Parsing the Penn Chinese Treebank with Semantic Knowledge

Parsing the Penn Chinese Treebank with Semantic Knowledge Deyi Xiong1,2, Shuanglong Li1,3, Qun Liu1, Shouxun Lin1, and Yueliang Qian1 1 Institute of Computing Technology, Chinese Academy of Sciences, P.O. Box 2704, Beijing 100080, China {dyxiong, liuqun, sxlin}@ict.ac.cn 2 Graduate School of Chinese Academy of Sciences 3 University of Science and Technology Beijing Abstract. We build a class-based selection preference sub-model to in- corporate external semantic knowledge from two Chinese electronic se- mantic dictionaries. This sub-model is combined with modifier-head gen- eration sub-model. After being optimized on the held out data by the EM algorithm, our improved parser achieves 79.4% (F1 measure), as well as a 4.4% relative decrease in error rate on the Penn Chinese Treebank (CTB). Further analysis of performance improvement indicates that se- mantic knowledge is helpful for nominal compounds, coordination, and N⋄V tagging disambiguation, as well as alleviating the sparseness of in- formation available in treebank. 1 Introduction In the recent development of full parsing technology, semantic knowledge is sel- dom used, though it is known to be useful for resolving syntactic ambiguities. The reasons for this may be twofold. The first one is that it can be very difficult to add additional features which are not available in treebanks to generative models like Collins (see [1]), which are very popular for full parsing. For smaller tasks, like prepositional phrase attachment disambiguation, semantic knowledge can be incorporated flexibly using different learning algorithms (see [2,3,4,5]). For full parsing with generative models, however, incorporating semantic knowl- edge may involve great changes of model structures. The second reason is that semantic knowledge from external dictionaries seems to be noisy, ambiguous and not available in explicit forms, compared with the information from treebanks. Given these two reasons, it seems to be difficult to combine the two different information sources–treebank and semantic knowledge–into one integrated sta- tistical parsing model. One feasible way to solve this problem is to keep the original parsing model unchanged and build an additional sub-model to incorporate semantic knowledge from external dictionaries. The modularity afforded by this approach makes it easier to expand or update semantic knowledge sources with the treebank R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 70–81, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 Parsing the Penn Chinese Treebank with Semantic Knowledge 71 unchanged or vice versa. Further, the combination of the semantic sub-model and the original parsing model can be optimized automatically. In this paper, we build a class-based selection preference sub-model, which is embedded in our lexicalized parsing model, to incorporate external seman- tic knowledge. We use two Chinese electronic dictionaries and their combi- nation as our semantic information sources. Several experiments are carried out on the Penn Chinese Treebank to test our hypotheses. The results indi- cate that a significant improvement in performance is achieved when seman- tic knowledge is incorporated into parsing model. Further improvement analy- sis is made. We confirm that semantic knowledge is indeed useful for nominal compounds and coordination ambiguity resolution. And surprisingly, semantic knowledge is …

Content truncated to 500 words