Introduction
In the evolving landscape of e-commerce, understanding consumer feedback is paramount for product development and marketing strategy. Traditional fine-grained sentiment analysis has primarily focused on extracting "triplets": the aspect (e.g., "sound"), the opinion (e.g., "loud"), and the sentiment polarity (e.g., "negative"). While these triplets provide a snapshot of user experience, they often fail to capture the actionable intent behind the feedback. For instance, if a customer complains that a washing machine is excessively noisy, the implicit demand is for the manufacturer to "improve product quality."
This research introduces the Mutually Enhanced Bidirectional Extraction (MEMB) framework, a novel approach designed to bridge this gap. By jointly identifying aspects, opinions, sentiments, and implicit user demands, the model provides a more comprehensive "quadruple" representation of consumer reviews. This bidirectional Machine Reading Comprehension (MRC) approach captures the dynamic interplay between product descriptions and underlying user needs.
Problem Statement
Mining implicit user demand from reviews presents two significant challenges:
- Vague Boundaries: Unlike simple adjectives, user opinions and demands are often expressed through complex, multi-word phrases. This makes it difficult for standard sequence labeling models to accurately identify where an aspect ends and an opinion begins.
- Hidden Dependencies: User demand is frequently triggered by dissatisfaction expressed through sentiment. Capturing this nuanced interaction requires a model capable of understanding the deep semantic relationships between disparate pieces of information within a sentence.
Methodology
The MEMB framework treats the extraction task as an MRC problem, where the model "asks" specific questions of the text rather than simply tagging tokens. The process is structured into two primary stages:
Stage 1: Initial Extraction
The model performs an initial pass over the text to identify potential aspects and opinions. This stage provides a foundation for the more complex reasoning that follows.
Stage 2: Bidirectional Enhancement
This stage represents the core innovation of the framework. The model utilizes the results from the first stage to construct more specific, context-aware queries. For example, if an aspect like "sound" is identified, the model generates a query such as: "What opinions are expressed regarding [sound]?"
Simultaneously, the model highlights previously extracted opinions using special tokens (e.g., [SEP]loud[SEP]) to guide its attention mechanism. This bidirectional flow—using aspects to find opinions and opinions to find aspects—allows the model to resolve ambiguities and refine entity boundaries that single-pass models might overlook.
Joint Training and Architecture
The system utilizes a pre-trained BERT backbone and is trained with four simultaneous classifiers. These classifiers are responsible for identifying the start and end positions of aspects and opinions, determining sentiment polarity, and categorizing the specific type of implicit demand (e.g., "product quality improvement").
Dataset and Evaluation
To evaluate the framework, the researchers developed a new dataset consisting of Chinese e-commerce reviews for electrical appliances (e.g., refrigerators, ovens, washing machines). The dataset includes 1,944 labeled reviews containing 4,849 "quadruples" (aspect-opinion-sentiment-demand). Annotation quality was verified with a Fleiss’ kappa coefficient of 0.77, indicating substantial agreement among human labelers.
Results
The MEMB model (referred to as MEBE in experimental results) was compared against several baseline architectures, including TSF (a pipeline model), RACL+ (a collaborative learning model), and BMRC (a standard bidirectional MRC model).
As shown in the results table, MEMB achieved superior performance across all primary metrics. Notably, it reached an F1 score of 67.71 for Aspect-Opinion Extraction and 64.32 for Demand Detection. The performance gains in demand detection are particularly significant, suggesting that the mutually enhanced mechanism is highly effective at capturing the relationship between specific complaints and underlying consumer needs.
Conclusion and Future Work
The transition from sentiment triplets to quadruples represents a significant step toward making sentiment analysis more actionable for businesses. By leveraging MRC as a form of prior knowledge, the MEMB framework effectively transfers information between sub-tasks, resolving the ambiguity inherent in natural language reviews.
Future research may address current limitations, such as the model's domain specificity to electronics and the use of predefined demand categories. Expanding the framework to support open-ended demand generation or more granular categories could further enhance its utility in diverse market sectors.
Summary updated