A fast partial parse of natural language sentences using a connectionist method

A fast partial parse of natural language sentences using a connectionist method Caroline Lyon Division of Computer Science University of Hertfordshire Hatfield ALl0 9AB, UK comrcml@herts, ac. uk Bob Dickerson Division of Computer Science University of Hertfordshire Hatfield ALl0 9AB, UK comqrgd@hert s. ac. uk Abstract The pattern matching capabilities of neural networks can be used to loc- ate syntactic constituents of natural lan- guage. This paper describes a fully auto- mated hybrid system, using neural nets operating within a grammatic frame- work. It addresses the representation of language for connectionist processing, and describes methods of constraining the problem size. The function of the network is briefly explained, and results are given. 1 Introduction The pattern matching capabilities of neural net- works can be used to detect syntactic constituents of natural language. This approach bears compar- ison with probabilistic systems, but has the ad- vantage that negative as well as positive inform- ation can be modelled. Also, most computation is done in advance, when the nets are trained, so the run time computational load is low. In this work neural networks are used as part of a fully automated system that finds a partial parse of declarative sentences. The connectionist pro- cessors operate within a grammatic framework, and are supported by pre-processors that filter the data and reduce the problem to a computation- ally tractable size. A prototype can be accessed via the Internet, on which users can try their own text (details from the authors). It will take a sen- tence, locate the subject and then find the head of the subject. Typically 10 sentences take about 2 seconds, 50 sentences about 4 seconds, to process on a Sparcl0 workstation. Using the prototype on technical manuals the subject and its head can be detected in over 90% of cases (See Section 7). The well known complexity of parsing is ad- dressed by decomposing the problem, and then locating one syntactic constituent at a time. The sentence is first decomposed into the broad syn- tactic categories pre-subject - subject - predicate by locating the subject. Then these constituents can be processed further. The underlying prin- ciple employed at each step is to take a sentence, or part of a sentence, and generate strings with the boundary markers of the syntactic constituent in question placed in all possible positions. Then a neural net selects the string with the correct placement. This paper gives an overview of how natural language is converted to a representation that the neural nets can handle, and how the problem is reduced to a manageable size. It then outlines the neural net selection process. A comprehensive account is given in (Lyon, 1994); descriptions of the neural net process are also in (Lyon, 1993; Lyon and Frank, 1992). This is a hybrid sys- tem. The core process is data driven, as the para- meters of the neural networks are derived from training text. The neural net is trained in super- vised mode on examples that have been manually …

Content truncated to 500 words