Use of a genetic algorithm in brill's transformation-based part-of-speech tagger

Abstract

The tagging problem in natural language processing is to find a way to label every word in a text as a particular part of speech, e.g., proper noun. An effective way of solving this problem with high accuracy is the transformation-based or "Brill" tagger. In Brill's system, a number of transformation templates are specified a priori that are instantiated and ranked during a greedy searchbased algorithm. This paper describes a variant of Brill's implementation that instead uses a genetic algorithm to generate the instantiated rules and provide an adaptive ranking. Based on tagging accuracy, the new system provides a better hybrid evolutionary computation solution to the part-of-speech (POS) problem than the previous attempt. Although not able to make up for the use of a priori knowledge utilized by Brill, the method appears to point the way for an improved solution to the tagging problem.

Evolutionary TBL template generation

“…The use of genetic algorithms in conjunction with TBL has already been examined in [13], where they are used in the TBL training process to generate the instantiated rules and to provide an adaptive ranking. Nevertheless, they have not been used in the evaluation of template sets what is our proposal.…”

An Approach to the POS Tagging Problem Using Genetic Algorithms

“…The first group is based on statistical data concerning the different context possibilities for a word [2,3,4,5,6,7], while the second group is based on rules, normally designed by human experts, that capture the language properties [8,9,10].…”

“…These approaches can also be divided by the type of information used to solve the problem, statistical information [3,4,5,6,7], and rule-based information [9]. Shortly, in the former, an evolutionary algorithm is used to assign the most likely tag to each word of a sentence, based on a context table, that basically has the same information that is used in the traditional probabilistic approaches.…”

“…In this case a genetic algorithm (GA) is used to evolve a set of transformations rules, that will be used to tag a text in much the same way as the tagger proposed by Brill. While in [3,4,5,6,7] the evolutionary algorithm is used to discover the best sequence of tags for the words of a sentence, using an information model based on statistical data, in [9] the evolutionary algorithm is used to evolve the information model, in the form of a set of transformation rules, that will be used to tag the words of a sentence.…”

“…These rules contemplate, not only context information, but also some information about the words' morphology. They are not oriented toward error correction, like in [8,9], instead they are a form of classification rules, which try to generalize the context information that is used by probabilistic taggers.…”