Compositional Matrix-Space Models: Learning Methods and Evaluation

There has been a lot of research on machine-readable representations of words for natural language processing (NLP). One mainstream paradigm for the word meaning representation comprises vector-space models obtained from the distributional information of words in the text. Machine learning techniques have been proposed to produce such word representations for computational linguistic tasks. Moreover, the representation of multi-word structures, such as phrases, in vector space can arguably be achieved by composing the distributional representation of the constituent words. To this end, mathematical operations have been introduced as composition methods in vector space. An alternative approach to word representation and semantic compositionality in natural language has been compositional matrix-space models. In this thesis, two research directions are considered. In the first, considering compositional matrix-space models, we explore word meaning representations and semantic composition of multi-word structures in matrix space. The main motivation for working on these models is that they have shown superiority over vector-space models regarding several properties. The most important property is that the composition operation in matrix-space models can be defined as standard matrix multiplication; in contrast to common vector space composition operations, this is sensitive to word order in language. We design and develop machine learning techniques that induce continuous and numeric representations of natural language in matrix space. The main goal in introducing representation models is enabling NLP systems to understand natural language to solve multiple related tasks. Therefore, first, different supervised machine learning approaches to train word meaning representations and capture the compositionality of multi-word structures using the matrix multiplication of words are proposed. The performance of matrix representation models learned by machine learning techniques is investigated in solving two NLP tasks, namely, sentiment analysis and compositionality detection. Then, learning techniques for learning matrix-space models are proposed that introduce generic task-agnostic representation models, also called word matrix embeddings. In these techniques, word matrices are trained using the distributional information of words in a given text corpus. We show the effectiveness of these models in the compositional representation of multi-word structures in natural language. The second research direction in this thesis explores effective approaches for evaluating the capability of semantic composition methods in capturing the meaning representation of compositional multi-word structures, such as phrases. A common evaluation approach is examining the ability of the methods in capturing the semantic relatedness between linguistic units. The underlying assumption is that the more accurately a method of semantic composition can determine the representation of a phrase, the more accurately it can determine the relatedness of that phrase with other phrases. To apply the semantic relatedness approach, gold standard datasets have been introduced. In this thesis, we identify the limitations of the existing datasets and develop a new gold standard semantic relatedness dataset, which addresses the issues of the existing datasets. The proposed dataset allows us to evaluate meaning composition in vector- and matrix-space models.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.