Source-linked AI summary
A Unified Model for Opinion Target Extraction and Target Sentiment Prediction
Xin Li, Lidong Bing, Piji Li, Wai Lam
TL;DR
Existing TBSA research often separated target extraction from sentiment classification, although practical use requires both outputs. This paper introduces an end-to-end unified tagging framework with stacked recurrent networks, boundary guidance, transition modeling, and sentiment-consistency gating. Across benchmark datasets, the framework consistently achieves the best F1 scores and a new state-of-the-art result.
Problem
Most existing TBSA studies address opinion target extraction or target sentiment classification alone, while complete applications require both target mentions and their sentiment predictions.
Method
The framework uses stacked recurrent networks with unified tags, auxiliary boundary prediction, boundary-to-sentiment transition modeling, and a gate for within-target sentiment consistency.
Results
The framework consistently gives the best F1 score across all datasets and achieves a new state-of-the-art result for complete TBSA.
Takeaways & Limitations
An integrated end-to-end model can outperform pipeline and unified sequence-tagging baselines on complete TBSA.
Abstract
from arXiv · showhide
Target-based sentiment analysis involves opinion target extraction and target sentiment classification. However, most of the existing works usually studied one of these two sub-tasks alone, which hinders their practical use. This paper aims to solve the complete task of target-based sentiment analysis in an end-to-end fashion, and presents a novel unified model which applies a unified tagging scheme. Our framework involves two stacked recurrent neural networks: The upper one predicts the unified tags to produce the final output results of the primary target-based sentiment analysis; The lower one performs an auxiliary target boundary prediction aiming at guiding the upper network to improve the performance of the primary task. To explore the inter-task dependency, we propose to explicitly model the constrained transitions from target boundaries to target sentiment polarities. We also propose to maintain the sentiment consistency within an opinion target via a gate mechanism which models the relation between the features for the current word and the previous word. We conduct extensive experiments on three benchmark datasets and our framework achieves consistently superior results.
Introduction
TBSA combines opinion target extraction with sentiment classification, but prior work often treated them separately or assumed target mentions were already given. The paper proposes an end-to-end unified framework that jointly uses target-boundary information and sentiment-aware tagging.
- Task motivation: TBSA detects explicitly mentioned opinion targets and predicts their sentiment polarities.The paper illustrates this with multiple targets in one sentence receiving different polarities.
- Task motivation: Traditional approaches separate opinion target extraction from target sentiment classification, with many classification methods assuming target mentions are given.This limits the completeness of the practical TBSA solution addressed by the paper.
- Research gap: Because the two sub-tasks are highly coupled, the paper argues that an integrated model is promising despite relatively limited prior work on complete TBSA.Earlier integrated attempts used either jointly trained boundary and sentiment tags or other combined formulations.
- Proposed framework: The proposed framework uses two stacked RNNs: an upper network predicts unified TBSA tags, while a lower network provides auxiliary target-boundary predictions.Boundary predictions guide the upper network through explicitly modeled dependencies between boundary tags and sentiment tags.
- Proposed framework: The model maintains sentiment consistency within target mentions through a gate and refines boundary information using supervision from words near opinion words.The gate consolidates current- and previous-word features, while target-word potential improves lower-level boundary prediction.
Our Proposed Framework
The framework jointly performs complete TBSA through unified tagging, stacked recurrent networks, boundary guidance, sentiment consistency, and opinion-enhanced target detection.
- Unified Task Formulation: The unified tagging scheme encodes target boundaries and sentiment polarities in one sequence-labeling output.Tags distinguish boundary positions such as B, I, E, and S from POS, NEG, or NEU sentiment.
- Stacked RNN Architecture: Two stacked LSTMs use the upper network for unified TBSA tagging and the lower network for auxiliary target-boundary prediction.The lower network supplies boundary information to guide the upper network’s unified-tag predictions.
- Boundary Guidance: Boundary Guidance encodes valid boundary-to-unified-tag transitions and weights boundary-derived scores according to boundary confidence.Invalid transitions receive zero probability, while uncertain boundary predictions contribute less to the final tagging decision.
- Sentiment Consistency: Sentiment Consistency gates previous-word features into the current prediction to reduce inconsistent sentiments within multi-word targets.The mechanism combines current and previous features through learnable parameters and element-wise gating.
- Opinion-Enhanced Detection: Opinion-Enhanced target-word detection adds an auxiliary binary classifier to refine boundary representations using target-word supervision.The classifier distinguishes target words from non-target words, supporting higher-quality boundary information.
- Training: All framework components are differentiable and trained jointly by aggregating the main TBSA loss with two auxiliary-task losses.The training objective uses token-level cross-entropy errors for the task-specific predictions.
Dataset
The experiments use two SemEval ABSA product-review datasets and a Twitter dataset, with auxiliary boundary annotations and exact-match evaluation of target spans and sentiments.
- Dataset: The benchmarks comprise laptop reviews (DL), merged restaurant reviews (DR), and tweets (DT).DL uses the original train-test split; DR merges SemEval 2014–2016 restaurant data; DT uses ten-fold cross-validation.
- Dataset: DL and DR reserve 10% of randomly held-out training data for development, while DT has no standard train-test split.DT therefore follows the ten-fold cross-validation protocol used in earlier work.
- Annotations: Gold boundary annotations support the auxiliary target-boundary prediction task.Opinion words for the auxiliary target-word detection task come from an existing opinion lexicon.
- Evaluation: Evaluation uses exact-match precision, recall, and F1, counting an output as correct only when both span and sentiment match the gold target.The metric requires simultaneous agreement on the target mention boundary and corresponding sentiment.
Compared Models
The study compares the proposed framework with CRF, neural-CRF, pipeline, unified-tagging, LSTM-CRF, and language-model-enhanced sequence taggers.
- CRF baselines: CRF-{pipeline, joint, unified} represents pipeline and joint or unified tagging approaches based on conditional random fields.These models are attributed to Mitchell et al. (2013).
- Neural-CRF baselines: NN-CRF-{pipeline, joint, unified} enhances CRF models with word embeddings and neural-network feature extractors.These models are attributed to Zhang, Zhang, and Vo (2015).
- Pipeline baseline: HAST-TNet is the pipeline combination of HAST for target-boundary detection and TNet for target-sentiment classification.The authors identify HAST and TNet as state-of-the-art models for their respective subtasks and use officially released code.
- Unified taggers: The unified-tagging sequence-tagger baselines include standard LSTM-unified, two LSTM-CRF variants, and language-model-enhanced LM-LSTM-CRF.The LSTM-CRF variants differ in their character-level representations: LSTM-CRF-1 uses LSTM, whereas LSTM-CRF-2 uses CNN.
Experiment Settings
Training uses pretrained word embeddings, randomized model initialization, Adam optimization, dropout, and development-set F1 selection, with fixed hidden dimensions and component hyperparameters.
- Initialization: LSTM weight matrices use Glorot Uniform initialization, other parameters use U(-0.2, 0.2), and all biases start at 0.These initialization settings apply to the model parameters described in the experiment setup.
- Optimization: Models train for up to 50 epochs with Adam, β1 = β2 = 0.9, and initial learning rate η0 = 10^-3.Dropout is applied to word embeddings and ultimate prediction features at rate 0.5; the best development-set F1 model produces test results.
- Hyperparameters: Both hidden-representation dimensions dimT and h are 50, with maximum boundary-score proportion ϵ = 0.5 and opinion-detection window size s = 3.The effects of ϵ and s are tuned later in the experiments.
Results and Analysis
The framework achieves the best complete-TBSA F1 across the benchmark datasets, while ablations and case analyses examine boundary guidance, sentiment consistency, opinion detection, and hyperparameter choices.
- Main Results: The proposed framework consistently achieves the best F1 across all datasets and significantly outperforms the strongest baselines in most cases.Comparisons use aligned train, development, and test configurations and common pretrained embeddings where applicable.
- Main Results: 2.6%, 2.4% and 0.40% absolute F1 gains over HAST-TNet occur on DL, DR and DT respectively.These results support the effectiveness of an integrated model over the compared pipeline approach.
- Baseline Analysis: CRF-based models show poor performance, particularly on recall, while embeddings and neural feature extractors provide only slight improvement.The paper reports that these scores remain unpromising after enhancement.
- Ablation Analysis: The stacked-LSTM base model outperforms LSTM-unified, indicating that auxiliary boundary predictions increase complete-TBSA F1.Adding BG improves performance further, and its boundary constraints yield more true positives.
- Ablation Analysis: The full model requires both SC and OE because their combination produces the new state-of-the-art result, whereas BG with only one component is weaker.The paper characterizes SC and OE as complementary within the boundary-guided base model.
- Case Analysis: Case analyses show that BG and the full model correct sentiment errors when the base model identifies target boundaries but predicts sentiments incorrectly.The examples also show that relying only on BG can inherit lower-level boundary-detection errors, motivating higher-quality boundary information from OE.
- Hyperparameter Analysis: ϵ = 0.5 gives the best development-set result on DR, while s = 3 is the best window size for the TBSA task.The paper attributes poorer performance at larger s to irrelevant opinion words and at smaller s to insufficient opinion-word coverage.
Related Works
Target-Based Sentiment Analysis comprises Opinion Target Extraction and Target Sentiment Classification, which are often treated and solved separately. A complete practical output should provide both extracted targets and their sentiment.
- Target-Based Sentiment Analysis is divided into Opinion Target Extraction and Target Sentiment Classification.
- Most existing approaches solve the two sub-tasks individually rather than within one framework.
- For practical applications, a method should output both opinion targets and their sentiment.
Conclusions
The framework formulates TBSA as unified sequence tagging with stacked LSTMs and auxiliary boundary information. Experiments report a new state-of-the-art result, while Figure 2 evaluates F1 across parameter settings.
- The framework uses two stacked LSTMs for auxiliary target boundary detection and the complete TBSA task.
- Its auxiliary components exploit target boundary information, refine boundary quality, and maintain sentiment consistency.
- Figure 2 reports F1 scores (%) on DR's development set across different ϵ and s values.
- Experimental results and case studies demonstrate the proposed framework's effectiveness and establish a new state-of-the-art result.