Grammar Checking In A Combined Model

Grammar checking is an interesting field of computational linguistics that is worthy of our exertion. Different technologies and methods have been used in this area. In projects by Kukich (1992), by Atwell and Elliott (1987), by Douglas and Dale (1992), by Golding and Schabes (1996), by Bustamante and Leon (1996) and by Gojenola and Oronoz (2000) this problem has been tackled from various points of view with different methodologies. And besides these, there are some valuable research projects that are directly aimed at conducting grammar checking on a practical scale. Kuenning's Ispell, Atkinson's Aspell, Bernth's EasyEnglish (2000), Critique of Jensen and his colleagues (1993), Bustamante and Leon's GramCheck (1996) and FLAG of Bredenkamp and his colleagues (2000) are examples of these projects.Even though the research on grammar checking can be carried out throughdeterministic or non-deterministic methods, checking the grammatically using an appropriate parsing technology is a way that deserves our interest and exertion. In order to build up a successful grammar checking program, we need a robust parsing system, which is reliably accurate and fast. Up till now numerous parsers have been developed under different frameworks using different strategies, among which, the link grammar parser developed by Sleator and Temperley (1991;Lafferty, Sleator, & Temperley, 1992) is one of the best candidates for our needs. Rather than examine the basic context of a word within a sentence, their Link Parser is based on a model that words within a text form "links" with one another. In this respect, the link grammar is theoretically linked with dependency grammar. The fundamental working mechanism of the Link Parser is that the linking requirement of all the words in the sentence should be satisfied. The importance of this mechanism lies in its ability to help determine the grammaticality of sentences. This is a quality that most other parsing systems lack, whether they are statistical-based or rule-based.And on the other hand, the Link Parser very often generates scores of parses (in link grammar terminology, linkages) with an order determined by the parser's evaluation system (in link grammar terminology, cost system). This may sometimes cause trouble when we try to find the correct parse among the bulk of them, which may be mis-ranked due to the parser's inefficiency. The current research question is directed to overcoming this deficiency. I have proposed a way of elevating the accuracy of the judgment of the link grammar parser. I argue in this dissertation that, with semantics and world knowledge not sufficiently exploited, at the current stage of computational linguistic development, using phrase structure knowledge in dependency grammar parsing may improve grammar checking.Therefore, this research aims at improving the ranking performance of theLink Parser, and as a result the parser is more "confident" in checking gram-maticality. In analyzing the performance of the parser, I have been trying to find out how to make the parser to be more powerful to select with phrase structure knowledge involved the (near-) correct linkage out of the potential candidate linkages for a sentence.Combined models have been adopted by researchers to fully exploit the advantages of cooperating systems. It is widely discussed in areas of linguistic theory and practice. In research areas of part-of-speech tagging, parsing and grammar checking, combined models have achieved great success.This research combines link grammar parsing with a phrase structure parsing model — transformation-based machine learning algorithm. As discussed in greater detail in 2.4.2.2, a sentence is considered by link grammar to be proper if it satisfies three conditions:1. Planarity: The link arcs above words do not cross.2. Connectivity: The links suffice to connect all the words of the sequence together.3. Satisfaction: The links satisfy the linking requirements of each word in the sequence.This process of judging the grammaticality of a sentence can be deemed one of the major differences between link grammar parsing and other methods. Once these criteria are satisfied, the Link Parser determines the sentence to be grammatical. With the null-link system incorporated into the parser proper, errors in an ungrammatical sentence —judged by the parser, of course — can be located. And this is the very rationale of the present research — a grammar checker based on the Link Parser.The above-mentioned deficiency of the Link Parser can be partially resolved by introducing transformation-based machine learning algorithm — a phrase structure parsing model. Then in the end, I shall bring about a grammar checking program based on the combined model. With the combined model, link grammar and transformation-based machine-learning parsing algorithm are successfully integrated to form an application-oriented grammar checking system. In the grammar checking system, the input is first analyzed by the Link Parser module. If ungrammaticality is detected for the input, the analysis is provided to the user as it is;if the input sentences are deemed grammatical, the transformation-based machine-learning parsing module is activated and the linkages generated by the Link Parser are re-ranked by the parse of this module. The re-ranked linkages will then be output to the user.The combined model system is superior to the Link Parser alone in the following aspect: The linkages generated by the Link Parser are re-ranked according to the transformation-based machine-learning parsing module's parse so that the combined model grammar checking system is able to yield more reasonable feedback when determining grammaticality of the sentences being checked. The superiority of the present system over the Link Parser in the ranking of linkages can be verified by the x2 *est in the experiment conducted in the present research. And moreover,with an evaluation experiment conducted with CET4 essays taken from ST3 of the CLEC Corpus1, the agreement of the grammaticality judgment of the grammar checker with humans is found to be 90.4%. This result suggests that the grammaticality judgment of the grammar checker agrees to much extent with human judgment and thepracticality of the grammar checker has been manifested.