On the contribution of word embeddings to temporal relation classification

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2818–2828, Osaka, Japan, December 11-17 2016. On the contribution of word embeddings to temporal relation classification 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 Temporal relation classification is a challenging task, especially when there are no explicit markers to characterise the relation between temporal entities. This occurs frequently in inter- sentential relations, whose entities are not connected via direct syntactic relations making clas- sification even more difficult. In these cases, resorting to features that focus on the semantic content of the event words may be very beneficial for inferring implicit relations. Specifically, while morpho-syntactic and context features are considered sufficient for classifying event-timex pairs, we believe that exploiting distributional semantic information about event words can ben- efit supervised classification of other types of pairs. In this work, we assess the impact of using word embeddings as features for event words in classifying temporal relations of event-event pairs and event-DCT (document creation time) pairs. 1 Introduction The classification of temporal relations between events in text has been long studied and attacked from different perspectives in the NLP community. However, existing approaches heavily rely on information overtly expressed in text, such as explicit temporal markers (e.g. before, during), the tense, aspect and modality of event words, as well as specific syntactic constructions. In case overt indicators are missing, the task becomes significantly more challenging, as is often the case when two events take place in different sentences, or in anchoring an event to the document creation time (DCT). See for example the sentences in (i), where the label for the event pair (e1, e2) is BEFORE, and the sentence in (ii), where e INCLUDES the DCT. (i) When Wong Kwan spent e1 seventy million dollars for this house, he thought it was a great deal. He sold e2 the property to five buyers and said he’d double his money. (ii) The U.N. Security Council on Aug. 6 ordered a global embargo e on trade with Iraq as punishment for seizing Kuwait. Inter-sentential event relations are quite frequent, covering for example 32.76% of the event pairs in the TempEval-3 evaluation corpus (UzZaman et al., 2013). Around 42.37% of pairs of an event and a time expression in the same corpus are actually pairs of an event and the DCT. Moreover, the TimeBank corpus contains 718 temporal relations which are co-ordinated by temporal signals, i.e., only 11.2% of all temporal links (Derczynski and Gaizauskas, 2013). These make research on implicit temporal ordering very relevant. One common approach, first proposed in Marcu and Echihabi (2002), incorporates word-based infor- mation in the form of word pair feature vectors. Conventionally, a word is converted into a symbolic ID, which is then transformed into a feature vector using a one-hot representation: the feature vector has the same length as the size of the vocabulary, and only one dimension is on. From …

Content truncated to 500 words