Subject Area

Statistics

Abstract

In this study, our main objective is to tackle the black-box nature of popular machine learning models in sentiment analysis and enhance model interpretability. We aim to gain more insight into the decision-making process of sentiment analysis models, which is often obscure in those complex models. To achieve this goal, we introduce two word-level sentiment analysis models.

The first model is called the attention-based multiple instance classification (AMIC) model. It combines the transparent model structure of multiple instance classification and the self-attention mechanism in deep learning to incorporate the contextual information from documents. As demonstrated by a wine review dataset application, AMIC can achieve state-of-the-art performance compared to a number of machine learning methods, while providing much improved interpretability.

The second model, AMIC 2.0, improves AMIC in two key aspects. Notably, AMIC is limited in integrating positional information in text because it ignores the order of words in documents. AMIC 2.0 comes up with a novel approach to incorporate relative positional information in the self-attention mechanism, enabling the model to capture more accurate sentiment that is position-sensitive. This modification enables the model to better understand how word order and proximity influence sentiment expressions. Secondly, AMIC 2.0 takes a step further by decomposing the sentiment score in AMIC into a context-independent score and a context-dependent score. This decomposition, along with the incorporation of two sentiment shifters linking these scores in a global environment and a local environment of text respectively, elucidate how context of document influences sentiment of words, leading to more interpretable results in sentiment analysis.

The utility of AMIC 2.0 is demonstrated by an application to a Twitter dataset. AMIC 2.0 has improved the overall performance of AMIC, with the additional capability of handling more intricate language subtleties, such as different types of negations. Both AMIC and AMIC 2.0 are trained without having to use pre-trained sentiment word dictionary or seeded sentiment words. Compared to some other big language models, their computation cost is relatively low and they are versatile to use conventional datasets to generate domain-specific sentiment dictionary and provide interpretable sentiment analysis results.

Degree Date

Winter 12-16-2023

Document Type

Dissertation

Degree Name

Ph.D.

Department

Statistics and Data Science

Advisor

Jing Cao

Number of Pages

81

Format

.pdf

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Share

COinS