CATENA: CAusal and TEmporal relation extraction from NAtural language texts

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 64–75, Osaka, Japan, December 11-17 2016. CATENA: CAusal and TEmporal relation extraction from NAtural language texts Paramita Mirza Max Planck Institute for Informatics Saarland Informatics Campus, Germany paramita@mpi-inf.mpg.de Sara Tonelli Fondazione Bruno Kessler Trento, Italy satonelli@fbk.eu Abstract We present CATENA, a sieve-based system to perform temporal and causal relation extraction and classification from English texts, exploiting the interaction between the temporal and the causal model. We evaluate the performance of each sieve, showing that the rule-based, the machine- learned and the reasoning components all contribute to achieving state-of-the-art performance on TempEval-3 and TimeBank-Dense data. Although causal relations are much sparser than temporal ones, the architecture and the selected features are mostly suitable to serve both tasks. The effects of the interaction between the temporal and the causal components, although limited, yield promising results and confirm the tight connection between the temporal and the causal dimension of texts. 1 Introduction When the Greek government missed its 1.6 billion euro payment to the IMF as its bailout expired on 30 June 2015, people started to look for information, such as What is going on? Why did it happen and what will happen next? A compact summary that represents the development of a story over time, highlighting not only the temporal connections between events but also cause-effect chains, would be very beneficial for providing information that the readers need. Besides, this kind of knowledge, derived from structured information about events and their temporal-causal relations, could be used in a number of applications, from tools for automated generation of timelines to question answering and decision support systems. While temporal relation classification is a well-studied task with a number of systems participating in the TempEval campaigns (Verhagen et al., 2010; UzZaman et al., 2013; Llorens et al., 2015), less attention has been devoted by the NLP community to the detection of causal links between events. Although recent attempts have tried to settle an annotation standard for causality inspired by TimeML (Mirza et al., 2014), the interactions between the temporal and the causal dimension of texts have been scarcely explored, especially from an empirical point of view. In this work, we face this challenge by presenting CATENA (CAusal and TEmporal relation extraction from NAtural language texts),1 a multi-sieve architecture for the extraction and classification of both relation types from English documents, which are pre-annotated with temporal entities, namely events and time expressions. 2 Related Work Our proposed approach for relation extraction is inspired by recent works on hybrid approaches for temporal relation extraction (D’Souza and Ng, 2013; Chambers et al., 2014). D’Souza and Ng (2013) introduce 437 hand-coded rules along with supervised classification models using lexical relation, semantic and discourse features. CAEVO, a CAscading EVent Ordering architecture by Chambers et al. (2014), combines rule-based and data-driven classifiers in a sieve-based architecture for temporal ordering. The classifiers (sieves) are ordered by their individual precision, and transitive closure is applied after each sieve to …

Content truncated to 500 words