Abstract
Aspect-Based Sentiment Analysis (ABSA) requires a fine-grained understanding of the relationship between specific aspects and their context. This paper proposes the Hierarchical Gated Deep Memory Network (HG-DMN), which integrates position-aware embeddings and a dual-attention mechanism to capture word-level interactions and spatial relevance. By explicitly weighing the distance between words and targets and employing a multi-hop gated memory update, the model achieves state-of-the-art performance on standard sentiment benchmarks, specifically the SemEval-2015 Task 12 datasets.
Introduction
Sentiment analysis has evolved from judging an entire review to identifying feelings about specific "aspects" (e.g., in the sentence "The food was great but the service was slow," the sentiment is positive for food but negative for service). Existing ABSA models often suffer from two major flaws: coarse-grained interaction, where context is compressed into a single vector too early, and the neglect of positional information. In natural language, words closer to an aspect are often more relevant to it. For example, in "The price is reasonable although the service is poor," the word "reasonable" is physically closer to "price" and clearly describes it.
The HG-DMN addresses these issues by treating spatial proximity as a fundamental feature and using iterative reasoning to filter out noise, ensuring that the model focuses on high-signal sentiment words relevant to the target aspect.
Methodology
The HG-DMN architecture consists of three sophisticated modules designed to preserve nuance and context:
- Position-Aware Embedding Layer: Unlike standard embeddings, this layer calculates the relative distance of every word from the target aspect. This "positional index" is converted into a vector and concatenated with the word's semantic embedding (GloVe), ensuring the model knows exactly how far away a descriptor is from its target.
- Information Interaction Layer (Dual Attention): The model employs a two-way attention mechanism. Sentence-to-aspect attention identifies which words within a multi-word aspect (like "battery life") are most important. Aspect-to-sentence attention then uses that refined aspect representation to hunt for the most indicative sentiment words in the surrounding context.
- Gated Memory Update Module: Borrowing from Gated Recurrent Units (GRUs), the model uses a "multi-hop" memory approach. It "reads" the sentence multiple times, using gates to filter out irrelevant information and iteratively refine the sentiment representation before passing it to the final classifier.
Experimental Setup
The researchers evaluated their model using the SemEval-2015 Task 12 datasets, specifically the Laptop Dataset (reviews concerning laptop features) and the Restaurant Dataset (reviews covering food, service, and ambiance). The data was preprocessed to focus on Positive, Neutral, and Negative polarities, removing "conflict" labels to isolate the core classification task.
Results and Discussion
The HG-DMN outperformed several strong baselines, including popular models like IAN, RAM, and AOA. On the Laptop dataset, it achieved 74.82% accuracy and a 72.53% Macro-F1 score. On the Restaurant dataset, it reached 81.95% accuracy and a 74.07% Macro-F1 score.
Ablation studies confirmed that removing position-aware embeddings dropped accuracy by roughly 0.4–0.6%, while removing fine-grained attention mechanisms led to even more significant performance hits. As shown in the performance analysis across computational layers, the model typically reaches peak performance at approximately three hops, after which additional reasoning layers may introduce noise or overfit.
Limitations and Future Work
Despite its success, the model faces challenges with neutral sentiment classification due to data imbalance. It also struggles with non-compositional expressions (e.g., "sushi is to die for") and idioms (e.g., "the service was on point"), where word-level processing fails to capture the holistic meaning. Future work will explore integrating pre-trained language models like BERT and incorporating syntactic structures, such as dependency trees, to better handle complex sentence logic.
Summary updated