Instruction Tuning for Few-Shot Aspect-Based Sentiment Analysis

• Full ABSA predicts aspect term, aspect category, opinion term, and sentiment polarity, whereas many prior approaches target only some elements.Source: §1 · p. 1
• The unified T5 multi-task prompting model addresses ASQP and four associated sub-tasks: AE, AESC, TASD, and ASTE.Source: §1 · pp. 1–2
• 8.29 F1 average improvement over the state of the art in few-shot learning is reported for the proposed model.Source: §1 · p. 2

2 Methods

• The method factorizes ABSA into five question-answering tasks, using instructional prompts to train a sequence-to-sequence model across tasks.The design reflects the hypothesis that interactions among sentiment elements matter both within subsets and at the quadruple level.Source: §2 · p. 2

2.1 Input Transformation

• Each task’s input is transformed with an instruction template, with paraphrased templates sampled during training and evaluation while the target stays unchanged.Source: §2.1 · p. 2

2.2 Model Training

• Training samples input-output combinations from all tasks and uses a token-probability loss on each task’s transformed input.The model parameters are updated with Adam and weight decay.Source: §2.2 · p. 2

2.3 Output Transformation

• Task-specific output templates encode predictions; multiple quadruples are separated with [SSEP], sentiment labels are mapped to great, bad, and ok, and reverse transformations recover quadruples at inference.Source: §2.3 · pp. 2–3

3 Experiments

• Because standard few-shot ABSA datasets are lacking, the experiments emulate few-shot data drawing on prior work.Source: §3 · p. 3

3.1 Datasets: Few-shot Preparation

• REST15 and REST16 samples are selected to include at least k examples per aspect category, while LAPTOP14 uses k examples per sentiment class because it lacks category annotations.The data are shuffled with a fixed random seed; exactly k samples cannot always be guaranteed when an example has multiple categories.Source: §3.1 · p. 3, §3.2 · p. 3

3.2 Baselines and Models for Comparison

• PARA is used for TASD, ASTE, and ASQP, while BARTABSA covers the other two tasks; PARA shares the proposed model’s backbone.The study uses its own PARA implementation for a fair comparison.Source: §3.2 · p. 3
• Two model ablations are included to assess the impact of components in the approach.Source: §3.2 · p. 3

2. IT: $TEXT is transformed to instructions

• IT-MTL is the full proposed model covering all tasks; Table 3 illustrates input prompts for ablations of the approach.Source: §2 · p. 3, Table 3 · p. 4

3.3 Experimental Setup

• The models use t5-base, average results over five random-seed runs, and are evaluated with Micro F1.Source: §3.3 · p. 3

3.4 Results

• Table 2 lists input instruction prompts for all five sub-tasks and identifies $TEXT as the text placeholder.Source: Table 2 · p. 4
• IT-MTL > IT > Text is the reported overall trend, while IT beats Text in most settings.On LAPTOP14, IT-MTL underperforms on AE; the authors suggest this may reflect the absence of TASD and ASQP.Source: §3.4 · p. 4
• On full training data, IT-MTL outperforms or matches SOTA baselines in 4 of 5 tasks, but trails BARTABSA by 3.5 F1 on AE.The authors attribute the AE difference to BARTABSA’s advanced decoding strategies, which they describe as orthogonal to their work.Source: §3.4 · p. 4
• Training is reasonably stable across tasks despite randomness from the seeds; the authors conclude IT-MTL improves the underlying seq-to-seq model’s few-shot generalization across ABSA tasks.Source: §3.4 · p. 4

4 Conclusion

• The authors pose ABSA as instruction-tuned sequence-to-sequence modeling and factorize quadruple prediction into five question-answering-like sub-tasks.They propose a multi-task learning approach using a pre-trained sequence-to-sequence model.Source: §4 · p. 4
• The approach gives superior performance over baselines in few-shot experiments and stays comparable in full fine-tuning scenarios.Experiments use customer reviews from two domains.Source: §4 · pp. 4–5

5 Limitations

• Malformed generative outputs can still reduce accuracy, although fixed-template matching prevents irrelevant generated text from reaching end users.The model relies on a generative language model to capture relationships between sentiment elements.Source: §5 · p. 5
• Arbitrarily long user-generated inputs may increase decoding time because the underlying model is generative.Source: §5 · p. 5
• The instruction templates are designed solely for English, leaving more language-inclusive instruction tuning as future work.Source: §5 · p. 5

A.2 Hyperparameters

• The models train for 20 epochs with learning rate 3e-4, maximum sequence length 160, and batch sizes varying by shot count.Full-shot runs use batch size 16; k=5, 10, 20, and 50 use batch sizes 2, 2, 4, and 8, respectively.Source: §A.2 · p. 8

A.3 Dataset Statistics

• LAPTOP14 few-shot samples are selected by sentiment class rather than aspect category because that dataset lacks category annotations.The same test set is used for few-shot and full-shot evaluation.Source: §A.3 · p. 8, Table 5 · p. 9

A.4 Results on Full Datasets

• Figure 3 reports averaged full-dataset results for REST15, REST16, and LAPTOP14.Source: §A.4 · p. 8

A.5 Implementation Issues

• Evaluation fixes duplicate-tuple counting, reducing PARA’s LAPTOP14 full-corpus F1 from 61.13 to 60.70 and the IT-MTL–PARA gap from 6.75 to 2.32 absolute F1.The authors report that the original logic inflated F1 by counting repeated predicted tuples more than once.Source: §A.5 · p. 8

A.6 Stochasticity in Few-shot Data Sampling

• Across ASQP runs, standard deviations stay under approximately 1.6 F1 points; resampling REST16 few-shot data five times preserves the observed trend.Table 7 reports ASQP results averaged across five k-shot samples.Source: §A.6 · p. 8, Table 7 · p. 8, Table 6 · p. 9