Abstract
We present ABSApp, a portable system for weakly-supervised aspect-based sentiment extraction. The system is interpretable and user friendly and does not require labeled training data, hence can be rapidly and cost-effectively used across different domains in applied setups. The system flow includes three stages: First, it generates domain-specific aspect and opinion lexicons based on an unlabeled dataset; second, it enables the user to view and edit those lexicons (weak supervision); and finally, it enables the user to select an unlabeled target dataset from the same domain, classify it, and generate an aspect-based sentiment report. ABSApp has been successfully used in a number of real-life use cases, among them movie review analysis and convention impact analysis.
Introduction
Standard sentiment analysis often provides a single aggregate score for an entire text, which can obscure nuanced feedback where different aspects of a product or service receive varying sentiments (e.g., "The food was amazing, but the service was glacial"). Aspect-Based Sentiment Analysis (ABSA) addresses this by identifying specific targets (aspects) and the sentiment directed at them.
However, high-performing ABSA systems typically rely on supervised learning, requiring thousands of hand-labeled examples. These models are often domain-specific; a model trained on restaurant reviews may fail on laptop reviews due to differing vocabularies. Labeling new data for every domain is prohibitively expensive and slow for many commercial applications. ABSApp addresses these challenges by providing a portable, human-in-the-loop system that performs ABSA without massive labeled training sets.
System Architecture
The ABSApp workflow prioritizes portability and interpretability through a six-step process categorized into three main stages.
Stage 1: Unsupervised Lexicon Extraction
The system utilizes a bootstrap approach to generate domain-specific lexicons. Starting with a small "seed" of generic opinion words, it employs dependency parsing to identify grammatical relationships (e.g., "adjective modifies noun") and discover new aspect terms and opinion words within an unlabeled dataset. To ensure quality, a Multi-Layer Perceptron (MLP) classifier, pre-trained on generic data, filters out noisy candidate terms. Sentiment polarity is estimated using pre-trained GloVe embeddings by calculating semantic distances to known positive and negative anchors.
Stage 2: Weak Supervision
The core innovation of ABSApp is the manual "weak supervision" phase. The system presents the generated lexicons to the user via a graphical interface. In approximately 15 minutes, a user can refine the lexicons by deleting irrelevant terms or grouping synonyms (e.g., "drinks" and "beverages"). This brief human intervention significantly enhances the system's accuracy compared to purely unsupervised methods.
Stage 3: Classification and Visualization
In the final stage, the system scans a target dataset and matches aspect-opinion pairs based on dependency rules. It accounts for linguistic nuances such as negations (e.g., "not good") to produce a final sentiment report. The results are then visualized to provide clear, actionable insights.
Evaluation and Results
The system was evaluated using the SemEval 2014 (Task 4) and SemEval 2015 (Task 12) datasets, focusing on the restaurant and laptop domains. These disjoint domains were chosen to test the system's robustness across unrelated vocabularies.
ABSApp demonstrated performance competitive with state-of-the-art supervised transfer learning models. In the restaurant domain, it achieved an F1 score of 51.1, nearly matching the supervised RNSCN model (51.5). In the laptop domain, it outperformed the Hier-Joint baseline with a score of 40.1 versus 34.2. Notably, the weak supervision phase provided a substantial boost over purely unsupervised versions. The system maintained high precision (over 70%) for sentiment polarity detection, which is often prioritized in industrial applications for reliable business insights.
Conclusion
ABSApp offers an efficient alternative to data-heavy supervised ABSA models by leveraging unsupervised extraction and minimal human intervention. While the current implementation faces limitations with context-dependent word polarities and complex grammatical structures, it provides a transparent and interpretable solution for cross-domain sentiment analysis. Future work may involve integrating contextual embeddings like BERT to further improve domain-specific polarity detection.
Summary updated