Elamur Research⌕
☰
home / papers / 2018 / Deep Reinforcement Learning for Chinese Zero Pronoun Resolution

Deep Reinforcement Learning for Chinese Zero Pronoun Resolution

Summary

Chinese often omits pronouns, leaving "zero pronouns" whose referents must be recovered from context. Earlier neural resolvers score each pair of a zero pronoun and one candidate antecedent in isolation. They therefore ignore what earlier decisions reveal about later ones. Yin and colleagues recast anaphoric zero pronoun resolution as a Markov decision process. A policy-network agent walks through the candidate noun phrases one by one and decides whether each one is an antecedent. At every step, the agent's state includes a pooled summary of the candidates it has already accepted. This lets earlier choices inform later ones, such as rejecting a nearby mention that clashes with the entity already identified. The agent is pre-trained with supervision and then optimized with a reward equal to the F-score of its selected antecedents. It therefore targets the evaluation metric directly rather than a per-decision loss. On OntoNotes 5.0, the model beats prior state-of-the-art systems. The authors present it as the first reinforcement learning approach to Chinese zero pronoun resolution.

Key points

  • Candidate selection: candidates are maximal or modifier noun phrases located at most two sentences from the zero pronoun. This follows the heuristic of Chen and Ng (2016) and Yin et al. (2017a,b).
  • Encoders: the zero pronoun is encoded with the ZP-centered neural network of Yin et al. (2017a), which uses two RNNs over the preceding and following context. Each candidate phrase is encoded by a left-to-right RNN over its content words.
  • State vector: each state combines four parts:
    • the zero pronoun vector,
    • the current candidate vector,
    • pooled information from previously predicted antecedents,
    • pairwise features.
  • Learning: the authors chose a stochastic policy network trained with SGD over a Deep Q-Network. Their reasons were to avoid getting stuck in intermediate states and to obtain a more interpretable policy. Training starts with supervised pre-training.
  • Robustness: across 38 random seeds, F-scores ranged from 56.5% to 57.5%, with a median of 57.1%.

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

Topics: Coreference resolution

Cited by

16 papers of our library cite it (of the 140673 whose bibliographies we read); 3 of them have a page here, 13 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/p18-1053 · arXiv 1806.03711

@inproceedings{yin2018deep,
  title = {{Deep Reinforcement Learning for Chinese Zero Pronoun Resolution}},
  author = {Yin, Qingyu and Zhang, Yu and Zhang, Wei-Nan and Liu, Ting and Wang, William Yang},
  booktitle = {Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2018},
  doi = {10.18653/v1/p18-1053},
  eprint = {1806.03711},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/p18-1053}
}

References

3 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Chen Chen and Vincent Ng. 2014. Chinese zero pronoun resolution: An unsupervised approach combining ranking and integer linear programming. In Twenty-Eighth AAAI Conference on Artificial Intelligence. find
  2. SRK Branavan, David Silver, and Regina Barzilay. 2012. Learning to win by reading manuals in a monte-carlo framework. Journal of Artificial Intelligence Research, 43:661–704. find
  3. Chen Chen and Vincent Ng. 2013. Chinese zero pronoun resolution: Some recent advances. In EMNLP, pages 1360–1365. \#\# Conclusion find our card

This page in Markdown → https://md.research.elamur.ai/free/papers/yin-2018-deep-reinforcement-learning-chinese-zero/