LCC-TE: A Hybrid Approach to Temporal Relation Identification in News Text

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 219–222, Prague, June 2007. c⃝2007 Association for Computational Linguistics LCC-TE: A Hybrid Approach to Temporal Relation Identification in News Text Congmin Min Language Computer Corporation 1701 N. Collins Blvd. Suite 2000 Richardson, TX 75080 cmin@languagecomputer.com Munirathnam Srikanth Language Computer Corporation 1701 N. Collins Blvd, Suite 2000 Richardson, TX 75080 Srikanth.munirathnam @languagecomputer.com Abraham Fowler Language Computer Corporation 1701 N. Collins Blvd, Suite 2000 Richardson, TX 75080 abraham@languagecomputer.com Abstract This paper explores a hybrid approach to temporal information extraction within the TimeML framework. Particularly, we focus on our initial efforts to apply machine learning techniques to identify temporal relations as defined in a constrained manner by the TempEval-2007 task. We explored several machine learning models and human rules to infer temporal relations based on the features available in TimeBank, as well as a number of other features extracted by our in-house tools. We participated in all three sub-tasks of the TempEval task in SemEval-2007 workshop and the evaluation shows that we achieved comparable results in Task A & B and competitive results in Task C. 1 Introduction There has been a growing interest in temporal information extraction in recent years, as more and more operational NLP systems demands dealing with time-related issues in natural languages. In this paper, we report on an end-to-end system that is capable of automating identification of temporal referring expressions, events and temporal relations in text by leveraging various NLP tools and linguistic resources at LCC. It has to be noted that the system we report here is not only intended for TempEval 2007 evaluation, but will also be used as a NLP tool for our other applications (e.g. temporal Question Answering). That is why we experimented to use our own temporal and event extraction capabilities in this work, although time and event tags have already been provided in the testing/training data. Another reason we use our own temporal tagging is that our temporal tagger extracts more information than that available in the training/testing data. For instance, temporal signals are removed from the data that the task organizers provide, but our temporal tagger detects that, as part of the tagging procedure. The following is an example for the tagged expression “on this coming Sunday”. <ArgStructure id="65" type="timex"> <argRef type="determiner" tokStr="this"/> <argRef type="directionIndicator” tokStr="coming"/> <argRef type="focus" tokStr="Sunday"/> <argRef type="prepSignal” tokStr="on"/> <argRef type="head" tokStr="this coming Sunday"/> <argRef type="root" tokStr="on this coming Sunday"/> <argValue type="focusType" value="weekOfDay"/> <argValue type="subType" value="Fuzzy"/> <argValue type="type" value="Date"/> </ArgStructure> Our data structure allows us to easily access and manipulate any part of the tagged chunk of text, which leaves the interpretation of whether the temporal signal on in the example is part of the temporal expression to users of temporal tagger. Taking as input this data structure, the normalization, including relative date resolution, is a straightforward process, provided that the reference time can be computed from the context. For temporal relation identification, by leveraging the capabilities of our temporal tagger, event tagger and several other in-house …

Content truncated to 500 words