Two-Level LSTM for Sentiment Analysis With Lexicon Embedding and Polar Flipping

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Sentiment analysis is crucial for text mining, but high-quality labeled training data is difficult to obtain due to subjective annotations and complex sentiment expressions.
• Existing sentiment classification methods often rely on pre-existing, high-quality datasets, which are challenging to construct in real-world scenarios.

Purpose of the Study:

• To develop a robust sentiment classification method that overcomes the limitations of traditional approaches, particularly concerning training data acquisition.
• To propose a novel labeling strategy and a two-level long short-term memory network for improved sentiment analysis.

Main Methods:

• A new labeling strategy was employed to construct high-quality datasets.
• A two-level long short-term memory (LSTM) network was utilized for sentiment classification.
• A novel ρ-hot encoding strategy was introduced to effectively incorporate lexical cues, addressing limitations of one-hot encoding.
• A flipping model was developed to account for context-dependent word polarity shifts.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art algorithms.
• Experiments were conducted on benchmark English datasets and newly compiled Chinese datasets.
• The effectiveness of the ρ-hot encoding and flipping model in handling lexical cues and sentiment polarity variations was validated.

Conclusions:

• The developed methodology offers a significant advancement in sentiment analysis, particularly in overcoming data labeling challenges.
• The proposed approach, incorporating a new labeling strategy, two-level LSTM, ρ-hot encoding, and a flipping model, achieves state-of-the-art results.
• This work contributes to more accurate and robust sentiment classification, especially for languages like Chinese where data annotation can be complex.