LCC-SRN: LCC’s SRN System for SemEval 2007 Task 4

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 215–218, Prague, June 2007. c⃝2007 Association for Computational Linguistics LCC-SRN: LCC’s SRN System for SemEval 2007 Task 4 Adriana Badulescu Language Computer Corporation 1701 N Collins Blvd #2000 Richardson, TX, 75080 adriana@languagecomputer.com Munirathnam Srikanth Language Computer Corporation 1701 N Collins Blvd #2000 Richardson, TX, 75080 srikanth@languagecomputer.com Abstract This document provides a description of the Language Computer Corporation (LCC) SRN System that participated in the SemE- val 2007 Semantic Relation between Nominals task. The system combines the outputs of different binary and multi-class classifiers build using machine learning al- gorithms like Decision Trees, Semantic Scattering, Iterative Semantic Specializa- tion, and Support Vector Machines. 1 Introduction The Semantic Relations between Nominals task from SemEval 2007 focuses on identifying the se- mantic relations that hold between two arguments manually annotated with word senses (Girju et al, 2007). The previous work in identifying semantic rela- tions between nominals focuses on finding one or more relations in text for specific syntactic patterns or constructions (like genitives and noun com- pounds) using semi-automated and automated sys- tems. An overview of some of these methods can be found in (Badulescu, 2004). The LCC SRN system, developed during the SRN training period, was for us, the beginning of a different approach to semantic relations detection: detecting semantic relations in text without using a syntactic pattern. Our existing work on semantic relation detection was on detecting semantic rela- tions in text (one or more at a time) at different levels in the sentence using different syntactic pat- terns like genitives, noun compounds, verb- arguments, etc. For SRN, we built a new system that combines the output of the pattern dependent classifiers with the new pattern-independent classifiers for better results. The remainder of this paper is organized as fol- lows: Section 2 describes our system, Section 3 details the experimental results, and Section 4 summarizes the conclusions. 2 System description The system consists of two types of classifiers: classifiers that do not use the syntactic parsed tree and that were built specifically for the SemEval 2007 Task 4(SRN) and classifiers that use specific syntactic pattern to determine the semantic rela- tions and there were previously developed at LCC and then adapted to the SRN task (SRNPAT). The classifiers for each type were built from an- notated examples using supervised machine learn- ing algorithms like Decision Trees (DT)1, Support Vector Machines (SVM) 2 , Semantic Scattering (SS) (Moldovan and Badulescu, 2005) , Iterative Semantic Specialization (ISS) (Girju, Badulescu, and Moldovan, 2006), Naïve Bayes (NB) 3 and Maximum Entropy (ME)4. The outputs of different classifiers (built using different types of machine learning algorithms were combined and ranked using predefined rules. Figure 1 shows the architecture of our SRN system. 1 C5.0., http://www.rulequest.com/see5-info.html 2 LIBSVM, www.csie.ntu.edu.tw/~cjlin/libsvm/ 3 jBNC, http://jbnc.sourceforge.net 4 http://homepages.inf.ed.ac.uk/s0450736/maxent_tool- kit.html 215 Pre- process Annotated tree Text Processing Tools Classification [Arg1, Arg2, Relation, Score,Classifier] [NPArg1, NPArg2, Relation, Score, Classifier] SRN REL=1, ..7, All REL1, …, REL7, or NUL DT SV ME …

Content truncated to 500 words