CREAD: Combined Resolution of Ellipses and Anaphora in Dialogues
Summary
Users of dialogue assistants often refer back to earlier entities with pronouns or short descriptions (anaphora), or leave information out entirely (ellipsis). Systems usually handle these separately: coreference resolution for the first and query rewriting for the second. The authors argue that a rewrite alone is not enough. It lacks explicit mention spans and cross-turn links, which help with entity disambiguation and downstream components. No English conversational corpus had both annotation types, so they added rewritten queries to MuDoCo, which already had coreference links.
They then built CREAD, a single GPT-2 model that works in four steps:
- It finds mentions in the latest user query.
- It links each mention to its antecedent in the context.
- It decides whether the query needs rewriting.
- If so, it generates a standalone version of the query.
Training both tasks together improved rewriting by +2.3% F1. The model also beat the previous best coreference system on MuDoCo by about +2% F1. The work shows that reference resolution signals help produce better self-contained queries.
Key points
- Dataset: MuDoCo has about 7.5k task-oriented dialogues across 6 domains, averaging 2.6 turns (maximum 5). Thirty annotators added rewrite labels, using a two-stage process with training sessions and author spot-checks of 5% of the results.
- Release: The augmented dataset and code are available at github.com/apple/ml-cread.
- Ablations:
- Removing the coref2qr attention layer lowered rewrite F1 by 2.9 points and the RM rate by 1.4 points.
- Dropping coreference supervision degraded results further.
- Binary rewrite classifier: It reached 93.9% accuracy. Without it, rewrite F1 fell by up to 5.9 points (60.2 to 54.3).
- Comparison with baselines:
- The joint model is strongest on coreference cases.
- On ellipsis it stays competitive with the QR-only variant.
- An LSTM seq2seq model with a pointer-generator often leaves elliptical queries unchanged.
- Limitation: The model only handles references that form contiguous word spans.
Cite
@inproceedings{tseng2021cread,
title = {{CREAD: Combined Resolution of Ellipses and Anaphora in Dialogues}},
author = {Tseng, Bo-Hsiang and Bhargava, Shruti and Lu, Jiarui and Moniz, Joel Ruben Antony and Piraviperumal, Dhivya and Li, Lin and Yu, Hong},
booktitle = {Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
year = {2021},
doi = {10.18653/v1/2021.naacl-main.265},
eprint = {2105.09914},
archivePrefix = {arXiv},
url = {https://doi.org/10.18653/v1/2021.naacl-main.265}
}References
- Raviteja Anantha, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2020. Open-domain question answering goes conversational via question rewriting. arXiv preprint arXiv:2010.04898. find
- Eric Bengtson and Dan Roth. 2008. Understanding the value of features for coreference resolution. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 294– 303. find our card
- Daniel M Bikel, Vittorio Castelli, Radu Florian, and Ding-jung Han. 2009. Entity linking and slot filling through statistical processing and inference rules. In TAC. find
- Anders Björkelund and Jonas Kuhn. 2014. find
This page in Markdown → https://md.research.elamur.ai/free/en/papers/tseng-2021-cread-combined-resolution-ellipses-anaphora/