Automatic Labeling of Semantic Roles on Chinese FrameNet

Abstract

In this paper, semantic role labeling(SRL) on Chinese FrameNet is divided into the subtasks of boundary identification(BI) and semantic role classification(SRC). These subtasks are regarded as the sequential tagging problem at the word level, respectively. We use the conditional random fields(CRFs) model to train and test on a two-fold cross-validation data set. The extracted features include 11 word-level and 15 shallow syntactic features derived from automatic base chunk parsing. We use the orthogonal array of statistics to arrange the experiment so that the best feature template is selected. The experimental results show that given the target word within a sentence, the best F-measures of SRL can achieve 60.42%. For the BI and SRC subtasks, the best Fmeasures are 70.55 and 81%, respectively. The statistical t-test shows that the improvement of our SRL model is not significant after appending the base chunk features.

Blocked 3×2 Cross-Validated t-Test for Comparing Supervised Classification Learning Algorithms

“…The blocked 3×2 crossvalidation first proposed by Li et al (2010) in the practical application of natural language processing is a method used for corpus partitioning. They proposed that data set D should first be split into four balanced blocks and then take either the two as the training set and the other two as a test set to implement the two-fold cross-validation.…”

“…Based on the same idea, Li, Wang, Wang, and Li (2010) applied this method to the practical application of natural language processing (NLP). The method of balancing data into four blocks essentially reflects the idea of statistical experimental design, which requires that the collected data should be designed beforehand.…”

An Extended Belief Network Approach for Keyword-Based Search over PLM Databases

“…Case [1] motivated the need for a search system that integrates DB and IR querying capabilities. BANKS [2] views a database as a graph where the database tuples are the nodes and application specific relationships are the edges. When the data graph is very large, the algorithm of BANKS can not fit into the limited memory and will make the query efficiency greatly reduced.…”

Research on Software Testability Requirement Analysis Method

“…In [2], the author suggested that we should consider the testability during the software development phase and proposed corresponding method, the method only involved controllability and observability but when consideration factors increase it will hard to carry out. In [3], the author designed methods for the satellite software, but it only applied to specific software so the portability was weak.…”