UNITOR-HMM-TK: Structured Kernel-based learning for Spatial Role Labeling

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 573–579, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UNITOR-HMM-TK: Structured Kernel-based Learning for Spatial Role Labeling Emanuele Bastianelli(†)(⋆), Danilo Croce(‡), Daniele Nardi(⋆), Roberto Basili(‡) (†) DICII University of Roma Tor Vergata Rome, Italy {bastianelli}@ing.uniroma2.it (‡) DII University of Roma Tor Vergata Rome, Italy {croce,basili}@info.uniroma2.it (⋆) DIAG University of Roma La Sapienza Rome, Italy {nardi}@dis.uniroma1.it Abstract In this paper the UNITOR-HMM-TK system participating in the Spatial Role Labeling task at SemEval 2013 is presented. The spatial roles classification is addressed as a sequence-based word classification problem: the SVMhmm learning algorithm is applied, based on a simple feature modeling and a ro- bust lexical generalization achieved through a Distributional Model of Lexical Semantics. In the identification of spatial relations, roles are combined to generate candidate relations, later verified by a SVM classifier. The Smoothed Partial Tree Kernel is applied, i.e. a con- volution kernel that enhances both syntactic and lexical properties of the examples, avoid- ing the need of a manual feature engineering phase. Finally, results on three of the five tasks of the challenge are reported. 1 Introduction Referring to objects or entities in the space, as well as to relations holding among them, is one of the most important functionalities in natural language understanding. The detection of spatial utterances thus finds many applications, such as in GPS navi- gation systems, or Human-Robot Interaction (HRI). In Computational Linguistics, the task of recog- nizing spatial information is known as Spatial Role Labeling (SpRL), as discussed in (KordJamshidi et al., 2010). Let us consider the sentence: [A man]TRAJECTOR is sitting [on]SPATIAL INDICATOR [a chair]LANDMARK and talking on the phone. (1) where three roles are labeled: the phrase “A man” refers to a TRAJECTOR, “a chair” to a LAND- MARK and they are related by the spatial expres- sion “on” denoted as SPATIAL INDICATOR. The last role establishes the type of the spatial relation, e.g. Regional. The ambiguity of natural language makes this task very challenging. For example, in the same Example 1, another preposition “on” can be considered, but the phrase “the phone” is not a spatial role, as it refers to a communication mean. This mainly depends on the semantics of the gram- matical head words, i.e. chair and phone. Such phe- nomena are crucial in many learning frameworks, as in kernel-based learning (Shawe-Taylor and Cris- tianini, 2004), where the decision is based on the similarity between training and testing data. This paper describes the UNITOR-HMM-TK sys- tem participating in the Semeval 2013 Spatial Role Labeling Task (Kolomiyets et al., 2013), addressing three of the five defined sub-tasks: • Task A: Spatial Role Classification. It con- sists in labeling short sentences with spatial roles among SPATIAL INDICATOR, TRAJEC- TOR and LANDMARK. • Task B: Relation Identification. It consists in the identification of relations among roles iden- tified in Task A. This task does not involve the semantic relation classification. …

Content truncated to 500 words