Methods for Improving Text Classification in Social Media Analysis

Introduction
Classifying social media posts into different categories is crucial for mental health professionals to monitor and support individuals’ emotional well-being. This article explores novel methods for improving text classification, specifically token attribution and fine-tuning pre-trained language models like BERT. Token attribution helps understand the relevance between a word and a specific document without considering its position in a sentence. Fine-tuning these models can improve their performance by incorporating contextual information.
Token Attribution

Token attribution is a method that assesses the relevance of each word to a specific document without accounting for its position in a sentence. This approach helps analyze text more accurately, as it considers the context in which words appear. By calculating the TF-IDF score for each token (i.e., word or punctuation), we can determine the relevance between tokens and documents.
For instance, if a document discusses both "depression" and "anxiety," the term "mental health" will have a higher TF-IDF score than if it only appeared in relation to one of those words. By weighting each token’s importance using its TF-IDF score, we can better understand how words contribute to a document’s overall meaning.
Fine-Tuning BERT for Improved Performance

Pre-trained language models like BERT have shown impressive performance in text classification tasks. However, they can benefit from fine-tuning to improve their performance even further. Fine-tuning involves adjusting the model’s weights to incorporate contextual information specific to each task at hand. By adapting the model to the particular classification task, it can better understand the nuances of language and capture subtle patterns in text data.
The authors fine-tuned BERT using a combination of token attribution and TF-IDF scaling. Token attribution helped identify the most relevant words for each document, while TF-IDF scaling adjusted their weights based on their importance in relation to other tokens. This approach enabled the model to better capture long-range dependencies and improve its overall performance.
Results and Discussion

The authors evaluated their methods using a social media dataset with four classes of risks (no risk, low risk, medium risk, and high risk). They compared the results to those obtained without fine-tuning BERT or using only token attribution. The fine-tuned model outperformed the others in both recall and F1-score, demonstrating its effectiveness in capturing subtle patterns in text data.
However, they found that token attribution alone resulted in lower performance compared to fine-tuning BERT with TF-IDF scaling. This suggests that incorporating contextual information into the model’s weights can significantly improve its ability to classify text accurately.
Conclusion and Future Work

In conclusion, this article explored novel methods for improving text classification in social media analysis. Token attribution helps understand the relevance between words and documents, while fine-tuning pre-trained language models like BERT can capture contextual information and improve performance. By combining these approaches, we can better monitor and support individuals’ emotional well-being by accurately classifying social media posts into different categories.
For future work, the authors suggest further investigating other techniques to improve text classification, such as incorporating domain knowledge or using multimodal features like images or audio. Additionally, they propose exploring new applications of these methods in areas like mental health diagnosis, treatment planning, and patient monitoring.