Aspect-Level Sentiment Classification Through Merged Dependency Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing aspect-level sentiment classification models are vulnerable to parsing errors, especially in out-of-domain settings, as they rely on syntactic structures that can be misleading due to imperfect dependency parses, leading to inaccurate sentiment predictions.
Innovation Solution
The GraphMerge technique combines dependency trees from multiple parsers to form a merged graph ensemble, using relational graph attention networks (RGAT) to learn suitable edges, reducing reliance on any specific parser and enhancing robustness to parsing errors while avoiding overfitting and computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If dependency trees from a single parser are used for aspect-level sentiment classification, then the model structure is simple and computational cost is low, but the model is vulnerable to parsing errors and produces inaccurate sentiment predictions
Solution Approach 1:
The patent merges dependency trees from multiple parsers into a unified graph structure, combining the strengths of different parsing approaches while mitigating individual parser errors. This merging process creates a more reliable syntactic representation for sentiment classification without requiring separate processing pipelines for each parser.
Solution Approach 2:
The patent employs multiple parsers (e.g., Stanford Parser, NLTK, spaCy) that can handle diverse sentence structures and linguistic patterns. Each parser serves as a universal tool that can process different types of syntactic constructions, and their combined output provides comprehensive coverage of various parsing hypotheses.
2Reliability
If multiple parsers are used to generate multiple dependency trees, then robustness to parsing errors is improved, but computational overhead increases
Solution Approach 1:
Instead of processing multiple dependency trees separately through independent model instances, the patent merges them into a single unified graph structure that can be processed by one model. This merging eliminates redundant computational operations while preserving the diversity of parsing hypotheses, thereby reducing overall computational cost.
Solution Approach 2:
The patent segments the processing into two distinct phases: (1) generating multiple dependency trees from different parsers, and (2) merging them into a unified graph for single-model processing. This segmentation allows the system to benefit from multiple parsers without requiring multiple model instances, thus controlling computational overhead.
3Adaptability or versatility
If multiple dependency trees are processed separately by separate model instances, then diverse parsing hypotheses are considered, but model overfitting occurs and computational efficiency decreases
Solution Approach 1:
The patent combines multiple dependency trees into a single unified graph that preserves diverse parsing hypotheses while enabling processing by a single model instance. This approach maintains the adaptability to handle multiple parsing interpretations without the computational inefficiency of multiple separate model instances.
Solution Approach 2:
The patent transitions from processing multiple independent dependency trees in separate model instances to merging them into a unified graph structure that incorporates multiple parsing hypotheses as alternative paths. This dimensional change allows the system to consider diverse hypotheses within a single processing framework, improving computational efficiency while maintaining versatility.
Data Source
AI summary
System and method for aspect-level sentiment classification. The system includes a computing device, the computing device has a processer and a storage device storing computer executable code. The computer executable code is configured to: receive an aspect term-sentence pair; embed the aspect term-sentence pair; parse the sentence using multiple parsers to obtain dependency trees, and perform edge union to obtain a merged graph; combine the embedding and the merged graph to obtain a relation graph; perform a relation graph neural network on the relation graph; extract hidden representation of the aspect term from updated relation neural network; and classify the aspect term based on the extracted representation to obtain a predicted classification label of the aspect term. During training, the computer executable code is further configured to calculate a loss function based on the predicted label and the ground truth label, and adjust parameters of models.