LVQ algorithm applied to natural language processing tasks

María Teresa Martín Valdivia. May 2004

Both Natural Language Processing (NLP) and Artificial Neural Networks (ANN) are two key areas in Artificial Intelligence. However, despite the large amount of work in both disciplines, attempts to combine them have been scarce.

On the one hand, the studies that include machine learning in NLP systems are numerous, and secondly, ANNs have been applied to a number of problems with similar features to those of PLN. Interestingly, however, the number of studies that use ANN in PLN systems is very small. This is more surprising when the results of the few existing studies show that the use of a neural approach is a good alternative method for building PLN systems based on learning.

The main purpose of this thesis is to demonstrate that it is possible to take advantage of features of ANNs to successfully address the development and implementation of systems that deal with language automatically.

To do this, a common formalism based on a neural model for solving various NLP tasks is proposed. Specifically three tasks will be discussed:
• Text categorization
• Resolution of lexical ambiguity
• Information retrieval.

While for the first two tasks complete systems will be developed, for information retrieval two specific issues related to these types of systems will be addressed:
• The recognition of multiword terms
• Collection fusion

The first problem is examined from a monolingual perspective while the second will be addressed to a multilingual environment.
The neural scheme used is based on the Kohonen model and more specifically his supervised release: the learning algorithm or algorithm for vector quantization LVQ (Learning Vector Quantization). It will be shown that this algorithm can be adapted to solving real applications of natural language processing, presenting it as a robust, flexible and effective method. Experiments show that the LVQ algorithm is easily adapted to the different scenarios used and the results obtained are comparable, and in many cases outperform the traditional methods used to solve each of the problems studied.

(Link TESEO)