2018-05-24

In the era of data-driven industry, the unstructured text, which is generated by human cognition skills, remains the main data format with a massive amount being generated from different sources of technology. The problem we are handling in our work is: How can machine more efficiently learn, represent, and grow semantics from unstructured text, as the written form of natural language? We propose a cognitive model, ICAN-2, inspired by human cognition skills, to learn/extract and represent semantics from text. ICAN-2 is an improved version of the ICAN cognitive model of semantic memory. The ICAN model is from Incremental Construction of an Associative Network model, and it aims at computationally modeling the development of semantic associations in the human semantic memory. Both the ICAN and ICAN-2 models use semantic-graphs to represent semantics. The traditional and yet widely used text representation model is the Vector Space Model (VSM), also known as Bag-Of-Words (BOW) model, in which documents are represented simply by n dimensional feature-vectors. The most widely used term weighting scheme in VSM model is called Term-Frequency/Inverse-Document-Frequency (TF/IDF), which is also used in the latent semantic analysis (LSA) model of semantic learning and representation. Both VSM-based approaches of the TF/IDF and LSA have some notable limitations such as neglecting the word order and other dependency relations among the terms appearing in the original text documents. The ICAN-2 model is an alternative cognitive-graph based model for the traditional VSM model of text representation. After a detailed survey of related works, the performance of the ICAN-2 model is compared against the two most closely related models of semantic modeling in the literature: (1) the LSA model as a cognitive model that has been applied in different text-mining tasks and is an alternative for the TF/IDF technique and (2) the ICAN model which is the seed model for our work and technically the most related model to our model. Experimental results show the ICAN-2 model outperforming both the semantic learning and representation models LSA and ICAN. Then we statistically analyze the cognitive-graphs generated by the ICAN-2 model to explore their structural characteristics and their meanings.

2018

PhD in Computer Science: Using and Improving Computational Cognitive Models for Graph-Based Semantic Learning and Representation from Unstructured Text with Applications