Abstract
Traditional sentiment analysis often identifies what a person feels but fails to capture why they feel it. This paper proposes a framework to detect underlying human motives—the fundamental drivers of opinions—across different consumer domains. By developing a taxonomy of core motives and annotating a new dataset of restaurant and laptop reviews, the authors demonstrate that motives are universal enough to support cross-domain transfer learning. The results show that while neural models benefit from transfer learning, a significant gap remains between machine performance and human reasoning.
Introduction
Standard sentiment analysis and Aspect-Based Sentiment Analysis (ABSA) provide a surface-level view of consumer feedback. While they can identify that a customer is dissatisfied with "service," they do not uncover the psychological driver behind that sentiment. For instance, a negative review of a slow waiter might stem from a frustrated need for efficiency (Ambition & Ability) or a feeling of being disrespected (Social Relation). Understanding these motives is essential for providing deeper business insights and creating AI that can truly justify human decisions.
Methodology: A Taxonomy of Motives
The authors condensed a complex psychological taxonomy of 161 motives into core categories relevant to consumer behavior: Self-Fulfillment (SF), Appreciating Beauty (BA), Social Relation (SR), Health (H), Ambition & Ability (AA), and Finance (F).
To detect these motives, the researchers treated the problem as a multi-label sentence classification task. They evaluated several architectures, including Support Vector Machines (SVM) with bag-of-n-grams and Multi-Layer Perceptrons (MLP) using various encoders: Simple Word-Embedding Averaging (SWEM), Convolutional Neural Networks (CNN), and Bidirectional Long Short-Term Memory (BiLSTM) networks.
Dataset and Domain Analysis
The study utilized 1,600 sentences sampled from the SemEval 2016 ABSA datasets, covering the Restaurant and Laptop domains. Annotation was performed via Amazon Mechanical Turk with rigorous quality controls.
Analysis of the dataset revealed distinct "hedonic" vs. "utilitarian" profiles. Restaurant reviews were primarily driven by Self-Fulfillment and Social Relation. In contrast, laptop reviews focused heavily on Ambition & Ability (efficiency) and Finance. Despite these distributional differences, the semantic expression of a motive (e.g., wanting to save money) remained consistent across domains, suggesting a degree of universality that can be exploited by machine learning models.
Results and Discussion
The experiments highlighted the effectiveness of transfer learning. By training models on both in-domain and out-of-domain data, performance improved across nearly all neural architectures. This confirms that knowledge about a motive in one context (e.g., "Finance" in laptops) is applicable to another (e.g., "Finance" in restaurants).
As shown in Table 2, the MLP-CNN with transfer learning achieved the highest F1 score in the Restaurant domain (0.583), while the MLP-LSTM led in the Laptop domain (0.518). However, a substantial "human-machine gap" persists; human annotators achieved F1 scores near 0.80, indicating that motive detection requires high-level reasoning that current models struggle to replicate fully.
Conclusion
This research shifts the focus of sentiment analysis from valence to motivation. By demonstrating the universality of human motives through transfer learning, the authors provide a path toward more explainable and deep consumer insights. Future work will focus on expanding the dataset size and exploring how the satisfaction or thwarting of specific motives directly translates into positive or negative sentiment valence.
Summary updated