BLCUNLP: Corpus Pattern Analysis for Verbs Based on Dependency Chain

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 325–328, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics BLCUNLP: Corpus Pattern Analysis for Verbs Based on Dependency Chain Yukun Feng, Qiao Deng and Dong Yu† College of Information Science, Beijing Language and Culture University No.15 Xueyuan Rd., Beijing, China, 100083. †The corresponding author. {fengyukun,dengqiao,yudong}@blcu.edu.cn Abstract We implemented a syntactic and semantic tagging system for SemEval 2015 Task 15: Corpus Pattern Analysis. For syntactic tagging, we present a Dependency Chain Search Algorithm that is found to be effective at identifying structurally distant subjects and objects. Other syntactic labels are identified using rules defined over dependency parse structures and the output of a verb classification module. Semantic tagging is performed using a simple lexical mapping table combined with post- processing rules written over phrase structure constituent types and named entity information. The final score of our system is 0.530 F1, ranking second in this task. 1 Introduction Corpus Pattern Analysis (CPA) is an important language analysis technique, which attempts to describe the patterns of word usage in text. In this paper, we present the system we developed for SemEval-2015 Task 15: CPA, Subtask1: CPA parsing. The system operates in two stages: syntactic tagging and semantic tagging. We first search for the syntactic roles of a verb’s arguments in a sentence. We use the following tag set for the syntactic roles: “subj” is for subject, “obj” is for object, “iobj” is for indirect object, “advprep” is for adverbial preposition or other adverbial/verbal link, “acomp” is for adverbial or verb complement, and “scomp” is for noun or adjective complement. For example, take a sentence whose core verb is “plan”: “Mr Eigen plans to wage his war diplomatically”. The correct tagging of syntactic and semantic roles is: Mr [subj/Human Eigen] plans [advprep/LexicalItem to] [acomp/Activity wage] his war diplomatically. Due to time constraints, we put more effort into improving the accuracy of syntactic tagging. We rely on simpler techniques for semantic tagging. For syntactic tagging, we use Stanford CoreNLP to extract linguistic attributes, deduce dependency chains through dependency relations and to classify verbs. When performing semantic tagging, we use a data driven mapping of words to their most frequent semantic tag in the task's training data in conjunction with a small number of post- processing rules. 2 Our Methods 2.1 System Framework Our system consists of five modules (Figure 1). The first module is Preprocessing, which generates input files with the correct format for Stanford CoreNLP to extract linguistic attributes. The second module is Linguistic Attributes. For the syntactic layer, tagged arguments must have direct or indirect dependency relations with the core verb. Dependency relations are thus a critical attribute for correctly selecting tagged units and types. We employ a number of additional linguistic attributes for our tagging rules: parts of speech (POS) provide useful information for syntactic tagging; direct dependency relations and phrase type are helpful in identifying and following a dependency chain. Last, named entity (NE) tags and phrase-structure …

Content truncated to 500 words