Enhancing Financial Sentiment Analysis with Lexicon-Based Prompts

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Jun 24 2024

In an article published in the journal Expert System with Applications, researchers proposed an innovative lexicon-based prompt method for financial sentiment analysis, specifically in domains lacking labeled data. This method introduced domain-specific lexicons to correct mispredicted words, improving sentiment detection without applying domain-adaptation approaches. It can be integrated with domain-adaptation approaches to improve performance.

Background

Sentiment analysis, a major application of natural language processing (NLP), is based on identifying the emotional tone of the text, usually categorized by valence (positive or negative) and arousal (excited or calm). In the financial domain, sentiment analysis has applications in areas like financial risk prediction, investment decision-making, and stock market trading. Traditionally, these tasks have utilized categorical approaches, which classify sentiments into discrete categories. However, categorical approaches often fail to capture the nuanced emotional states present in financial texts.

Dimensional approaches can provide more granular, continuous sentiment analysis. These methods represent sentiments along continuous scales, offering finer detail than categorical methods. However, applying dimensional sentiment analysis in the financial domain poses challenges due to the scarcity of domain-specific data. Financial texts often contain specialized vocabulary and references, such as "Tulip mania" or "Bretton Woods Agreement," which are not adequately addressed by models trained on general domain corpus. As a result, sentiment models pre-trained on general text data often underperform when applied to financial texts.

To bridge this gap, two primary approaches were explored, pre-training language models on large financial text corpus and domain adaptation methods. Pre-training requires considerable computing resources, while domain adaptation attempts to adjust models trained on general data to perform well in the financial domain. Despite their potential, traditional domain adaptation methods have not fully overcome the domain-specific challenges, as they often fail to accurately capture the specialized financial vocabulary and context.

This paper addressed these shortcomings by proposing a lexicon-based prompt method specifically for financial sentiment analysis. Leveraging the richness of existing categorical sentiment lexicons in the financial domain, this method integrated domain-specific lexicons into pre-trained language models to correct mispredictions. This method did not require extensive labeled data and could be easily combined with other domain-adaptation strategies to improve the model's validity and accuracy.

Lexicon-based Prompt Method

The lexicon-based prompt method aimed to enhance the performance of pre-trained language models in financial sentiment analysis by utilizing domain-specific lexical information. This approach addressed the gap in financial domain knowledge that often resulted in poor performance of general pre-trained models. The method involved two steps: dimensional fine-tuning and prompt fine-tuning, both utilizing the same pre-trained language model weights.

In the first step, dimensional fine-tuning, the pre-trained language model was fine-tuned to predict sentiment values accurately. This process used the mean square error (MSE) as the loss function to optimize the model's weights for sentiment prediction. The goal was to find the best weight configuration that minimized the error between the predicted and actual sentiment values for a given text.

Once the model was fine-tuned for general domain data, the second step, prompt fine-tuning, was employed to adapt the model further to the financial domain. This step required creating prompt sentences using a financial lexicon and a prompt template. The financial lexicon contained domain-specific words annotated with sentiment labels (e.g., positive or negative). The prompt template transformed this lexical information into training sentences that help the model better understand and predict financial sentiments.

The method generated sentences where the lexicon's words were integrated into prompts, thus correcting mispredictions and refining the model's understanding of financial terminology. By leveraging the rich lexical information specific to the financial domain, this method significantly improved the model's accuracy and performance in financial sentiment analysis.

Experimental Evaluation of Lexicon-Based Prompt Method

The lexicon-based prompt method was evaluated against traditional domain adaptation techniques like domain-adversarial neural network (DANN), central moment discrepancy (CDM), and task refinement learning (TRL) using bidirectional encoder representation from transformers (BERT), robustly optimized BERT pre-training approach (RoBERTa), Electra, and text-to-text transfer transformer (T5) models for dimensional sentiment analysis. EmoBank served as the general domain dataset, while Semeval-2017 task5 provided the financial corpus. A binary financial lexicon was integrated to enhance the models' understanding of financial sentiment.

Evaluation metrics included mean absolute error (MAE) and Pearson's correlation coefficient. Results indicated that the lexicon-based prompt method consistently improved sentiment prediction accuracy across all models compared to traditional domain adaptation methods. Notably, T5 showed the best overall performance. Despite the method's success in correcting sentiment mispredictions related to financial vocabulary, challenges remained in fully capturing nuanced sentiment variations.

Conclusion

In conclusion, the lexicon-based prompt method presented a robust approach to enhancing dimensional sentiment analysis in the financial domain. By integrating domain-specific lexical information, the method effectively corrected sentiment predictions without relying on extensive labeled data. Experimental results demonstrated superior performance compared to traditional domain adaptation methods across various pre-trained models.

While successful in improving sentiment prediction accuracy, challenges remain in capturing nuanced financial sentiments comprehensively. Future research will focus on refining domain-specific lexicons and exploring unsupervised learning techniques to further improve model capabilities in understanding and predicting sentiment nuances in financial texts.