Maximum Entropy Markov Models for Semantic Role Labelling

Maximum Entropy Markov Models for Semantic Role Labelling Phil Blunsom Department of Computer Science and Software Engineering University of Melbourne, Vic 3010, Australia email: pcbl@cs.mu.oz.au Abstract This paper investigates the application of Max- imum Entropy Markov Models to semantic role labelling. Syntactic chunks are labelled according to the semantic role they fill for sentence verb pred- icates. The model is trained on the subset of Prop- bank data provided for the Conference on Compu- tational Natural Language Learning 2004. Good precision is achieved, which is of key importance for information extraction from large corpora con- taining redundant data, and for generalising systems beyond task specific, hand coded template meth- ods. 1 Introduction In recent years, much progress has been made in the fields of information extraction and question answering. Research systems developed for confer- ence competitions have graduated into the commer- cial world in such applications as air travel infor- mation and booking, call handling, and banking. Although these systems perform well on their cho- sen tasks, they are generally based on a frame– and–slot approach. This approach uses application– dependent frames defined for propositions and then attempts to fill slots from words surrounding the proposition that triggered the frame. For exam- ple, in a financial system we may be interested in extracting company mergers from newswires. We could define a frame for the verb stem merge, and slots in that frame for the companies involved in the merger. Although this could be effective for the chosen domain, each time we want to develop a system for a new domain we need to start from scratch. In order to build broad coverage systems capable of generalising, we need a way of defining and labelling propositions and their arguments that is not tied to a particular application. Systems have been developed to address the task of semantic role labelling (SRL) using a variety of machine learning techniques and features ranging from simple lexical information to those derived # ↵⌧%⇧⌦⌃% ⌘⌅- ! ⌦$⇣$⇤⌦ ⇧⌦⇧$⇥*⇧+*⇧,⌘**.⇣✓⇧&⇤$⇥⇧◆*⌘.⇧⇣⌘2 $⌘⌥⇢$⌥⌘* ! D⌘⇣↵*⇧%⌦⌘⇧%⌅⌫⇥⇧⇡ <!DOCTYPE frameset SYSTEM "frameset.dtd"> <frameset> <predicate lemma="begin"> <roleset id="begin.01" name="start" vncls="55.1"> <roles> <role descr="Agent" n="0"> <vnrole vncls="55.1" vntheta="Agent"/></role> <role descr="Theme(-Creation)" n="1"> <vnrole vncls="55.1" vntheta="Theme"/></role> <role descr="Instrument" n="2"/> </roles> Figure 1: Propbank XML frame for begin from parse trees. These semantic roles identify arguments of verb predicates and can be general, such as agent, theme, or verb specific, such as A0, A1. This paper presents a discriminative Markov model for the semantic role labelling task proposed for the Conference on Computational Natural Lan- guage Learning 2004 (CoNLL). Firstly, recent work on SRL is examined and the data available for this task is presented. A discriminative maximum entropy framework is defined and a SRL model described. This model trains maximum entropy classifiers for each state in the Markov model, rep- resenting the probability distribution for transitions from that state for a given observation feature vec- tor. This approach combines the advantages of tran- sition sequence based models with the ability of maximum entropy classifiers to handle a …

Content truncated to 500 words