Source-linked AI summary
Grid Tagging Scheme for Aspect-oriented Fine-grained Opinion Extraction
Zhen Wu, Chengcan Ying, Fei Zhao, Zhifang Fan, Xinyu Dai, Rui Xia
TL;DR
AFOE extracts aspect–opinion pairs or sentiment-bearing triplets, but existing pipelines can suffer error propagation and inconvenience. GTS uses a unified grid-tagging task and an inference strategy that exploits indications among opinion factors; its models outperform strong baselines across OPE and OTE and achieve state-of-the-art performance.
Problem
Joint methods can extract separate aspect and opinion sets without their pair relation, while pipeline approaches can suffer error propagation and inconvenience in real-world scenarios.
Source: §1 · pp. 1–2Method
GTS tags word-pair relations in a unified grid task, uses inference to exploit mutual indications among opinion factors, and extends to OTE by substituting sentiment-polarity detection for pair-relation detection.
Source: §1 · p. 2Results
Across OPE and OTE, GTS models outperform strong baselines and achieve state-of-the-art performance; on OTE, GTS-CNN and GTS-BiLSTM improve F1-score by about 3% over IMN+IOG on 14res and 15res.
Source: §4.4 · p. 7, §6 · p. 9Takeaways & Limitations
The authors report that GTS is easily extended to other pair- or triplet-extraction tasks from text.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Aspect-oriented Fine-grained Opinion Extraction (AFOE) aims at extracting aspect terms and opinion terms from review in the form of opinion pairs or additionally extracting sentiment polarity of aspect term to form opinion triplet. Because of containing several opinion factors, the complete AFOE task is usually divided into multiple subtasks and achieved in the pipeline. However, pipeline approaches easily suffer from error propagation and inconvenience in real-world scenarios. To this end, we propose a novel tagging scheme, Grid Tagging Scheme (GTS), to address the AFOE task in an end-to-end fashion only with one unified grid tagging task. Additionally, we design an effective inference strategy on GTS to exploit mutual indication between different opinion factors for more accurate extractions. To validate the feasibility and compatibility of GTS, we implement three different GTS models respectively based on CNN, BiLSTM, and BERT, and conduct experiments on the aspect-oriented opinion pair extraction and opinion triplet extraction datasets. Extensive experimental results indicate that GTS models outperform strong baselines significantly and achieve state-of-the-art performance.
1 Introduction
AFOE extracts aspect–opinion pairs or sentiment-bearing triplets, but prior joint methods can leave their pair relations unmodeled and pipeline approaches can suffer error propagation. GTS makes extraction a unified grid-tagging task and adds an inference strategy to use relationships among opinion factors.
1 Introduction
- Joint extraction methods can return separate aspect and opinion sets, omitting pair relations crucial for downstream sentiment analysis.Those relations can support aspect-level sentiment classification, fine-grained opinion summarization, and in-depth opinion analysis.Source: §1 · p. 1
- Pipeline approaches to opinion-pair extraction can suffer error propagation and inconvenience in real-world scenarios.The introduction describes co-extraction followed by pair-relation detection, and aspect extraction followed by opinion-term extraction.Source: §1 · p. 2
- GTS transforms opinion-pair extraction into one grid-tagging task, tagging word-pair relations to decode all opinion pairs simultaneously.This extracts all OPE opinion factors in one step rather than through pipelines.Source: §1 · p. 2
- GTS extends to opinion-triplet extraction by replacing pair-relation detection with sentiment-polarity detection.The cited prior approach still used a two-stage framework and extracted aspect–opinion pairs in a pipeline.Source: §1 · p. 2
- The proposed scheme addresses complete AFOE, including OPE and OTE, as a unified tagging task rather than pipelines.The authors describe GTS as their first such approach and say the scheme extends to other text pair- or triplet-extraction tasks.Source: §1 · p. 2
- GTS uses an inference strategy to exploit mutual indications among opinion factors for more accurate extractions.For example, identifying “average” as an opinion term can suggest “coffee” as its aspect because it is the modifier.Source: §1 · p. 2
2 Grid Tagging Scheme
The section defines OPE and OTE, represents word-pair relations in a unified grid, and decodes the resulting tags into opinion pairs or triplets.
2.1 Task Definition
- OTE adds each opinion pair’s sentiment polarity—positive, neutral, or negative—to the OPE extraction of aspect–opinion pairs.OPE pairs non-overlapping aspect and opinion spans; OTE additionally assigns a polarity to each pair.Source: §2.1 · pp. 2–3
2.2 Grid Tagging
- GTS labels unordered word pairs with A, O, P, or N in an upper-triangular grid, converting AFOE into a unified tagging task.A and O mark words within the same aspect or opinion term; P marks a word pair linking an aspect and an opinion, and N indicates no relation.Source: §2.2 · p. 3
2.3 Decoding Algorithm
- For OPE, decoding identifies aspect and opinion spans from diagonal tags and links them when any cross-span word pair has tag P.The decoder relaxes strict matching because abundant N tags may lower recall.Source: §2.3 · pp. 3–4
- For OTE, the decoder assigns each extracted pair the most frequent predicted sentiment tag across its cross-span word pairs, rejecting pairs without a sentiment tag.Source: §2.3 · p. 4
3 Validation Models
GTS validation models use three encoder options, an inference strategy that exploits mutual indications among opinion factors, and a unified training loss over word-pair tags.
3 Validation Models
- The validation models use CNN, LSTM, and BERT as GTS encoders and add an inference strategy to exploit indications among opinion factors.The architecture is presented as a shared GTS framework with alternative neural encoders.Source: §3 · p. 4, §3.2 · p. 5
3.1 Encoding
- CNN, BiLSTM, or BERT encodes sentence words; GTS forms word-pair representations, adding pairwise attention except for BERT, whose encoder already has self-attention.CNN uses stacked convolutional layers, BiLSTM concatenates forward and backward hidden states, and BERT produces contextual representations with a bidirectional Transformer.Source: §3.1 · pp. 4–5
3.2 Inference on GTS
- Iterative inference uses predictions involving each word and other words, plus the pair's prior prediction, to refine word-pair tags before final-turn extraction.The final-turn predictions are used to extract fine-grained opinions.Source: §3.2 · p. 5
3.3 Training Loss
- Cross-entropy trains final-turn word-pair tag distributions against ground-truth tags, with different label sets for OPE and OTE.OPE uses {A, O, P, N}; OTE uses {A, O, Pos, Neu, Neg, N}.Source: §3.3 · p. 5
4 Experiments
The experiments evaluate GTS on opinion-pair and opinion-triplet extraction, alongside aspect- and opinion-term extraction and ablations. The reported results favor GTS over strong pipeline baselines, while the ablations examine attention and inference settings.
4.3 Results of Opinion Pair Extraction
- Across OPE datasets, all three GTS models improve over strong pipelines, and GTS-BERT achieves the best performance.The authors report especially clear improvements on 15res and 16res; they also note that GTS-CNN and GTS-BiLSTM improve without the additional training data used by RINANTE.Source: Table 3 · p. 6, §4.3 · p. 7
4.1 Datasets and Metrics
- The datasets align Fan et al. (2019) opinion-pair data with SemEval aspect-term and sentiment annotations to add sentiment polarity, and sentences can contain multiple aspects or opinions.Table 2 reports counts of sentences, aspect terms, opinion terms, pairs, and triplets across restaurant and laptop datasets.Source: §4.1 · pp. 5–6, Table 2 · p. 5
4.2 Experimental Settings
- Evaluation uses precision, recall, and F1-score, counting aspect and opinion spans as correct only when their boundaries exactly match the gold spans.CNN and BiLSTM use double embeddings, while BERT uses uncased BERTBASE; each model is run five times and results are averaged.Source: §4.1 · p. 6, §4.2 · p. 6
4.4 Results of Opinion Triplet Extraction
- On OTE, GTS-CNN and GTS-BiLSTM improve F1-score by about 3% over IMN+IOG without its auxiliary document-level sentiment data; GTS-BERT leads them by about 2%-3% on 15res and 16res.The authors say the latter comparison indicates a contribution from the tagging scheme beyond the BERT encoder.Source: Table 4 · p. 7, §4.4 · p. 7
4.5 Results of Aspects Term Extraction and Opinion Term Extraction
- On 15res, GTS improves aspect-term and opinion-term extraction as well as addressing OPE and OTE end to end.The authors attribute this to the tagging scheme and inference strategy exploiting potential connections among opinion factors.Source: §4.5 · p. 7, Table 5 · p. 7
4.6 Ablation Study
- Removing attention slightly lowers CNN and BiLSTM performance, while full models with inference outperform versions without inference across all datasets.CNN and BiLSTM perform best with two and three inference times, respectively; BERT peaks at one, and CNN outperforms BiLSTM across the tested inference times.Source: Table 6 · p. 8, §4.6 · p. 8, Figure 5 · p. 8
5 Related Work
Prior opinion-pair extraction work includes early rule-based approaches, neural pipelines, and joint methods using hand-crafted features. Opinion-triplet extraction also used a two-stage pipeline.
5 Related Work
- Earlier OPE methods mined frequent patterns or used dependency templates, while later neural approaches combined aspect and opinion extraction subtasks in pipelines.Joint inference with factor graphs or integer linear programming sought to avoid pipeline error propagation, but depended heavily on hand-crafted features and sometimes performed worse than pipelines.Source: §5 · p. 8
- Peng et al. (2019) extract aspect-sentiment pairs and unpaired opinion terms before detecting aspect-opinion relations in a second stage, leaving the key opinion-pair extraction pipelined.The authors identify this approach as also susceptible to error propagation.Source: §5 · p. 8
6 Conclusions
The conclusion presents GTS as an end-to-end unified tagging approach for aspect-oriented opinion extraction, with a designed inference strategy. Experiments with CNN, BiLSTM, and BERT models report strong-baseline and state-of-the-art performance on OPE and OTE.
6 Conclusions
- GTS converts AFOE into one grid-tagging task by labeling relations between all word pairs, then decodes the tags into opinion pairs or triplets.The scheme covers both OPE and OTE, which pipeline approaches typically divide into subtasks.Source: §6 · pp. 8–9
- The inference strategy is designed to exploit mutual indications among different opinion factors.Across CNN, BiLSTM, and BERT implementations, the paper reports strong-baseline and state-of-the-art performance on both OPE and OTE.Source: §6 · p. 9