APPLICATION OF LINK GRAMMAR IN SEMI-SUPERVISED NAMED ENTITY RECOGNITION FOR ACCIDENT DOMAIN

Accident document typically contains some crucial information that might be useful for analysis process for future accident investigation i.e. date and time when the accident happened, location where the accident occurred and also the person involved in the accident. This document is largely available in free text; it can be in the form of news wire articles or accident reports. Although it is possible to identify the information manually, due to the high volumes of data involved, this task can be time consuming and prone to error. Information Extraction (IE) has been identified as a potential solution to this problem. IE has the ability to extract crucial information from unstructured texts and convert them into a more structured representation. This research is attempted to explore Name Entity Recognition (NER), one of the important tasks in IE research aimed to identify and classify entities in the text documents into some predefined categories. Numerous related research works on IE and NER have been published and commercialized. However, to the best of our knowledge, there exists only a handful of IE research works that are really focused on accident domain. In addition, none of these works have attempted to either explore or focus on NER, which becomes the main motivation for this research. The work presented in this thesis proposed an NER approach for accident documents that applies syntactical and word features in combination with Self-Training algorithm. In order to satisfy the research objectives, this thesis comes with three main contributions.
The first contribution is the identification of the entity boundary. Entity segmentation or identification of entity boundary is required since named entity may consist of one or more words. We adopted Stanford Part-of-Speech (POS) tagger for the word POS tag and connectors from the Link Grammar (LG) parser to determine the starting and stopping word. The second contribution is the extraction pattern construction. Each named entity candidate will be assigned with an extraction pattern constructed from a set of word and syntactical feature. Current NER system used
restricted syntactical features which are associated with a number of limitations. It is therefore a great challenge to propose a new NER approach using syntactical features that could capture all syntactical structure in a sentence. For the third contribution, we have applied the Self-Training algorithm which is one of the semi-supervised machines learning technique. The algorithm is utilized for predicting a huge set of unlabeled data, given a small number of labelled data. In our research, extraction pattern from the first module will be fed to this algorithm and is used to make the prediction of named entity candidate category. The Self-Training algorithm greatly benefits semi-supervised learning which allows classification of entities given only a small-size of labelled data. The algorithm reduces the training efforts and generates almost similar result as compared to the conventional supervised learning technique. The proposed system was tested on 100 accident news from Reuters to recognize three different named entities: date, person and location which are universally accepted categories in most NER applications. Exact Match evaluation method which consists of three evaluation metrics; precision, recall and F-measure is used to measure the proposed system performance against three existing NER systems. The proposed system has successfully outperforms one of those systems with an overall F-measure of approximately 9% but in the other hand it shows a slight decrease as compared to other two systems identified in our benchmarking. However, we believe that this difference is due to the different nature and techniques used in the three systems. We consider our semi-supervised approach as a promising method even though only two features are utilized: syntactical and word features. Further manual inspection during the experiments suggested that by using complete word and syntactical features or combination of these features with other features such as the semantic feature, would yield an improved result.