Dependency-graph Based Chinese Semantic Parsing

Semantic Dependency Parsing (SDP) is deep semantic analysis of Chinese. A well-formed semantic dependency scheme is the foundation of SDP. However, public andcanonical dependency scheme is rare. The semantic dependency scheme from HIT is thefirst one used to organize the international public assessment in SemEval-2012withnormalization and usability. However, flaws of HIT semantic dependency are exist, suchas there are too many semantic labels, some of which are rarely mentioned or used;there are much overlapping between each semantic labels and so on. Thus it needsfurther improvement.In this paper, we refine the HIT dependency scheme using stronger linguistictheories, and we propose a scheme with more clear hierarchy and more canonical labelset. To cover Chinese semantics more comprehensively, we make a break away from theconstraints of dependency trees, and expand our dependency structure to graphs,allowing the existence of crossed arcs and more than one head on certain node. Thisbreak is useful, because in some circumstance loss or transformations of dependencyarcs always cause loss of semantics. So dependency graph is essential.There are two questions in semantic dependency parsing, one is to design asemantic representation system, and the other one is how to build an auto parsingalgorithm. So, another work this paper focused on is how to parse the semanticdependency graphs. Unitil now, most previous studies on dependency parsing arefocused on trees, no parser for dependency graphs exists. With the characteristic ofdependency graphs containing entire dependency trees, we utilize SVM classification toparse semantic dependency graphs based on the parsing of dependency trees.The parsing system works in a pipelined manner, and the parsing of trees is thebasis, so we improve the performance of the tree parser. One way is to consider thesyntactic dependency information, because the similarity between syntactic dependencyand semantic dependency can help to find arcs more accurately. Another way is tointroduce more semantic information of words, brown cluster classifies words intodifferent categories in accordance with semantics. This information will bring help forwords that appear fewer times in training set. The third way is to add high precisioncollocation of words extracted from large-scale unlabeled corpus, which is the samewith the nature of semantic dependency parsing, finding word pairs with direct semanticrelations. So collocation of words will supply guidance. Experiments show that thesethree methods improve the baseline by0.62%,0.74%and1.65%respectively.