Actionable Review Intelligence Using Aspect-Based Sentiment Analysis
Disciplines
Computer Engineering
Abstract (300 words maximum)
Customer feedback provides valuable insights that extend beyond basic positive or negative sentiment. For example, “the pasta was cold” and “I don’t like Italian food” are both negative statements, but only the first offers an actionable suggestion for improvement. This project develops an advanced natural language processing framework that performs aspect-based sentiment analysis along with a new dimension called actionability classification. The goal is to identify which aspects of a product or service are being discussed, determine the sentiment associated with each aspect, and evaluate whether the feedback contains a constructive and actionable point. The model is designed using a multi-task learning architecture based on BERT and RoBERTa, with three components: a Conditional Random Field layer for aspect extraction, an attention-based classifier for sentiment analysis, and a regression head for predicting actionability scores. These components are trained jointly using a weighted loss function to balance accuracy and interpretability. The model is evaluated on the Amazon Fine Food Reviews dataset, which includes over 500,000 entries, and a manually annotated subset for actionability assessment. Evaluation metrics include F1-score for aspect extraction, macro F1 for sentiment classification, and Pearson correlation and mean absolute error for actionability prediction. The proposed system helps organizations focus on specific and constructive feedback that can lead to practical improvements. This research contributes to the field of sentiment analysis by integrating actionability into the interpretation of customer reviews, creating a more comprehensive and meaningful understanding of user opinions.
Use of AI Disclaimer
no
Academic department under which the project should be listed
CCSE – Computer Science
Primary Investigator (PI) Name
Md Abdullah Al Hafiz Khan
Actionable Review Intelligence Using Aspect-Based Sentiment Analysis
Customer feedback provides valuable insights that extend beyond basic positive or negative sentiment. For example, “the pasta was cold” and “I don’t like Italian food” are both negative statements, but only the first offers an actionable suggestion for improvement. This project develops an advanced natural language processing framework that performs aspect-based sentiment analysis along with a new dimension called actionability classification. The goal is to identify which aspects of a product or service are being discussed, determine the sentiment associated with each aspect, and evaluate whether the feedback contains a constructive and actionable point. The model is designed using a multi-task learning architecture based on BERT and RoBERTa, with three components: a Conditional Random Field layer for aspect extraction, an attention-based classifier for sentiment analysis, and a regression head for predicting actionability scores. These components are trained jointly using a weighted loss function to balance accuracy and interpretability. The model is evaluated on the Amazon Fine Food Reviews dataset, which includes over 500,000 entries, and a manually annotated subset for actionability assessment. Evaluation metrics include F1-score for aspect extraction, macro F1 for sentiment classification, and Pearson correlation and mean absolute error for actionability prediction. The proposed system helps organizations focus on specific and constructive feedback that can lead to practical improvements. This research contributes to the field of sentiment analysis by integrating actionability into the interpretation of customer reviews, creating a more comprehensive and meaningful understanding of user opinions.