Elamur Research⌕
☰
home / papers / 2024 / EDDA: An Encoder-Decoder Data Augmentation Framework for Zero-Shot Stance Detection

EDDA: An Encoder-Decoder Data Augmentation Framework for Zero-Shot Stance Detection

Summary

The paper addresses zero-shot stance detection (ZSSD), where a model must classify an attitude (favor, against, neutral) towards targets it never saw in training. The authors identify two weaknesses in existing data augmentation for ZSSD. Target augmentation can produce new targets that have no real semantic link to the source text. Text augmentation draws only on training data, so the synthetic samples drift back toward the original distribution and generalize poorly.

Their EDDA framework has two stages:

  • Encoder: an LLM with one-shot chain-of-thought prompting condenses each training example into a target-specific "if (reason) then (stance)" rationale.
  • Decoder: the LLM writes new samples from these rationales. Before generation, emotion words in the reason part are randomly swapped using SenticNet, which yields more varied syntax while keeping the logic intact.

A rationale-enhanced network (REN) is then trained to exploit the augmented data. On standard benchmarks, the authors report substantial gains over state-of-the-art ZSSD methods. The approach also adds interpretability, because predictions are tied to explicit rationales.

Key points

  • The if-then rationales come from a one-shot prompt that frames the task as inserting "[RULE: If (A) then (B)]" API calls. A second prompt asks the LLM for one to three topics linked to each rationale.
  • The decoder's random replacement strategy (RRS), inspired by Huang et al. (2023), uses the SenticNet lexicon. It prevents near-identical outputs from the same rationale.
  • Baselines include TDDA (Xu et al.'s prompt-based augmentation re-implemented with GPT-3.5) and TarBK. The authors note that prior text augmentation gives only marginal improvements.
  • Evaluation relies on widely used ZSSD benchmarks; the cited datasets include VAST and SemEval-2016 Task 6. Exact scores were not in the supplied excerpt.
  • The if-then representation is model-agnostic and can be plugged into other stance detectors. Code is available at github.com/Szu-Ddj/EDDA. The paper is licensed CC BY-NC 4.0.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{ding2024edda,
  title = {{EDDA: An Encoder-Decoder Data Augmentation Framework for Zero-Shot Stance Detection}},
  author = {Ding, Daijun and Dong, Li and Huang, Zhichao and Xu, Guangning and Huang, Xu and Liu, Bo and Jing, Liwen and Zhang, Bowen},
  booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  year = {2024},
  doi = {10.63317/5fvfixthhpye},
  url = {https://doi.org/10.63317/5fvfixthhpye}
}

References

2 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Emily Allaway and Kathleen Mckeown. 2020. Zero-shot stance detection: A dataset and model using generalized topic representations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913–8931. find our card
  2. Emily Allaway, Malavika Srikanth, and Kathleen Mckeown. 2021. Adversarial learning for zero-shot stance detection on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4756–4767. *Tabl… find our card

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ding-2024-edda-encoder-decoder-data-augmentation/