Overview
Aspect-Level Sentiment Classification (ALSC) is a fine-grained sentiment analysis task that aims to determine the sentiment expressed toward specific entities or "aspects" within a sentence. For instance, in the sentence "the battery life is great but the screen is dim," a model must identify positive sentiment for the first aspect and negative for the second.
Despite the success of BERT-based models, two major challenges persist: data scarcity for rare aspects and the isolated context of entities. AR-BERT addresses these by incorporating semantic relationships from Knowledge Graphs (KGs) like DBpedia, using a novel embedding scheme and a "probing function" to handle noisy data.
Methodology
AR-BERT integrates external knowledge through a multi-stage architecture designed for scalability and robustness.
Two-Level Global-Local Embedding
To make Knowledge Graph integration computationally feasible, the authors avoid using the entire graph. Instead, they employ a two-level approach:
- Cluster Graph: Captures global connectivity using Louvain clustering to group related entities.
- Subgraph: Captures local relationships specifically between aspects present in the dataset. These are combined into a compact embedding that is trained jointly with the ALSC model, allowing the system to "borrow" sentiment signals from related entities (e.g., learning from "MacOS" to improve predictions for the rarer "OSX").
Incorrect Disambiguation Detection (IDD)
Automated "wikification" (linking text to KG nodes) is often error-prone. AR-BERT introduces a probing function that compares BERT’s internal representations with KG embeddings. If the representations diverge significantly—suggesting a mislink, such as mistaking "Vibe" the magazine for "Vibe" the phone—the model "zeroes out" the KG information to prevent it from corrupting the sentiment prediction.
Multi-modal Explanations
The model includes a post-hoc explanation component that identifies whether the text context or the KG relations were more significant for a specific prediction. This provides transparency into why the model relied on external knowledge versus the immediate sentence context.
Key Results and Insights
The model was evaluated on standard benchmarks including LAPTOP, REST (SemEval 2014), and TWITTER. AR-BERT demonstrated a consistent improvement of 2.5 to 4.1 percentage points in Macro-F1 score over strong baselines like BERT-ADA.
Solving the Long Tail
The most significant impact of AR-BERT is seen in "long-tail" aspects—entities that appear very few times in the training data.
As shown in the performance analysis, the gap between AR-BERT and the baseline is widest when training examples are scarce (0–20 examples). In these scenarios, the KG relations provide the necessary context that the training data lacks. As the number of training examples increases, the model naturally relies more on the text, but the KG remains a vital safety net for rare entities.
Robustness and Filtering
The IDD component proved highly effective, successfully filtering out over 80% of incorrectly linked entities across the datasets. This ensures that the inclusion of external knowledge remains a benefit rather than a source of noise.
Limitations and Future Work
While effective, the system requires significant one-time preprocessing to cluster the 22-million-node DBpedia graph. Additionally, it remains dependent on the initial "wikifier" to suggest entity candidates. Future research aims to move toward "inherently interpretable" architectures where explanations are generated as part of the primary classification process, rather than post-hoc.
Summary updated