Unsupervised Clustering of Structured and Unstructured Text Collections

This thesis explores the application of unsupervised clustering for domain adaptation of machine translation systems. As in many artificial intelligence areas, creating a system that generalizes to any domain is a hard problem in machine translation. Domain adaptation, in contrast, aims to specialize a generic (or otherwise intended) system for a particular domain and translate text within that domain better. In this thesis, experiments on using unsupervised learning as a first step in solving this problem are explored, posing the research questions a) how unstructured data could be used for domain adaptation and b) how a bespoke translation of an input document could be provided. In the first part of the thesis, background theory is presented and related work reviewed. In the second, experimental part, preliminary experiments on building n-gram models and multiword expression detection are presented before experiments into clustering of structured and unstructured document collections are conducted. Finally, the parts are brought together in experiments on using these input factors for domain adaptation of machine translation systems, with end-to-end evaluation. Some of the clusters identified in the clustering experiments on unstructured web collections were used as auxiliary language models in machine translation, in the experiments on domain adaptation. Self-Organizing Maps are used in the first phase of unsupervised clustering before a hierarchical agglomerative clustering algorithm is applied to extract tangible clusters from the map, with the number of clusters determined by the knee method. By creating a mapping between the input document and one of the auxiliary language models, translation is aided by this language model. Using the language model perplexity on the input documents to select the auxiliary language model for domain adaptation links the clusters to the translation process. Results show that the performance according to metrics such as BLEU, TER, and Meteor were on-par, and in some cases better than the results from leveraging all the available supplementary text as an auxiliary language model. The difference when using different auxiliary LM could be up to 1 BLEU points and 0.9 Meteor points