence. 3 Approach We were curious to explore the tradeoff between ad- ditional context on the one hand, and additional lay- ers of representation on the other, for the event and time expression extraction tasks. Researchers have investigated the impacts of different sets of features (Adafre and de Rijke, 2005; Angeli et al., 2012; 20 Feature type Features Used in Lexical 1 token ATT1, ATT2, ATT3 Lexical 2 lemma …

Content truncated to 500 words