Recognizing semantic relations within Polish noun phrase: A rule-based approach

Proceedings of Recent Advances in Natural Language Processing, pages 342–349, Hissar, Bulgaria, 7-13 September 2013. Semantic relation recognition within Polish noun phrase: A rule-based approach Paweł K˛edzia Institute of Informatics Wrocław University of Technology pawel.kedzia@pwr.wroc.pl Marek Maziarz Institute of Informatics Wrocław University of Technology marek.maziarz@pwr.wroc.pl Abstract The paper1 presents a rule-based approach to semantic relation recognition within the Polish noun phrase. A set of semantic re- lations, including some thematic relations, has been determined for the need of ex- periments. The method consists in two steps: first the system recognizes word pairs and triples, and then it classifies the relations. Evaluation was performed on random samples from two balanced Polish corpora. 1 Introduction Semantic relation recognition is a well-known task in natural language processing. Although the relation recognition within noun phrase and be- tween nominals was studied intensely, the task is still challenge for semantic analysis of Polish. We are aware of few papers and projects deal- ing with Semantic Role Labelling between pred- icates and their arguments, cf. (Gołuchowski and Przepiórkowski, 2012) or (Lun, 2009), but of none concerning semantic relation recognition inside Polish noun phrase. 2 Related work In (Nastase et al., 2006) authors classify semantic relations between a head and a modifier of a noun phrase. Number of all relation types was equal to 30. These relations were grouped into 5 more gen- eral groups. The authors experimented with de- cision trees, instance-based learning and Support Vector Machines. For each relation they learnt the binary classifier; as the baseline for F-measure they used the model with all of examples classified as positive and recall being equal to 100%. With 1Work financed by The National Centre for Research and Development project SP/I/1/77065/10. regard to the semantic relation the baseline ranged between 17.78% and 60.35%. Identifying the semantic relations inside com- pound nouns was presented in (Uchiyama et al., 2008). The authors used SVM classifier and in the best configuration of features, they achieved accu- racy of about 84%. In (Rosario and Hearst, 2001) authors used neu- ral networks to determine 20 semantic relations – similarily to (Nastase et al., 2006) – between a head and a modifier of noun phrase. They used a domain-specific lexical hierarchy of medicine. The authors achieved accuracy of about 60%. The workshop SemEval-2010 (task 8) con- cerned the recognition of semantic relations be- tween nominals. In (Tratz and Hovy, 2010) the authors developed a system based on the Maxi- mum Entropy classifier, able to detect 10 bidirec- tional semantic relations Achieved F-measures de- pended on the system configuration and lay be- tween 66, 68% and 77, 75%. The same set of se- mantic relations was used in (Rink and Harabagiu, 2010). The authors used Support Vector Machines classifier and a very rich set of features (i.e., part of speech for all constituents of a semantic relation pair, number of words between the nominals, fea- tures based on paths in the dependency tree from Stanford dependency parser). F-measure of this approach was 82.19%. Authors …

Content truncated to 500 words