Source-linked AI summary
Parameterized Convolutional Neural Networks for Aspect Level Sentiment Classification
Binxuan Huang, Kathleen M. Carley
TL;DR
Aspect-level sentiment classification requires identifying polarity toward a specific aspect rather than treating the sentence as a whole. The paper introduces CNNs with aspect-conditioned parameterized filters and gates, and reports strong performance, including 2%-5% improvements over AF-LSTM. The authors also identify dataset imbalance as a scope limitation that can make vanilla CNNs appear competitive.
Problem
Aspect-level sentiment classification must determine sentiment for a target aspect when a sentence contains multiple aspects with potentially different polarities.
Method
The paper introduces CNN-based models whose parameterized filters and gates incorporate aspect-specific information into sentence processing.
Results
The models achieve the best performance against the reported baselines, with 2%-5% improvements over AF-LSTM.
Takeaways & Limitations
Aspect-conditioned convolutional neural units effectively learn aspect-specific sentiment expressions and support CNNs as an approach for this task.
Takeaways & Limitations
The datasets are imbalanced: in the restaurant test set, 1034 of 1117 points share sentence-level and aspect-level labels, motivating more balanced future datasets.
Abstract
from arXiv · showhide
We introduce a novel parameterized convolutional neural network for aspect level sentiment classification. Using parameterized filters and parameterized gates, we incorporate aspect information into convolutional neural networks (CNN). Experiments demonstrate that our parameterized filters and parameterized gates effectively capture the aspect-specific features, and our CNN-based models achieve excellent results on SemEval 2014 datasets.
1 Introduction
Aspect-level sentiment classification must identify sentiment toward a specific aspect rather than an entire sentence. The paper addresses this challenge with aspect-aware convolutional neural networks using parameterized filters and gates.
- Multiple aspects in one sentence can express different polarities, making whole-sentence sentiment insufficient.For example, food can be positive while service is negative in the same sentence.
- Aspect-level classification identifies opinions about specific entities and aspects, unlike general sentiment classification.Ignoring aspect information makes target-specific polarity difficult when a sentence contains multiple aspects.
- Neural approaches learn representations without manual feature engineering and have become increasingly popular for this task.Prior neural methods primarily use long short-term memory or memory-network architectures.
- The paper proposes two convolutional neural networks that incorporate aspect information through parameterized filters and parameterized gates.Both neural units are generated from aspect-specific features and applied to the sentence.
2 Related Work
Prior aspect-level sentiment systems used rules, target-dependent machine-learning features, LSTMs, and memory networks. This paper differs by introducing CNNs for the task, motivated by their ability to capture local patterns.
- Aspect-level sentiment classification identifies the polarity of one aspect target within a context sentence.
- Early systems used dependency parsing and predefined language rules, while SVMs incorporated target-dependent and content features.
- Recent neural approaches predominantly use LSTMs with attention or aspect-sentence fusion mechanisms.Examples include attention vectors from aspect embeddings, word-aspect fusion, and attention-over-attention layers.
- Memory-network approaches treat the aspect as a query or add associative layers to attention-based computational architectures.
- The paper presents the first reported use of convolutional neural networks for aspect-level sentiment classification.Its architecture differs substantially from the preceding rule-based, recurrent, and memory-network approaches.
3 Parameterized Convolutional Neural Networks
The paper adapts CNNs for aspect-level sentiment classification by incorporating aspect information through parameterized filters and gates. PF-CNN uses aspect-derived filters, while PG-CNN uses aspect-derived gates to control sentence processing before classification.
- Aspect-level sentiment classification requires identifying sentiment toward a specific target aspect, such as distinguishing opinions about food from opinions about service.
- Standard CNNs do not use aspect-term information, motivating aspect-conditioned convolutional architectures.
- Parameterized Filters: PF-CNN extracts an aspect feature with CNNt, transforms it into parameterized filters, and applies those filters to the sentence.
- Parameterized Filters: PF-CNN combines aspect-specific sentence features with general sentiment features to form the final classification representation.
- Parameterized Gates: PG-CNN uses an aspect-derived feature as a sigmoid gate that controls how much sentence information passes through each convolutional layer.
- Classification and Training: The models generate classification features, project them through a linear layer, compute class probabilities with softmax, and train using cross-entropy loss with L2 regularization.
4 Experiments
Experiments evaluate the proposed CNNs on SemEval 2014 laptop and restaurant datasets against accuracy-based baselines. The parameterized models distinguish aspect-specific sentiment expressions and outperform AF-LSTM, while vanilla CNN remains competitive with some LSTM models.
- Experimental setting: The experiments use SemEval 2014 laptop and restaurant datasets, where each example pairs a sentence with an aspect term and sentiment polarity.Five hundred training instances are used as the development set.
- Baselines: Accuracy is used to compare the proposed models with TD-LSTM, AT-LSTM, ATAE-LSTM, AF-LSTM, and vanilla CNN baselines.Baseline descriptions specify context modeling, attention, aspect fusion, and aspect-agnostic convolutional architectures.
- Results: The two proposed models achieve the best performance among the baselines and improve over AF-LSTM by 2%-5%.The comparison results are reported in Table 2.
- Case study: Parameterized filters and gates distinguish describing words for the relevant aspect, assigning neutral and negative labels to “appetizers” and “service” respectively in a mixed-aspect sentence.Vanilla CNN assigns the same negative label to both aspects in that example.
- Discussion: Vanilla CNN has comparable performance with some well-designed LSTM models despite ignoring aspect information.The discussion attributes this partly to overlap between sentence-level and aspect-level labels.
- Discussion: The restaurant data contain 1034 of 1117 test points with matching sentence-level and aspect-level labels under the majority-label assumption.The authors suggest more balanced datasets for future research.
5 Conclusion
The paper proposes parameterized filters and gates that incorporate aspect information into CNNs for aspect-level sentiment classification. Comparisons show effective learning of aspect-specific expressions, strong improvements over neural baselines, and a first CNN-based attempt for this task.
- Contribution: The paper introduces parameterized filters and parameterized gates to incorporate aspect information into CNNs for aspect-level sentiment classification.Both neural units are generated from aspect-specific features and applied to the sentence.
- Findings: The experiments show significant improvement over multiple strong neural baselines and effective learning of aspect-specific sentiment expressions.The conclusion summarizes the comparison-based evaluation of the proposed method.
- Novelty: The authors identify the model as the first CNN attempt for aspect-level sentiment classification to their knowledge.They suggest exploring CNNs and parameterized CNN architectures in future work.