Augmenting Semantic Lexicons Using Word Embeddings and Transfer Learning.

Area of Science:

• Natural Language Processing
• Computational Linguistics

Background:

• Sentiment-aware intelligent systems rely on language models, primarily lexicon-based and contextual.
• Lexicon-based models offer interpretability but require continuous lexicon expansion.
• Automatic expansion is crucial for maintaining the effectiveness of lexicon-based sentiment analysis.

Purpose of the Study:

• To propose and evaluate two novel models for automatic lexicon expansion in sentiment analysis.
• To address the challenge of updating lexicons with new words and expressions.
• To provide a cost-effective solution for lexicon enrichment.

Main Methods:

• A baseline model using a shallow neural network with pre-trained word embeddings (non-contextual).
• An advanced model employing a deep Transformer network that incorporates word definitions.
• Both models aim to estimate the lexical polarity of new words.

Main Results:

• Both proposed models achieved accuracy comparable to human annotators (Amazon Mechanical Turk).
• The Transformer-based model demonstrates improved performance over the baseline.
• The automatic methods significantly reduce the cost of lexicon expansion.

Conclusions:

• Automatic lexicon expansion is feasible and accurate for sentiment analysis.
• The developed models offer a scalable and economical approach to maintaining sentiment lexicons.
• These methods enhance the utility of lexicon-based sentiment analysis systems.