Representing semantic relatedness

Publication Type:
Thesis
Issue Date:
2016
Full metadata record
Files in This Item:
Filename Description Size
01front.pdf237.13 kB
Adobe PDF
02whole.pdf1.65 MB
Adobe PDF
To do text mining, the first question we must address is how to represent documents. The way a document is organised reflects certain explicit and implicit semantic and syntactical coupling relationships which are embedded in its contents. The effective capturing of such content couplings is thereby crucial for a genuine understanding of text representations. It has also led to the recent interest in document similarity analysis, including semantic relatedness, content coverage, word networking, and term-term couplings. Document similarity analysis has become increasingly relevant since roughly 80% of big data is unstructured. Accordingly, semantic relatedness has generated much interest owing to its ability to extract coupling relationships between terms (words or phrases). Existing work has focused more on explicit couplings and this is reflected in the models that have been built. In order to address the research limitations and challenges associated with document similarity analysis, this thesis proposes a semantic coupling similarity measure and the hierarchical tree learning model to fully enrich the semantics within terms and documents, and represent documents based on the comprehensive couplings of term pairs. In contrast to previous work, the models proposed can deal with unstructured data and terms that are coupled for various reasons, thereby addressing natural language ambiguity problems. Chapter 3 explores the semantic couplings of pairwise terms by involving three types of coupling relationships: (1) intra-term pair couplings, reflecting the explicit relatedness within term pairs that is represented by the relation strength over probabilistic distribution of terms across document collection; (2) the inter-term pair couplings, capturing the implicit relatedness between term pairs by considering the relation strength of their interactions with other term pairs on all possible paths via a graph-based representation of term couplings; and finally, (3) semantic coupling similarity, which effectively combine the intra- and inter-term couplings. The corresponding term semantic similarity measures are then defined to capture such couplings for the purposes of analysing term and document similarity. This approach effectively addresses both synonymy (many words per sense) and polysemy (many senses per word) in a graphical representation, two areas that have up until now been overlooked by previous models. Chapter 4 constructs a hierarchical tree-like structure to extract highly correlated terms in a layerwise fashion and to prune weak correlations in order to maintain efficiency. In keeping with the hierarchical tree-like structure, a hierarchical tree learning method is proposed. The main contributions of our work lie in three areas: (1) the hierarchical tree-like structure featuring hierarchical feature extraction and correlation computation procedures whereby highly correlated terms are merged into sets, and these are associated with more complete semantic information; (2) each layer is a maximal weighted spanning tree to prune weak feature correlations; (3) the hierarchical treelike structure can be applied to both supervised and unsupervised learning approaches. In this thesis, the tree is associated with Tree Augmented Naive Bayes (TAN) as the Hierarchical Tree Augmented Naive Bayes (HTAN). All of these models can be applied in the text mining tasks, including document clustering and text classification. The performance of the semantic coupling similarity measure is compared with typical document representation models on various benchmark data sets in terms of document clustering and classification evaluation metrics. These models provide insightful knowledge to organise and retrieve documents.
Please use this identifier to cite or link to this item: