Experiments with Reasoning for Temporal Relations between Events

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 857–864 Manchester, August 2008 Experiments with Reasoning for Temporal Relations between Events Marta Tatu and Munirathnam Srikanth Lymba Corporation Richardson, Texas, United States marta,srikanth@lymba.com Abstract Few attempts have been made to inves- tigate the utility of temporal reasoning within machine learning frameworks for temporal relation classification between events in news articles. This paper presents three settings where temporal reasoning aids machine learned classifiers of tempo- ral relations: (1) expansion of the dataset used for learning; (2) detection of inconsis- tencies among the automatically identified relations; and (3) selection among multiple temporal relations. Feature engineering is another effort in our work to improve clas- sification accuracy. 1 Introduction In recent years, there has been a growing inter- est in temporal information extraction, as more and more operational natural language processing (NLP) systems demand dealing with time-related issues in natural language texts. Machine learning- based temporal relation identification has been ex- plored by only a few researchers, including Bogu- raev and Ando (2005), Mani et al. (2006), Cham- bers et al. (2007), and the TempEval 2007 partici- pants (Verhagen et al., 2007). For a given ordered pair of elements (x1, x2), where x1 and x2 are events or times, temporal re- lation resolution is the task of automatic identifi- cation of the relation ri ∈ TempRel that tem- porally links x1 and x2. For example, given the statement Mr. Antar was chargede137 last montht237 in a civil suite138 filede140 in federal c⃝ Lymba Corporation 2008. Licensed under the Cre- ative Commons Attribution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by- nc-sa/3.0/). Some rights reserved. court in Newark by the Securities and Exchange Commission (wsj 07781) and the pairs (e137, t137), (e137, e138), and (e138, e140), the task is to automat- ically label the given pairs with the is included, is included, and simultaneous relations, respec- tively. We note that the granularity of the temporal relations (TempRel) varies from TimeML’s 14 re- lations to TempEval’s three coarse-grain relations. While machine learning approaches attempt to improve classification accuracy through feature engineering, Mani et al. (2006) introduced a tem- poral reasoning component to greatly expand the training data. By computing the temporal closure of the training data relations, they increased the training set by a factor of 10. They reported en- couraging accuracy of classification on event-event and event-time relations. According to their ex- periments, the event-event relation accuracy goes from 62.5% to 94.95% and the event-time rela- tion accuracy ranges from 73.68% to 90.16%. Re- cently, extensions of Mani et al. (2006)’s research is briefly described in (Mani et al., 2007). This technical report addresses two problems found in (Mani et al., 2006): (1) feature vector dupli- cation caused by the data normalization process (once fixed, the accuracy drops to 76.56% and 83.23%) and (2) a somewhat unrealistic evaluation scheme (we describe Mani et al. (2007)’s results in Section 4.1). TempEval 2007 is the first standard evaluation arena that consists of three temporal relation clas- …

Content truncated to 500 words