PaperlayerView exact revision on arXiv

Source-linked AI summary

Position-Aware Tagging for Aspect Sentiment Triplet Extraction

Lu Xu, Hao Li, Wei Lu, Lidong Bing

arXiv:2010.02609v3cs.CL

TL;DR

ASTE requires jointly extracting targets, sentiment, and explanatory opinion spans, yet existing pipeline and tagging approaches do not adequately represent their connections. The paper proposes JET, an end-to-end position-aware tagging model that encodes triplet structure and interactions. Across the reported experiments, JET outperforms strong baselines significantly, while prior tagging approaches cannot handle overlapping opinion spans in some cases.

  • Problem

    Existing ASTE approaches use pipelines, while BIOES tagging lacks positional information for representing connections among targets, sentiments, and opinion spans.

  • Method

    JET uses a position-aware tagging scheme with enriched labels and factorized feature representations to jointly extract ASTE triplets.

  • Results

    JET models significantly outperform strong baselines in the reported extensive experiments.

  • Takeaways & Limitations

    Jointly encoding triplet elements provides a model that captures their interactions and handles cases involving overlapping opinion spans.

  • Takeaways & Limitations

    The position-aware tagging scheme assumes non-overlapping targets and one corresponding opinion span per target.

Abstract

from arXiv · show

Aspect Sentiment Triplet Extraction (ASTE) is the task of extracting the triplets of target entities, their associated sentiment, and opinion spans explaining the reason for the sentiment. Existing research efforts mostly solve this problem using pipeline approaches, which break the triplet extraction process into several stages. Our observation is that the three elements within a triplet are highly related to each other, and this motivates us to build a joint model to extract such triplets using a sequence tagging approach. However, how to effectively design a tagging approach to extract the triplets that can capture the rich interactions among the elements is a challenging research question. In this work, we propose the first end-to-end model with a novel position-aware tagging scheme that is capable of jointly extracting the triplets. Our experimental results on several existing datasets show that jointly capturing elements in the triplet using our approach leads to improved performance over the existing approaches. We also conducted extensive experiments to investigate the model effectiveness and robustness.

1 Introduction

ASTE extracts targets, their sentiments, and corresponding opinion spans, but existing pipeline and BIOES-based approaches struggle to learn their connections jointly. The paper introduces JET, a position-aware tagging approach designed to encode triplet structure and capture interactions among its elements.

  • Task motivation: ASTE extracts targets, associated sentiment polarities, and opinion spans that explain each target’s sentiment.The three elements form an aspect sentiment triplet.
  • Task motivation: Existing approaches use a two-stage pipeline that separately labels targets, sentiments, and opinion spans before pairing targets with opinions.The prior sequence-labeling stage uses LSTM, CRF, or GCN models, followed by an MLP pairing classifier.
  • Research gap: Because triplet elements are highly related, jointly modeling their interactions may better capture sentiment, target context, and opinion wording.For example, opinion wording can depend on the target, while sentiment polarity depends on the opinion span, target, and context.
  • Research gap: BIOES tagging lacks positional information needed to specify connections between targets and their corresponding opinion spans.This limitation prevents the tagging scheme from representing the richer structural interactions within a triplet.
  • Proposed approach: JET introduces a position-aware tagging scheme that enriches label semantics to encode triplet connections and jointly predict ASTE triplets.The approach is presented as an end-to-end model for the task.
  • Reported outcome: Extensive experiments report that the joint JET approach significantly outperforms baselines.The reported results support improved modeling of interactions among triplet elements.

2 Our Approach

JET jointly extracts target, sentiment, and opinion-span triplets with a position-aware tagging scheme that encodes their structural relationships. Its CRF/Semi-Markov CRF architecture combines token- and segment-level features to model these interactions.

  • 2.1 Position-Aware Tagging Scheme: JET jointly extracts Target, Target Sentiment, and Opinion Span using a position-aware tagging scheme.The model is designed to capture interactions among all three triplet elements.
  • 2.1 Position-Aware Tagging Scheme: The scheme extends BIOES tags with sentiment polarity and offsets encoding an opinion span’s position relative to a target.Offsets are distances between the opinion span’s two ends and the target’s starting position.
  • 2.1 Position-Aware Tagging Scheme: A tag sequence corresponds one-to-one with aspect sentiment triplets when targets do not overlap and each target has one corresponding opinion span.The construction process is reversible under these conditions.
  • 2.1 Position-Aware Tagging Scheme: The position-aware scheme handles overlapping opinion spans that previous tagging approaches cannot represent.For “The salad is cheap with fresh salmon,” it represents both (“salad”, “cheap with fresh salmon”, positive) and (“salmon”, “fresh”, positive).
  • 2.2 Our JET Model: JET combines CRF and Semi-Markov CRF components to encode token-level target features and segment-level opinion-span features.Its factorized feature score includes local target context, opinion-span representations, and structural information.
  • 2.2.1 Neural Module: The neural module uses a bidirectional LSTM and boundary-based segment representations for opinion spans.The span representation is computed from forward and backward hidden states at the span boundaries.

3 Experiments

The experiments evaluate JET against pipeline baselines on four datasets using exact triplet matching, with results reported on the refined ASTE-Data-V2 dataset. JET generally outperforms the strongest baseline, while performance depends on the maximum offset and dataset.

  • Datasets: The refined ASTE-Data-V2 dataset adds missing triplets when one opinion span is associated with multiple targets.The authors give “Best” associated with both “service” and “atmosphere” as an example and report ASTE-Data-V1 separately in the Appendix.
  • Experimental setup: JET is compared with pipeline baselines on the 14Rest, 15Rest, 16Rest, and 14Lap datasets.The datasets cover restaurant and laptop domains and originate from SemEval datasets.
  • Evaluation: A correct triplet requires the target boundary, opinion-span boundary, and sentiment polarity to all be correct simultaneously.Precision, recall, and F1 are reported for these exact triplets; a boundary is the beginning and ending positions of a text span.
  • Main results: JETt and JETo outperform Peng et al. (2019) on all four datasets, with significant improvements on 14Rest and 16Rest at p < 10^-5.The models are selected using the best development-set F1 score, and F1 generally increases as M increases when M ≤5.
  • Main results: Nearly 7 F1 points separate JETo(M = 6) from Peng et al. (2019) on 14Rest.JETt(M = 6) reaches 56.58 F1, while models with M ≥4 improve precision by more than 15 points while maintaining acceptable recall.
  • Main results: JETt(M = 6)+ BERT and JETo(M = 6)+ BERT achieve new state-of-the-art performance on all four datasets.Performance is better on 14Rest and 16Rest than on 14Lap and 15Rest, which the authors relate to sentiment-distribution differences between training and test sets.

4 Analysis

The analysis examines robustness across span lengths, evaluation criteria, qualitative examples, ablations, and ensembles. Results show that model variants differ in boundary handling and complementarity, while longer spans remain harder to extract.

  • Robustness Analysis: Performance drops as target, opinion-span, and offset lengths increase, indicating that longer boundaries are harder to model.This pattern is reported on 14Rest and similarly observed on 14Lap, 15Rest, and 16Rest.
  • Robustness Analysis: JETo(M = 6)+ BERT better extracts targets of length ≤3 and offsets of length 4 or more than JETt(M = 6)+ BERT.JETo also performs better for opinion spans of lengths 1 and 4, while the models are comparable at lengths 2 and 3.
  • Robustness Analysis: Partial-boundary evaluation shows larger F1 improvements when target or opinion boundaries are relaxed, depending on the model variant.For JETt, (T, Op, S) generally improves more; for JETo, (Tp, O, S) generally improves more, with exceptions on 15Rest.
  • Qualitative Analysis: JETo correctly predicts two triplets in an example where JETt predicts only one because JETt cannot handle one target linked to multiple opinion spans.The qualitative examples also show both joint models avoiding incorrect target–opinion connections made by the earlier approach.
  • Ablation Study: Removing offset features causes larger F1 drops for JETt than for JETo, while character embeddings mostly provide little positive impact.The authors attribute the limited character-embedding benefit to data sparsity.
  • Ensemble Analysis: JETt→o achieves better F1 than JETt and JETo on 14Rest, 14Lap, and 15Rest, but the simple ensemble is less effective on 16Rest.The ensemble models significantly improve recall, reflecting more extracted gold triplets and supporting complementarity between the base models.

5 Related Work

Related work connects ASTE to aspect-based and targeted sentiment analysis, while distinguishing ASTE by requiring targets, sentiments, and corresponding opinion spans. Earlier approaches commonly use staged sequence labeling and pairing.

  • Related Tasks: ASTE extends aspect-based sentiment analysis by extracting target phrases, associated sentiment, and the opinion spans expressing that sentiment.These three elements form each aspect sentiment triplet.
  • ASTE Approaches: Earlier ASTE approaches first label targets, sentiments, and opinion spans, then use an MLP classifier to pair targets with corresponding opinion spans.The described sequence-labeling stage uses models including LSTM, CRF, and GCN.

6 Conclusion

The conclusion presents JET as a joint ASTE approach built on position-aware tagging and factorized feature representations. Experiments report significant improvements over strong baselines, while future work targets broader applications.

  • Conclusion: JET jointly extracts targets, target sentiment, and opinion spans using a position-aware tagging scheme with enriched label expressiveness.The scheme specifies connections among the three triplet elements.
  • Conclusion: Factorized feature representations are designed to capture interactions among elements in an aspect sentiment triplet.
  • Conclusion: Experiments with detailed analysis show that the proposed models significantly outperform strong baselines.
  • Future Work: Future work includes applying the tagging scheme to other triplet-extraction tasks and extending the approach to additional sentiment-analysis tasks.

A More Data Statistics

The appendix reports cumulative training-data statistics for target, opinion-span, and offset lengths across four datasets. These statistics support comparisons of length distributions used in the robustness analysis.

  • Data Statistics: Figure 7 reports cumulative percentages by target, opinion-span, and offset length in the training data across four datasets.The datasets named in the passage are 14Rest, 14Rest, 15Rest, and 16Rest.
  • Data Statistics: The appendix also reports counts of targets and opinions associated with single versus multiple spans.The supplied passage introduces these statistics but does not provide their values.

B Experimental Details

The model is evaluated with specified hardware, software, optimization, embedding, and architectural settings. Its reported runtime varies across datasets, while the model contains 2.5M parameters.

  • The experiments use an Intel(R) Xeon(R) Gold 6132 CPU and PyTorch version 1.40.
  • 2.5M parameters are reported, with average runtimes ranging from 1170 to 3300 sec/epoch across the listed datasets when M = 6.The passage lists runtimes of 3300, 1800, 1170, and 1600 sec/epoch for 14Rest, 14Rest, 15Rest, and 16Rest, respectively.
  • The model uses 300d pre-trained GloVe embeddings, 100-dimensional offset embeddings, and a bidirectional LSTM with hidden size 300.Out-of-vocabulary words and offset embeddings are randomly initialized from U(−0.1, 0.1).

C Experimental Results

The experiments use ASTE-Data-V1 and describe the position-aware tag representation and Viterbi decoding procedure. The decoding objective selects an optimal tag sequence using transition, tagging, sentiment, offset, and relation scores.

  • Experimental setup: Table 7 reports experimental results on the previously released ASTE-Data-V1 datasets.The table caption distinguishes best development-set scores from their corresponding test results.
  • Tagging scheme: The position-aware tag set encodes sentiment polarity together with distances between opinion-span ends and the target start.The sentiment symbol denotes target polarity, while j and k provide the position information.
  • Tagging scheme: The model defines B and S as sub-tags of Bε_j,k and retains I, O, and E as sub-tags.A bar denotes the sub-tag of a tag, such as the subtag of u ∈ T.
  • Decoding: Viterbi decoding computes π(i, v) as the best sequence score ending with tag v and recovers the optimal sequence y*.The recurrence combines transition and factorized feature scores, including tagging, sentiment, offset, and relation terms.
  • Decoding: The decoding algorithm initializes boundary cases, loops forward across tokens and tags, and then backtracks from the optimal final sequence.START precedes the input sentence and STOP follows it; the reported time complexity is O(n|T|) = O(nM^2).

E.1 Robustness Analysis

Robustness analysis compares JETt and JETo with BERT across target, opinion-span, and offset lengths on three datasets. Performance is reported using F1(%) as a function of element length.

  • Datasets and models: The analysis evaluates JETt and JETo with BERT on 14Lap, 15Rest, and 16Rest.The corresponding plots are Figures 8, 9, and 10, respectively.
  • Length-based analysis: F1(%) is plotted against different lengths for targets, opinion spans, and offsets.The analysis examines length-specific performance rather than only aggregate scores.

E.2 Qualitative Analysis

The qualitative analysis compares predictions from Peng et al. (2019), JETt, and JETo on an example containing two gold triplets. JETt recovers both triplets, whereas the comparison systems make different errors, illustrating model differences.

  • Qualitative Analysis: The example contains two gold triplets, but Peng et al. (2019) predicts only one opinion span.It incorrectly assigns “Good” to the target “price.”
  • Qualitative Analysis: JETt makes the correct predictions for the example, while JETo predicts only one triplet correctly.The qualitative analysis is presented to clarify differences among the compared models.
  • Qualitative Analysis: Table 8 presents the example sentence and predictions from Peng et al. (2019), JETt, and JETo.
  • Related work: The related task of joint entity and relation extraction differs by extracting entity pairs and their relations rather than targets, opinion spans, and sentiment.Prior approaches include tablefilling-based methods and tree-structured neural networks.
  • Robustness figures: Figures 8–10 encode F1(%) on the y-axis and element length on the x-axis for targets, opinion spans, and offsets.The figures cover 14Lap, 15Rest, and 16Rest, respectively.
Loading 2010.02609v3…