Abstract

The unprecedented expansion of user-generated content in recent years demands more attempts of information filtering in order to extract high-quality information from the huge amount of available data. In this dissertation, we begin with a focus on topic detection from microblog streams, which is the first step toward monitoring and summarizing social data. Then we shift our focus to the authorship attribution task, which is a sub-area of computational stylometry. It is worth mentioning that determining the style of a document is orthogonal to determining its topic, since the document features which capture the style are mainly independent of its topic. We initially present a frequent pattern mining approach for topic detection from microblog streams. This approach uses a Maximal Sequence Mining (MSM) algorithm to extract pattern sequences, where each pattern sequence is an ordered set of terms. Then we construct a pattern graph, which is a directed graph representation of the mined sequences, and apply a community detection algorithm to group the mined patterns into different topic clusters. Experiments on Twitter datasets demonstrate that the MSM approach achieves high performance in comparison with the state-of-the-art methods. For authorship attribution, while previously proposed neural models in the literature mainly focus on lexical-based neural models and lack the multi-level modeling of writing style, we present a syntactic recurrent neural network to encode the syntactic patterns of a document in a hierarchical structure. The proposed model learns the syntactic representation of sentences from the sequence of part-of-speech tags. Furthermore, we present a style-aware neural model to encode document information from three stylistic levels (lexical, syntactic, and structural) and evaluate it in the domain of authorship attribution. Our experimental results, based on four authorship attribution benchmark datasets, reveal the benefits of encoding document information from all three stylistic levels when compared to the baseline methods in the literature. We extend this work and adopt a transfer learning approach to measure the impact of lower-level linguistic representations versus higher-level linguistic representations on the task of authorship attribution. Finally, we present a self-supervised framework for learning structural representations of sentences. The self-supervised network is a Siamese network with two components; a lexical sub-network and a syntactic sub-network which take the sequence of words and their corresponding structural labels as the input, respectively. This model is trained based on a contrastive loss objective. As a result, each word in the sentence is embedded into a vector representation which mainly carries structural information. The learned structural representations can be concatenated to the existing pre-trained word embeddings and create style-aware embeddings that carry both semantic and syntactic information and is well-suited for the domain of authorship attribution.

Notes

If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu.

Graduation Date

2021

Semester

Fall

Advisor

Hua, Kien

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Degree Program

Computer Science

Format

application/pdf

Identifier

CFE0008855; DP0026134

Language

English

Release Date

December 2021

Length of Campus-only Access

None

Access Status

Doctoral Dissertation (Open Access)

Share

COinS