Author ORCID Identifier
Biosketch
D.Rekha received her B.Sc degree in Mathematics from Bharathidasan University in 2002. She received the M.Sc degree in Information Technology from Bharathidasan University in the year 2004. She has been with SASTRA University since 2006. She has published papers in international journals. Her research interests include Natural Language Processing, Machine Learning & Deep Learning, Theoretical Computer Science, and Computational Linguistics.
She completed her Bachelor of Science degree in Mathematics in 2002 and subsequently earned her Master of Science degree in Information Technology in 2004. With a strong academic foundation that combines mathematical reasoning and computing technologies, she developed a keen interest in interdisciplinary research areas involving language technologies, intelligent systems, and computational models.
Since 2006, she has been serving at SASTRA University, where she has actively contributed to teaching, research, and academic development in the field of Computer Science and Engineering. Over the years, she has guided undergraduate students and participated in various academic and research-oriented activities. Her dedication to teaching and research has enabled her to build expertise in emerging areas of Artificial Intelligence and data-driven computing methodologies.
D. Rekha has published several research papers in reputed international journals and conferences. Her primary research interests include Natural Language Processing, Machine Learning, Deep Learning, Theoretical Computer Science, Computational Linguistics, and Digital Document Analysis. She has also worked on applications of intelligent systems for text analytics, pattern recognition, and language understanding.
Date of Award
17-8-2025
Document Type
Thesis
School
School of Computing
Programme
Ph.D.-Doctoral of Philosophy
First Advisor
Dr.V.Ramasamy
Keywords
Natural language Processing, Logistic Regression, Token Generation, Document Analysis, Digital Document Parsing
Abstract
Digital document dependency parsing is a significant task in natural language processing. Dependency parsing supports verification of the grammatical correctness of a sentence besides enabling extraction of relevant documents. Inappropriate extraction of features may result in falsely parsing a document, leading to decreased accuracy. Machine learning methods have been employed to perform feature extraction. However, selecting pertinent features was never achieved which minimizes the time consumption and overhead.
Hence, novel machine learning and deep learning techniques have been designed in our work for accurate and computationally efficient digital document analytics through dependency parsing. Four different contributions have been proposed for digital document analytics. A linear approach is designed for automatic analysis of English text. Sentences are initially preprocessed via tokenization, syntactical analysis and stemming word elimination. This is followed by grammatical analysis for parsing the documents. The grammatical analysis tool tokenizes and patterns are applied to rearrange the words which results in better analysis with higher accuracy.
It is observed that linear approach increases accuracy by 4%, precision by 22%, recall by 22% and less time by 19% and overhead by 7% when compared to existing methods. Logistic Regression and Deep Dependency Parsing (LRDDP) is developed for digital document analytics. Efficient feature xv representation using Logistic Regressive Token generation model is carried out by partitioning the documents into important features. Naperian logarithms are used in this model to generate robust tokens with minimum complexity. The probability of each class is computed and based on the probability outcomes, efficient feature representation is achieved in a computationally efficient manner.
A deep transition-based dependency parser is employed for parsing digital documents. Two-layer perceptron using ReLU activation function computes the dependency score of the textual distances to obtain efficient and accurate parsing. LR-DDP enhances the digital document parsing accuracy by 18%, parsing time by 41%, and overhead by 22% compared to existing methods. Weighted Score Convolutional Network and Arc-factored Graph-based Dependency Parsing (WSCN-AGDP) models are next developed for document parsing with less falsification. Important features are extracted by convoluting the baseline embedding vector obtained from the input dataset. Spearman Correlated Arc-Factored Graph-based Dependency Parsing is designed to parse digital documents precisely by applying the scoring function.
This results in precise and accurate digital document parsing with less false positive rate. Experimental results of WSCN-AGDP model exhibit that precision is increased by 20%, parsing time is reduced by 30%, overhead by 22% and false positive rate by 38%. Homing Sequence and Deep Convolution Finite State Automatabased Dependency Parsing (HSDCFS-ADP) technique has been introduced for accurate digital document analysis with finite memory. In order to guarantee precise and timely digital document analysis, feature extraction and dependency parsing are carried out. A Deterministic Finite State Automata-based Relevant Data Extraction model is being employed for extracting robust features. Homing xvi sequence is determined for each word in the sentence and extracts pertinent data.
Deep Convolution Neural Network is applied in order to perform accurate dependency parsing. This in turn increases precision rate while retrieving relevant documents. It is observed that HSDCFS-ADP achieves 14% precision, 19% accuracy and 13% true positive rate. Also, time is reduced by 52% and overhead by 29% when compared with the existing techniques.
Recommended Citation
D, Rekah Ms, "Investigation of Dependency Parsing Techniques for Digital Document Analysis through Deep Learning Approach" (2025). Theses and Dissertations. 191.
https://knowledgeconnect.sastra.edu/theses/191