JU_CSE_TAC: Textual Entailment Recognition System at TAC RTE-6

摘要

The note describes the Recognizing Textual Entailment (RTE) system developed at the Computer Science and Engineering Department, Jadavpur University, India. In this competition, we participated and submitted the results in the RTE-6 Main Task (3 runs), Novelty Task (3 runs) and RTE-6 KBP task (3 runs for generic task and 3 runs for tailored task). For the Main and the Novelty Tasks, the corpus was a collection of news wire documents from various sources and arranged into particular topics, a hypothesis H and a set of "candidate" sentences retrieved by Lucene from that corpus for the hypothesis H. Each sentence in the set of documents associated with a given topic was involved in an entailment relationship with each hypothesis for the topic. RTE systems are required to identify all the sentences that entail H among the candidate sentences. For the Main and the Novelty Tasks, the system is based on the composition of lexical entailment module, lexical distance module, Chunk module, Named Entity module and syntactic text entailment (TE) module. Our TE system is based on the Support Vector Machine (SVM) that uses twenty five features for lexical similarity, the output tag from a rule based syntactic two-way TE system as a feature and the outputs from a rule based Chunk Module and Named Entity Module as the other features. For the Main task test set, the following micro-average results were obtained for Run 1: F-Score 34.79, Run 2: F-Score 26.78 and Run 3 : F-score 31.19. For the novelty task test set, the following micro-average results were obtained for Run 1: Novelty Evaluation F-Score 81.77 and Justification Evaluation F-Score 34.35, Run 2: Novelty Evaluation F-Score 78.18 and Justification Evaluation 26.87 and Run 3: Novelty Evaluation F-score 78.69 and Justification Evaluation 24.57 were obtained. The KBP Slot Filling task is focused on the searching a collection of news wire and Web documents and extracting values for a predefined set of attributes ("slots") for the target entities. The RTE KBP Validation Pilot is based on the assumption that extracted slot filler is correct if and only if the supporting document entails an hypothesis created on the basis of the slot filler. In RTE KBP, we participated for generic task and tailored task. For the RTE-6 KBP test set for Generic Task, micro-average results for Run 1: F-Score 0.1403, Run 2: F-Score 0.172 and Run 3: F-score 0.1531 were obtained. For RTE-6 KBP test set for Tailored Task, micro-average results for Run 1: F-Score 0.3, Run 2: F-Score 0.3307 and Run 3: F-score 0.3288 were obtained.