Pith. sign in

REVIEW 4 major objections 6 minor 49 references

Contrastive Learning on LLM Back Generation Treebank for Cross-domain Constituency Parsing

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Reversing constituency parsing — feeding an LLM a bare parse skeleton with only domain keywords and letting it fill in words — builds a 10,000-sentence treebank that lifts average F1 on five MCTB domains to 88.52, beating the prior 88.31.

desk verdict The LLM back-generation idea is a genuine twist on prior corpus generation, but the reported SOTA margin is within plausible seed noise and needs a variance estimate before the central claim holds. read the letter →

arxiv 2505.20976 v1 pith:ESTJQLIB submitted 2025-05-27 cs.CL

Peiming Guo , Meishan Zhang , Jianling Li , Min Zhang , Yue Zhang This is my paper
classification cs.CL
keywords cross-domainconstituencyparsingLLMbackgenerationtreebankspan-levelcontrastivelearningMCTBchart-basedparserin-contextdomainvocabulary
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that a large language model can create a useful cross-domain constituency treebank even though the same kind of model is bad at parsing directly. Its method, LLM back generation, reverses the task: instead of asking the model to build a tree for a sentence, it gives the model a bare constituency tree whose non-keyword words have been removed and asks it to fill in plausible words, producing a sentence-tree pair in the target domain. A span-level contrastive pre-training step then teaches a chart parser to tell valid constituent spans from nearby invalid spans on those generated trees. On the five-domain MCTB benchmark the resulting parser reaches 88.52 average F1, above the previous best method's 88.31. If this holds, automatic treebank generation needs only about 10,000 LLM-generated trees rather than hundreds of thousands of raw sentences.

What carries the argument

LLM back generation is the central object: an incomplete constituency tree with only domain-keyword leaf words is given to an LLM, which fills in the missing words to produce a complete sentence and tree. The second piece is span-level contrastive pre-training on the resulting treebank: for every constituent span, the model is trained to pull its span representation close to those of its left child, right child, parent, and brother, and away from fifteen invalid spans with shifted boundaries. These two mechanisms together let a small generated treebank of 10,000 trees, with about 25 spans each and so roughly 250,000 span-level training examples, transfer syntactic knowledge to the chart-based parser.

What would settle it

Re-parse a sample of the back-generated sentences with an independent high-quality parser and measure bracket-overlap between the re-parsed trees and the skeletons that were fed to the LLM; if that overlap is low, the generated treebank is not actually skeleton-conditioned and the reported gains cannot be attributed to LLM back generation.

Watch

Extended reading notes

Core claim

The central discovery is that syntactic skeletons plus domain keywords are a better condition for LLM treebank generation than unconstrained sentence generation. The pipeline parses raw target-domain sentences with the same chart parser family that serves as its base parser, keeps only the tree structure and roughly 25 percent of words chosen by keyword extraction, and feeds the masked tree to GPT-4 with demonstrations; the LLM fills in the missing words and returns a complete tree. The resulting treebank improves every pre-training variant it is tested with, and the new contrastive pre-training—which treats left child, right child, parent, and brother spans as positives and fifteen boundary-shifted invalid spans as negatives—adds further gains, averaging 88.52 F1 across the dialogue, forum, law, literature, and review domains of MCTB.

Load-bearing premise

The approach assumes the out-of-domain parse trees produced by the baseline chart parser are sound enough to act as teaching signals, and the paper does not measure how often the LLM honored the supplied skeleton.

Editorial extensions

If this is right

  • Roughly 10,000 generated trees suffice for cross-domain gains, compared with the 200,000 raw sentences used by the closest prior method, cutting the cost of LLM-based treebank construction.
  • A single parser trained on the generated treebank covers all five MCTB domains, whereas the prior method trains one parser per domain.
  • The gains are not confined to one pre-training setup: the generated treebank outperforms the natural-corpus treebank under domain-adaptive pre-training, no pre-training, and the new contrastive pre-training.
  • When matched to the prior method's settings (GPT-3.5, 200,000 generated trees, per-domain parsers), the proposed approach still averages 88.42 F1 versus 88.31.
  • Source-domain PTB performance is preserved at 95.71 F1, so the cross-domain gains do not come at the expense of in-domain accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The span-level contrastive pre-training is logically independent of LLM back generation; it could be applied to any treebank. That independence means the gains from contrastive pre-training may transfer to human-annotated or other automatically built treebanks, a testable extension the paper does not run.
  • Because the skeletons come from the same chart-parser family as the final parser, part of the gain could be the parser learning its own output distribution. A control using human-annotated or independently produced skeletons would separate the LLM's lexical contribution from the parser's self-consistency.
  • The method is presented as language-agnostic because GPT-4 is multilingual, but all experiments are English-only. The natural next test is a non-English multi-domain treebank to see whether the 25 percent keyword retention and the span-contrastive objective hold under different morphology.
  • The mask-rate curve, where 25 percent masking is best and 0 percent is worst, suggests the LLM's lexical filling does real work beyond resampling the raw corpus; measuring lexical diversity of back-generated sentences against the original domain sentences would sharpen this claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a method for automatic cross-domain constituency treebank construction called LLM back generation. The idea is to take an incomplete constituency tree, produced by running a chart parser on unlabeled target-domain sentences and then removing all words except domain keywords, and prompt an LLM to fill in the missing words while preserving the given syntactic skeleton. The resulting 10,000-tree treebank is used for a span-level contrastive learning pre-training stage, followed by fine-tuning of a Kitaev-Klein style chart parser on PTB plus the generated treebank. Experiments on the five MCTB domains report an average F1 of 88.52, which is claimed to be state-of-the-art compared with baselines including Li et al. (2023). The paper also includes analyses of the mask rate, treebank size, convergence, and a fair comparison with Li et al. under matched settings.

Significance. If the central claim holds, the paper makes a useful contribution: it shows that a relatively small LLM-generated treebank can improve cross-domain constituency parsing, and it introduces a span-level contrastive pre-training objective that appears to help beyond standard fine-tuning. The paper reports several strengths: the code is public, the method is evaluated on a standard five-domain benchmark, a natural-corpus control is included, a matched-setting comparison with Li et al. is provided in Table 5, and the source-domain PTB performance is maintained. However, the main quantitative claim is currently not supported with sufficient statistical rigor, and the quality of the pseudo-parsed skeletons is not directly assessed. These issues are load-bearing because the headline contribution is a numerical state-of-the-art claim and a proposed data-generation pipeline whose value depends on the reliability of its pseudo-labels.

major comments (4)
  1. [§4.2, Table 2, and §A.4] The headline claim of state-of-the-art average F1 (88.52 vs. 88.31 for Li et al.) is not supported by the reported statistics. The authors state that experiments were run with three random seeds, but no standard deviations, confidence intervals, or per-seed scores are reported anywhere. The only significance statement is restricted to three of five domains (dialogue, forum, review), and in the two remaining domains (law and literature) the proposed method is numerically below Li et al. in Table 2. The average gap is 0.21 F1, and in the matched setting of Table 5 (same GPT-3.5-turbo, 200k sentences, per-domain parser) the gap shrinks to 0.11. Without variance estimates or a joint test over all five domains, the headline SOTA claim is indistinguishable from seed-level noise. Please report per-seed scores and standard deviations for the main results, and either provide a significance test covering all five domains or soften the state-of-the-art claim.
  2. [§3.1, Figure 2] The syntactic skeletons in the LLM back generation treebank are produced by running the same Kitaev-Klein chart parser that is later improved, and the paper reports no quality check on these pseudo-parses or on the LLM's fidelity to the masked skeletons. If the out-of-domain parses are noisy, the generated treebank inherits those errors, and part of the observed gain may be self-distillation toward the prior parser's output distribution rather than acquisition of genuinely new target-domain syntax. This does not invalidate the relative comparison against the natural-corpus control, since that control uses the same parser for its skeletons, but it is load-bearing for the interpretation of the treebank-generation contribution. Please report parse-quality statistics on the target-domain raw sentences (e.g., F1 against a held-out gold set or human evaluation on a sample), and verify that the LLM-generated trees preserve the masked skeleton structure at a high rate.
  3. [§4.2 and Table 5] The main comparison with Li et al. (2023) in Table 2 is not matched: the proposed method uses GPT-4-1106-preview and 10k generated sentences with a single shared parser, while Li et al. use GPT-3.5-turbo, 200k sentences, and a separate parser per domain. The fair comparison in Table 5 uses the same settings and still reports a gain, which is reassuring, but no significance test or variance estimate is given for that table either; the text only says 'p < 0.05' without specifying the test, the number of seeds, or which domains are included. Please provide the statistical details for Table 5 and make clear in the main results section that the headline comparison is not apples-to-apples.
  4. [§3.3 and Table 1] The contrastive pre-training contribution would benefit from ablations that isolate its effect from the choice of positive and negative instances. The paper compares CTPT against DAPT and NOPT, which shows the combined effect of the contrastive objective, but it does not test, for example, a non-contrastive span-classification pre-training objective on the same treebank, nor does it ablate the four positive-instance types or the set of fifteen negatives. Such ablations are not strictly necessary for the paper's central claim, but they would strengthen the evidence that the specific contrastive formulation, rather than pre-training on the treebank in any form, drives the improvement over NOPT.
minor comments (6)
  1. [Figure 2] The figure contains the typo 'pround' in both the target domain constituency tree and the masked tree; it should be 'proud'.
  2. [Related Work] The word 'retrivel' appears in the sentence about contrastive learning applications; it should be 'retrieval'.
  3. [Table 2] In the CTPT and Li et al. rows, several F1 values are run together without spaces (e.g., '93.22 87.50' is rendered as '93.2287.50'); please fix the formatting for readability.
  4. [§4.3, Figure 6] The y-axis label says 'F1 score' but the text describes average F1 over five domains; please make the axis label consistent with the text.
  5. [§A.2] The appendix states that a BERT-LSTM-CRF model is trained for POS tagging from PTB, but it does not report this model's accuracy or how its errors might propagate into the LLM back generation prompts; a sentence noting the POS tagger accuracy would be helpful.
  6. [§4.1] The sentence 'We conduct the experiments on three different random seeds and report the average results' should mention that no variance is reported; at minimum, please add standard deviations to Table 2.

Circularity Check

0 steps flagged · score 1.0 of 10

No by-construction circularity; the generated-treebank skeleton comes from the same parser family, but evaluation on held-out gold MCTB and the natural-corpus control make this a transparent self-training design, not a circular derivation.

full rationale

The only potential circularity is that the LLM back generation treebank's syntactic skeletons are produced by the same Kitaev-Klein chart parser that the paper later fine-tunes (Section 3.1: 'we first train the state-of-the-art chart-based parser (Kitaev and Klein, 2018) and then parse the constituency tree corresponding to the unlabeled target domain raw sentence'; Section 3.3: 'we equip the constituent representation model with max-margin tree loss and fine-tune it on the combination of the limited source domain human annotation treebank and the target domain LLM back generation treebank'). This is a pseudo-labeling / self-distillation setup, not a by-construction reduction: the final parser is evaluated on held-out gold MCTB test trees, and the natural-corpus treebank control is parsed with the same basic parser, so the comparison isolates the LLM's contribution to surface generation. No equation in the paper equates the reported F1 prediction with a fitted input, and no fitted parameter is renamed as a prediction. The SOTA claim's 0.21-point average margin and the significance test restricted to three of five domains are statistical robustness concerns, not circularity. Self-citations (Li et al. 2023, Guo et al. 2024, Yang et al. 2022) are used as baselines or motivation, not as load-bearing justification, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the method's choice. The derivation chain is therefore self-contained with respect to the external benchmark; any weakness is empirical, not circular.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central method rests on a small number of tuned hyperparameters and on assumptions about the quality of parser-produced skeletons and LLM fidelity. No new theoretical entities are proposed; the 'LLM back generation treebank' is a dataset artifact, not an invented entity.

free parameters (5)
  • mask_rate = 0.25 (25% of words retained as domain keywords)
    Selected via Figure 4; 25% gives the best average F1 on MCTB among 0/25/50/100% rates, so the main result depends on this tuning choice.
  • contrastive_temperature_tau = 0.05
    Hyperparameter for the contrastive objective (Appendix A.2); no sensitivity analysis reported.
  • contrastive_sample_ratio = 20% of constituents per tree
    In each batch only 20% of constituents are sampled as contrastive examples; chosen without reported ablation.
  • num_demonstrations = 2
    LLM prompt uses 2 demonstration pairs (Appendix A.2); no ablation on demonstration count.
  • generated_treebank_size = 10,000 trees
    Cost-limited scale; Figure 6 shows performance still increasing at 10k, so the final number might not be the saturation point.
assumptions (4)
  • domain assumption The chart parser (Kitaev and Klein) produces adequate syntactic skeletons for in-the-wild target-domain sentences.
    Section 3.1 builds every generated tree on such parses; if these are poor, the whole treebank inherits the errors.
  • domain assumption GPT-4 follows instructions and preserves the masked tree structure while filling in fluent words.
    The method assumes the LLM output tree has the same labeled skeleton and only different lexemes; no validity rate or filtering is reported.
  • domain assumption Span-level contrastive learning on the synthetic treebank transfers to labeled parsing performance.
    CTPT gains over NOPT are reported, but the mechanism is not otherwise verified (e.g., no direct span-quality metric).
  • domain assumption KeyBERT keyword selection identifies words that carry domain signal.
    Domain keywords are selected by embedding similarity and retained in the mask; if the selector picks generic words, the domain signal would be lost.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Learning on LLM Back Generation Treebank for Cross-domain Constituency Parsing." pith.science (2026). https://pith.science/paper/ESTJQLIB

@misc{pith2026250520976,
  author       = {Pith},
  title        = {Pith review of: Contrastive Learning on LLM Back Generation Treebank for Cross-domain Constituency Parsing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ESTJQLIB}},
  note         = {Machine review of arXiv:2505.20976}
}
read the original abstract

Cross-domain constituency parsing is still an unsolved challenge in computational linguistics since the available multi-domain constituency treebank is limited. We investigate automatic treebank generation by large language models (LLMs) in this paper. The performance of LLMs on constituency parsing is poor, therefore we propose a novel treebank generation method, LLM back generation, which is similar to the reverse process of constituency parsing. LLM back generation takes the incomplete cross-domain constituency tree with only domain keyword leaf nodes as input and fills the missing words to generate the cross-domain constituency treebank. Besides, we also introduce a span-level contrastive learning pre-training strategy to make full use of the LLM back generation treebank for cross-domain constituency parsing. We verify the effectiveness of our LLM back generation treebank coupled with contrastive learning pre-training on five target domains of MCTB. Experimental results show that our approach achieves state-of-the-art performance on average results compared with various baselines.

Figures

Figures reproduced from arXiv: 2505.20976 by the authors.

Figure 1
Figure 1. LLM constituency parsing usually predicts b [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of LLM Back Generation. We first extract the target domain constituency tree and domain [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Positive instances (green node) and indexes for [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: F1 score of three pre-training strategies on [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: F1 score of contrastive learning pre-training [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: The prompt and output of LLM back generation. The blue texts are only shown for illustrative purposes [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: A constituency tree. A.5 Performance on Source Domain Though our approach achieves impressive results on target domain, performances on source domain standard benchmarks are also important. We test our final parser on the PTB dataset, which achieves an F1-score of 95.7…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 32 canonical work pages

  1. [1]

    Xuefeng Bai, Jialong Wu, Yulong Chen, Zhongqing Wang, and Yue Zhang. 2023. https://arxiv.org/abs/2310.19462 Constituency parsing using llms . arXiv

    arXiv 2023

  2. [2]

    Parikshit Bansal and Amit Sharma. 2023. https://arxiv.org/abs/2306.15766 Large language models as annotators: Enhancing generalization of nlp models at minimal cost . arXiv

    arXiv 2023

  3. [3]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf Language models are few-shot learners . In Advances in neural information processing systems

    work page 2020

  4. [4]

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. 2020. https://proceedings.neurips.cc/paper/2020/hash/70feb62b69f16e0238f741fab228fec2-Abstract.html Unsupervised learning of visual features by contrasting cluster assignments . Advances in neural information processing systems

    work page 2020

  5. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. https://proceedings.mlr.press/v119/chen20j.html A simple framework for contrastive learning of visual representations . In Proceedings of the 37th International Conference on Machine Learning

    work page 2020

  6. [6]

    Sumit Chopra, Raia Hadsell, and Yann LeCun. 2005. https://doi.org/10.1109/CVPR.2005.202 Learning a similarity metric discriminatively, with application to face verification . In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05)

    doi:10.1109/cvpr.2005.202 2005

  7. [7]

    Michael Collins. 1997. https://doi.org/10.3115/976909.979620 Three generative, lexicalised models for statistical parsing . In 35th Annual Meeting of the Association for Computational Linguistics and 8th Conference of the E uropean Chapter of the Association for Computational Linguistics

    work page arXiv 1997

  8. [8]

    Leyang Cui, Sen Yang, and Yue Zhang. 2022. https://aclanthology.org/2022.acl-long.146 Investigating non-local features for neural constituency parsing . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics

    work page 2022

Show all 49 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://www.aclweb.org/anthology/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Ass...

    2019

  2. [10]

    Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. https://openreview.net/forum?id=r1l73iRqKm Wizard of wikipedia: Knowledge-powered conversational agents . In International Conference on Learning Representations

    2019

  3. [11]

    Daniel Fried, Nikita Kitaev, and Dan Klein. 2019. https://aclanthology.org/P19-1031 Cross-domain generalization of neural constituency parsers . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics

    2019

  4. [12]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.552 S im CSE : Simple contrastive learning of sentence embeddings . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing

    2021 doi

  5. [13]

    Maarten Grootendorst. 2020. https://doi.org/10.5281/zenodo.4461265 Keybert: Minimal keyword extraction with bert

    2020 doi

  6. [14]

    Peiming Guo, Meishan Zhang, Yulong Chen, Jianling Li, Min Zhang, and Yue Zhang. 2024. Cross-domain constituency parsing by leveraging heterogeneous data. Journal of Artificial Intelligence Research, 81:771--791

    2024

  7. [15]

    Suchin Gururangan, Ana Marasovi \'c , Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. https://aclanthology.org/2020.acl-main.740 Don ' t stop pretraining: Adapt language models to domains and tasks . In Proceedings of the 58th Annual Meeting of t...

    2020

  8. [16]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. https://openaccess.thecvf.com/content_CVPR_2020/html/He_Momentum_Contrast_for_Unsupervised_Visual_Representation_Learning_CVPR_2020_paper.html Momentum contrast for unsupervised visual representation learni...

    2020

  9. [17]

    Ruining He and Julian McAuley. 2016. https://doi.org/10.1145/2872427.2883037 Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering . In proceedings of the 25th international conference on world wide web

    2016

  10. [18]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  11. [19]

    Vidur Joshi, Matthew Peters, and Mark Hopkins. 2018. https://aclanthology.org/P18-1110 Extending a parser to distant domains using a few dozen partially annotated examples . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics

    2018

  12. [20]

    Nikita Kitaev, Steven Cao, and Dan Klein. 2019. https://aclanthology.org/P19-1340 Multilingual constituency parsing with self-attention and pre-training . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics

    2019

  13. [21]

    Nikita Kitaev and Dan Klein. 2018. https://aclanthology.org/P18-1249 Constituency parsing with a self-attentive encoder . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics

    2018

  14. [22]

    Jianling Li, Meishan Zhang, Peiming Guo, Min Zhang, and Yue Zhang. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.508 LLM -enhanced self-training for cross-domain constituency parsing . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processin...

    2023 doi

  15. [23]

    Jiangming Liu and Yue Zhang. 2017. https://doi.org/10.1162/tacl\_a\_00070 In-order transition-based constituent parsing . Transactions of the Association for Computational Linguistics

    2017 doi

  16. [24]

    Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz

    Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. https://aclanthology.org/J93-2004 Building a large annotated corpus of E nglish: The P enn T reebank . Computational Linguistics, 19(2):313--330

    1993

  17. [25]

    David McClosky, Eugene Charniak, and Mark Johnson. 2006. https://aclanthology.org/N06-1020 Effective self-training for parsing . In Proceedings of the Human Language Technology Conference of the NAACL , Main Conference

    2006

  18. [26]

    David McClosky, Eugene Charniak, and Mark Johnson. 2008. https://aclanthology.org/C08-1071 When is self-training effective for parsing? In Proceedings of the 22nd International Conference on Computational Linguistics

    2008

  19. [27]

    David McClosky, Eugene Charniak, and Mark Johnson. 2010. https://aclanthology.org/N10-1004 Automatic domain adaptation for parsing . In Human Language Technologies: The 2010 Annual Conference of the North A merican Chapter of the Association for Computational Linguistics

    2010

  20. [28]

    OpenAI. 2023. https://doi.org/10.48550/arXiv.2303.08774 Gpt-4 technical report . arXiv

    2023 doi arXiv

  21. [29]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/b1efde53be364a73914f58805a001731-Paper-Conference.pdf Training...

    2022

  22. [30]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. https://proceedings.mlr.press/v139/radford21a.html Learning transferable visual model...

    2021

  23. [31]

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. 2015. https://www.cv-foundation.org/openaccess/content_cvpr_2015/html/Schroff_FaceNet_A_Unified_2015_CVPR_paper.html Facenet: A unified embedding for face recognition and clustering . In Proceedings of the IEEE conferenc...

    2015

  24. [32]

    Tianyu Shi, Zhicheng Wang, Liyin Xiao, and Cong Liu. 2022. https://arxiv.org/abs/2212.08458 Fast rule-based decoding: Revisiting syntactic rules in neural constituency parsing . arXiv

    2022 arXiv

  25. [33]

    Harman Singh, Pengchuan Zhang, Qifan Wang, Mengjiao Wang, Wenhan Xiong, Jingfei Du, and Yu Chen. 2023. https://aclanthology.org/2023.emnlp-main.56 Coarse-to-fine contrastive learning in image-text-graph space for improved vision-language compositionality . In Proceedings of th...

    2023

  26. [34]

    Kihyuk Sohn. 2016. https://proceedings.neurips.cc/paper/2016/hash/6b180037abbebea991d8b1232f8a8ca9-Abstract.html Improved deep metric learning with multi-class n-pair loss objective . Advances in neural information processing systems

    2016

  27. [35]

    Mitchell Stern, Jacob Andreas, and Dan Klein. 2017. https://aclanthology.org/P17-1076 A minimal span-based neural constituency parser . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics

    2017

  28. [36]

    yvind Stiansen and Erik Voeten. 2019. https://doi.org/10.7910/DVN/OBYUO5 ECtHR judgments . Harvard Dataverse

    2019 doi

  29. [37]

    Zhiyang Teng and Yue Zhang. 2018. https://aclanthology.org/C18-1011 Two local models for neural constituent parsing . In Proceedings of the 27th International Conference on Computational Linguistics

    2018

  30. [38]

    Yuanhe Tian, Yan Song, Fei Xia, and Tong Zhang. 2020. https://aclanthology.org/2020.findings-emnlp.153 Improving constituency parsing with span attention . In Findings of the Association for Computational Linguistics: EMNLP 2020

    2020

  31. [39]

    Petter T \"o rnberg. 2023. https://arxiv.org/abs/2304.06588 Chatgpt-4 outperforms experts and crowd workers in annotating political twitter messages with zero-shot learning . arXiv preprint arXiv:2304.06588

    2023 arXiv

  32. [40]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. https://arxiv.org/abs/2307.09288 Llama 2: Open foundation and fine-tuned chat models . arXiv preprint

    2023 arXiv

  33. [41]

    Michael V \"o lske, Martin Potthast, Shahbaz Syed, and Benno Stein. 2017. https://aclanthology.org/W17-4508 TL ; DR : Mining R eddit to learn automatic summarization . In Proceedings of the Workshop on New Frontiers in Summarization

    2017

  34. [42]

    Ziqi Wang, Xiaozhi Wang, Xu Han, Yankai Lin, Lei Hou, Zhiyuan Liu, Peng Li, Juanzi Li, and Jie Zhou. 2021. https://aclanthology.org/2021.acl-long.491 CLEVE : C ontrastive P re-training for E vent E xtraction . In Proceedings of the 59th Annual Meeting of the Association for Co...

    2021

  35. [43]

    Junhan Yang, Zheng Liu, Chaozhuo Li, Guangzhong Sun, and Xing Xie. 2023. https://aclanthology.org/2023.emnlp-main.223 Longtriever: a pre-trained long text encoder for dense document retrieval . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

    2023

  36. [44]

    Sen Yang, Leyang Cui, Ruoxi Ning, Di Wu, and Yue Zhang. 2022. https://aclanthology.org/2022.findings-acl.11 Challenges to open-domain constituency parsing . In Findings of the Association for Computational Linguistics: ACL 2022

    2022

  37. [45]

    Ruoyu Zhang, Yanzeng Li, Yongliang Ma, Ming Zhou, and Lei Zou. 2023 a . https://aclanthology.org/2023.findings-emnlp.872 LLM a AA : Making large language models as active annotators . In Findings of the Association for Computational Linguistics: EMNLP 2023

    2023

  38. [46]

    Xin Zhang, Zehan Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, and Min Zhang. 2023 b . https://arxiv.org/abs/2310.08232 Language models are universal embedders . arXiv

    2023 arXiv

  39. [47]

    Yu Zhang, Houquan Zhou, and Zhenghua Li. 2020. https://doi.org/10.24963/ijcai.2020/560 Fast and accurate neural crf constituency parsing . In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence

    2020 doi

  40. [48]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  41. [49]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.