Analysis of Text-Semantics via Efficient Word Embedding using Variational Mode Decomposition

Analysis of Text-Semantics via Efficient Word Embedding using Variational Mode Decomposition Rohith Ramakrishnan, Anirudh Vadakedath, U Vamsi Krishna, Premjith B and K.P Soman Centre for Computational Engineering and Networking Amrita School of Engineering Coimbatore Amrita Vishwa Vidyapeetham b premjith@cb.amrita.edu Abstract In this paper, we propose a novel method which establishes a newborn relation between Signal Processing and Natural Language Pro- cessing (NLP) method via Variational Mode Decomposition (VMD). Unlike the modern Neural Network approaches for NLP which are complex and often masked from the end user, our approach involving Term Fre- quency - Inverse Document Frequency (TF- IDF) aided with VMD dials down the com- plexity retaining the performance with trans- parency. The performance in terms of Ma- chine Learning based approaches and seman- tic relationships of words along with the methodology of the above mentioned ap- proach is analyzed and discussed in this paper. 1 Introduction Regardless of numerous years of research and devel- opment in Natural Language Processing, the models developed for various applications are primitive with limited linguistic knowledge. Over the years of dis- pute over how best to improve the NLP models, nu- merous approaches and techniques were proposed. During this period, the amount of text data for NLP research sky-rocketed. However, such an increase in the corpus didn’t help NLP models developed for several languages and applications to escalate at the same rate. Distinguishing the local and global scope of words is a crucial part of the NLP problems. It is because a word can possess different senses accord- ing to the contexts. The local scope of the word may be entirely different from its global scope. This in- crease in the ambiguity in perceiving the sense of words increases the complexity of the NLP tasks. In addition to these hassles, the major obstacle is deriving synonyms, semantic and lexical relation- ship between words. It has been a holy principle of NLP to evaluate the synonymity via annotation and semantical-labels. Even with the introduction of numerous neural network-based text classifica- tion such as BERT and fastText, the above-stated problems haven’t been solved. TF-IDF was one of the numerous techniques sur- faced to generate a vectorial representation for the text data (Sparck Jones, 1972) (Barathi Ganesh et al., 2017). TF-IDF quantifies the significance of a word to a document (Ullman, 2011), by computing a weighting factor, which helps to identify the most frequently occurring words in the corpus. Another popular and most sought after model is fastText. It is a word embedding algorithm developed by Face- book’s AI Reach lab (FAIR) for learning word em- bedding and text classification. fastText was able to retrieve character-level information for word repre- sentations and sentence classification. For example, considering the word ”bunny” and n=3, the n-grams would be <bu, bun, unn, nny, ny>. The symbols < and > are added so that n-grams are not mixed with shorter words such as <bun> in this case. It helps to generate meaningful representations for suffixes and prefixes. Like fastText, GloVe (Global …

Content truncated to 500 words