Hidden Markov Model Parameters Estimation For Part-of-Speech Tagging

Part-of-Speech Tagging is the fundamental problems in natural language processing. Because it requires no manual rules of natural language and has a high level of accuracy, the statistical language model has gradually become a hot topic. For its better performance, Hidden Markov Model (HMM), one of the statistical models, has been the recent trend in Part-of-Speech Tagging.In a HMM, parameters are very important constituent parts. As a result, parameters estimation becomes a premise to build a HMM. Based on the work of previous analysis of current HMM researches, two improvements for the supervised parameters estimation are presented in this paper. One is adding previous tag while estimating the words' output probability; the other is adjusting the HMM parameters using Perceptron Algorithm.Contrary to the Output Dependency Assumption of a traditional HMM, which is that the probability of a word depends only on its own tag, we assume that the probability of a word depends not only on its own tag, but also on the previous tag while estimating the words' output probability. By doing this, we can employ more grammatical information in the HMM.The Perceptron Algorithm is one of the algorithms that can adjust the parameters of a model by comparing the results produced by the model and the expected output results. In this paper, this algorithm is used to improve the quality of the HMM parameters. At first, we use Viterbi algorithm to get the words and their tags, then, by comparing these results and the correct ones, we can optimize the parameters if there is any difference.With the parameters produced by the improved HMM Parameters Estimation Model, in close tests, the F-value of the HMM can arrive at 96.78%, which is 1.84% higher than that of the traditional HMM with the unimproved ones, in open tests, the F-value of the HMM can arrive at 92.79%, which is 3.44% higher than that of the traditional HMM with the unimproved ones. The results of the experiments also show that the quality of the HMM parameters can improve a lot, by modifying the Output Dependency Assumption and adjusting the HMM parameters with Perceptron Algorithm.