Combining POS Tagging, Dependency Parsing and Coreferential Resolution for Bulgarian

Proceedings of Recent Advances in Natural Language Processing, pages 755–762, Hissar, Bulgaria, 7-13 September 2013. Combining POS Tagging, Dependency Parsing and Co-referential Resolution for Bulgarian Valentin Zhikov and Georgi Georgiev Ontotext AD Sofia, Bulgaria valentin.zhikov@ontotext.com georgiev@ontotext.com Kiril Simov and Petya Osenova Linguistic Modelling Department IICT-BAS, Sofia, Bulgaria kivs@bultreebank.org petya@bultreebank.org Abstract This paper proposes a combined model for POS tagging, dependency parsing and co-reference resolution for Bulgarian — a pro-drop Slavic language with rich mor- phosyntax. We formulate an extension of the MSTParser algorithm that allows the simultaneous handling of the three tasks in a way that makes it possible for each task to benefit from the information available to the others, and conduct a set of experi- ments against a treebank of the Bulgarian language. The results indicate that the pro- posed joint model achieves state-of-the- art performance for POS tagging task, and outperforms the current pipeline solution. 1 Introduction Advanced language technology applications de- pend on various forms of preprocessing, such as POS tagging, parsing, co-reference resolution, word sense disambiguation, etc. Although in ideal settings these tasks have satisfactory solutions on their own, their combination in a pipeline is re- lated to a significant decrease in accuracy at each consequent stage of analysis. Recently, models that enable the single-step handling of multiple tasks have gained popularity, as they improve on the performance achieved by pipeline approaches. They take advantage of the interaction among the various levels of linguistic knowledge. Here we propose a model that challenges three tasks si- multaneously: POS tagging, dependency parsing and co-reference resolution (within a sentence). The experiments are performed on data from the Bulgarian HPSG-based treebank — BulTreeBank. Our motivation to attempt solving these particu- lar problems via a single model is many-fold: (1) avoiding the accumulation of errors inherent to pipeline processing, (2) overcoming the low speed of model-chaining approaches, (3) confirming the success of previous developments in joint model- ing; and last but not least, (4) assessing the bene- fits of modeling the interactions that exist among morphology, syntax and discourse. Pipeline approaches follow a sequence of pro- cessing that reflects the traditional levels of analy- sis within linguistics: syntax depends on morphol- ogy; co-reference resolution depends on morphol- ogy and syntax. Thus, dependency arcs are deter- mined by the grammatical features of the word- forms; co-reference chains depend on the gram- matical features of wordforms and the configura- tion of the dependency arcs. Unfortunately, this style of processing does not necessarily lead to op- timal results. One should keep in mind that some alternative interaction paths and interdependencies exist among the linguistic levels, and this interde- pendence can be accounted for in order to achieve a better solution for each task. Two phenomena in Bulgarian that illustrate this statement are: (1) co- reference links between dative verbal clitics and nouns (within a prepositional phrase, expressing the indirect object of the same verb) have common number and gender features; (2) unexpressed sub- jects participate in co-reference chains of control, binding, etc. constructions. …

Content truncated to 500 words