Time expression and named entity recognition for sentiment analysis

• Main Author: Tan, Jordan Rei Yao
• Other Authors: Erik Cambria
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Document and text processing
• Online Access: https://hdl.handle.net/10356/157140
• Institution: Nanyang Technological University

Sentiment analysis has become an important area of natural language processing (NLP). Today, when it comes to sentiment analysis tasks, sub-symbolic artificial intelligence (AI) approaches such as neural networks and deep learning are much more popular and widely used compared to classic symbolic AI approaches. Despite this, there have also been attempts to bridge the gap between the two approaches and integrate them. The Sentic Computing framework is a novel approach that leverages on the strengths of both symbolic and sub-symbolic AI for sentiment analysis. It identifies the core concepts in text using linguistic patterns, generalizes them through deep learning, and identifies the polarities associated with them in a knowledge base to determine the overall text sentiment. This ensemble application taps on the strengths of both top-down and bottom-up learning, and gives machines some logical reasoning ability for the natural language. However, it does not rely on any prior training for polarity prediction on a text. On the other hand, sub-symbolic approaches learn the polarity of text based on identifying patterns in training data. Hence, they are able to take the particular context or lingo of a text type into account when predicting its polarity. Understanding this difference, one aim of this research is to explore how Sentic Computing—an ensemble application of symbolic and sub-symbolic AI—compares with popular sub-symbolic approaches when it comes to sentiment analysis. In addition, it is noted that most research on sentiment analysis is focused on analysis at the sentence-level. However, most of the time, we wish to understand what is the entity mentioned in a text that a positive, neutral, or negative sentiment is for. This task gets more complicated when a text contains multiple named entities, with different sentiments for each of them. This highlights the importance of named entity recognition for sentiment analysis. Therefore, in this research, we also propose a novel method to perform entity-level sentiment analysis. The method looks at the grammatical structure of sentences to extract named entities as well as their corresponding descriptions, in order to identify the sentiments for each of them. Experimental results show that our proposed method for entity-level sentiment analysis yields more insights compared to traditional sentence-level sentiment analysis.