Pith. sign in

REVIEW 4 major objections 4 minor 49 references

Def-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation

T0 review · 4 major / 4 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 Def-DTS claims that a three-step deductive reasoning prompt—context summary, intent classification, and a deterministic topic-shift rule—outperforms trained segmentation models on open-domain dialogue benchmarks without task-specific…

desk verdict Training-free prompt recipe with a real TIAGE win, but the Dialseg711 SOTA claim rests on omitting the stronger turn-based S3-DST baseline and the paper's own ablation undercuts the mechanism. read the letter →

arxiv 2505.21033 v1 pith:YIEG72NG submitted 2025-05-27 cs.CL

Seungmin Lee , Yongsang Yoo , Minhwa Jung , Min Song This is my paper
classification cs.CL
keywords dialoguetopicsegmentationlargelanguagemodelsdeductivereasoningutteranceintentclassificationpromptengineeringshiftdetectionauto-labelingopen-domain
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

This paper argues that dialogue topic segmentation—deciding where a conversation changes topic—can be solved by teaching a large language model to reason in three steps instead of training a special-purpose segmenter. The steps are: summarize a short window of context before and after the target utterance, decide which of five general intents the utterance serves, and then read the answer off a deterministic rule (introduce-topic or change-topic means a shift; otherwise not). The authors claim this prompt chain, called Def-DTS, outperforms prior unsupervised and supervised baselines on the TIAGE benchmark, sets a new state of the art on Dialseg711, and beats all unsupervised methods on SuperDialseg, while needing no task-specific training. If true, the result matters because it turns a data-hungry, label-ambiguous task into an inspectable reasoning task that any instruction-following LLM can run.

What carries the argument

The load-bearing object is the general intent pool: five mutually exclusive utterance intents, each defined by a description and an illustrative dialogue, plus the hard-coded deduction rule 'intent in {INTRODUCE_TOPIC, CHANGE_TOPIC} implies topic shift; otherwise no shift' (Algorithm 1). The intent labels are deliberately domain-agnostic for open dialogue and are swapped per dataset, with document-grounded intents for SuperDialseg and no INTRODUCE_TOPIC for Dialseg711. Around this sits an XML structured prompt that outputs the intent and label for every utterance, and a fixed bidirectional context window (two turns before, three after) that is summarized before classification. Together they convert topic segmentation from a fuzzy similarity judgement into a transparent two-class decision that an LLM can execute from instructions alone.

What would settle it

Take a set of dialogues whose gold topic boundaries fall inside utterances that answer or comment, for example an answer that first replies and then pivots to a new subject. Since Algorithm 1 maps JUST_ANSWER and JUST_COMMENT to NO by construction, Def-DTS cannot mark those boundaries as shifts; checking whether such boundaries exist in TIAGE and are systematically missed would settle the completeness claim.

Watch

Extended reading notes

Core claim

The central claim is that the hard part of dialogue topic segmentation is not scoring topical coherence but naming what an utterance is doing. Def-DTS asks the LLM to classify each utterance into one of five intents—JUST_COMMENT, JUST_ANSWER, DEVELOP_TOPIC, INTRODUCE_TOPIC, CHANGE_TOPIC—after summarizing two preceding and three following utterances, and then maps INTRODUCE_TOPIC and CHANGE_TOPIC to a topic shift and everything else to no shift. Using gpt-4o as the main model, the paper reports on TIAGE a $P_k$ of 0.232, a WD of 0.256, and an $F_1$ of 0.699, outperforming supervised BERT, RoBERTa, and RetroTS-T5; on Dialseg711 it reports $P_k$ 0.015 and $F_1$ 0.979; on SuperDialseg it is the best unsupervised method, though below supervised models. Ablations show each stage contributes, with intent classification the largest single contributor: removing it drops TIAGE $F_1$ from 0.699 to 0.524. The paper also reports that the same pipeline agrees with human labels on TIAGE at Cohen's kappa 0.485, higher than the 0.479 agreement between human annotators, and treats this as evidence the method could serve as a minimal auto-labeler.

Load-bearing premise

The five intents (or their per-dataset variants) are complete enough that every real topic shift shows up as INTRODUCE_TOPIC or CHANGE_TOPIC, so the hard-coded rule cannot miss a shift that the LLM labels otherwise.

Editorial extensions

If this is right

  • A training-free, prompt-only pipeline can outperform fine-tuned segmentation models on at least two public benchmarks, which would shift effort in DTS from architecture design to prompt and intent-pool design.
  • The deterministic intent-to-label map makes every prediction auditable: a user can see the intent behind a YES/NO decision, enabling error analysis and manual correction.
  • Since the intent pool is domain-agnostic and only the descriptions and examples change per dataset, the method can be adapted to new dialogue settings by editing a prompt rather than retraining a model.
  • The kappa results suggest LLM reasoning can produce labels of human-level quality on TIAGE and near-perfect quality on Dialseg711, making semi-automated dataset construction a realistic use case.

Reading between the lines

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

  • The ceiling of Def-DTS is set by the intent classifier's recall on shift-bearing utterances, not by the reasoning prompt: because the final rule is hard-coded, any genuine shift that the model files under JUST_COMMENT or JUST_ANSWER is unrecoverable, and the paper's own Table 6 shows exactly this failure pattern.
  • A natural stress test the paper does not run is longer-range topic shifts: the fixed two-back, three-forward window may hide shifts that only become visible over several turns, so an adaptive or memory-augmented context stage would be a direct extension.
  • The distinct success on Dialseg711 (clear, synthetic boundaries) versus moderate agreement on TIAGE hints that the method is strongest where topic shifts are linguistically marked; naturalistic implicit shifts may need richer intent taxonomies or a probabilistic mapping instead of a hard rule.
  • The auto-labeling result could be turned into a human-in-the-loop workflow: accept Def-DTS proposals on easy dialogues and route low-confidence utterances to annotators; the paper only reports kappa, not such a protocol.
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 / 4 minor

Summary. The paper proposes Def-DTS, an LLM-prompting method for utterance-level dialogue topic segmentation. The method structures each utterance through three subtasks in a single prompt: bidirectional context summarization, classification of the utterance into a hand-crafted intent pool, and a deterministic topic-shift label derived from the intent. The approach is evaluated on TIAGE, SuperDialseg, and Dialseg711 with Pk, WD, and F1, including comparisons with unsupervised, supervised, and LLM-based baselines, ablations, experiments with local and closed-source LLMs, and a preliminary auto-labeling analysis. The paper claims that Def-DTS consistently outperforms traditional and state-of-the-art approaches, achieves state-of-the-art results on TIAGE and Dialseg711, and that each subtask contributes to the overall improvement.

Significance. If the main empirical claims hold, the paper would show that a zero-training prompting method can outperform strong supervised baselines on TIAGE, which is an interesting and potentially useful result for the DTS community. The paper has real strengths: the code and prompts are publicly released, the experiments cover multiple LLMs, the ablation study is reported on all three datasets, and a statistical linguistic test and an annotation-agreement analysis provide additional evidence. However, the broad state-of-the-art claim is overstated as written, and one of the central ablation conclusions is contradicted by the paper's own table on Dialseg711. The core idea is defensible after these claims are qualified and the missing comparison is addressed.

major comments (4)
  1. [Abstract and §4.4, Table 3] The abstract and Section 4.4 claim that Def-DTS "consistently outperforms traditional and state-of-the-art approaches," but on SuperDialseg the supervised RoBERTa baseline achieves F1=0.784 and BERT achieves F1=0.725, while Def-DTS achieves F1=0.686. The claim should be restricted to LLM-based and unsupervised methods on this dataset, or the comparison should be qualified explicitly.
  2. [§4.4, Tables 3 and 15] The state-of-the-art claim on Dialseg711 is not supported by the comparisons shown in Table 3. Table 15 reports the original turn-based S3-DST with Pk=0.009 and WD=0.008 on Dialseg711, both better than Def-DTS's Pk=0.015 and WD=0.018, yet Table 3 compares only the utterance-level reimplementation S3-DST uttr. The authors should either include the original configuration in the main comparison or revise the SOTA claim to the error metrics on which it actually holds.
  3. [Appendix B.1, Table 11] In the Dialseg711 ablation, the w/o all variant (F1=0.987) and the w/o context variant (F1=0.990) both outperform the full Def-DTS model (F1=0.982), and w/o examples is also better on Pk and F1. This directly contradicts the paper's statements that each subtask contributes to improved performance and that the intent classification module elevates performance across all datasets. The ablation claim needs to be qualified to the datasets that actually support it, and the Dialseg711 result should be discussed as a limitation of the multi-step design rather than evidence for it.
  4. [§3.4, §3.5, Algorithm 1, Table 1] The intent descriptions in Table 1 already contain the topic-shift label within them, e.g., "Not a topic shift" and "A topic shift," and Algorithm 1's CLASSIFY TOPIC SHIFT is a deterministic lookup that maps INTRODUCE_TOPIC and CHANGE_TOPIC to YES and all other intents to NO. The so-called deductive topic shift detection is therefore not an independent reasoning step; the topic-shift decision is effectively made during intent classification. The paper should state this explicitly and explain what the multi-step framing contributes beyond prompt structuring, since the current wording overstates the deductive component.
minor comments (4)
  1. [§5.6 and Table 7] The model name is misspelled as "Lama 3.1" in the text; it should be "Llama 3.1."
  2. [Appendix A.2 and Table 12] The intent label is written as "RELEV ANT_QUESTION" in Table 12 and as "RELEVANT_QUESTION" in the text; the spelling should be consistent.
  3. [Abstract] The phrase "On the other hand, Despite advances" contains a capitalization and punctuation error; "Despite" should not be capitalized mid-sentence.
  4. [§5.3, Table 6] The caption or surrounding text should make more prominent that the intents have no gold standard and that TP/FP/TN/FN are defined relative to the alignment of the predicted intent with the topic-shift label, since this is essential for interpreting the confusion matrix.

Circularity Check

1 steps flagged · score 4.0 of 10

Topic-shift label is read off intent definitions that already encode the label; the rest of the evaluation is independent evidence.

  1. self definitional [Section 3.4, Table 1; Section 3.5; Algorithm 1 (lines 11-13)]
    "Using this list, we categorize the utterance and deductively classify the topic shift label. ... INTRODUCE TOPIC: Introducing a relevant but different topic. A topic shift ... CHANGE TOPIC: Completely changing the topic. A topic shift ... return x ∈ {"introduce_topic", "change_topic"} ? "YES" : "NO""

    The final topic-shift label is not derived from independent premises: the intent definitions already contain the target classification. INTRODUCE_TOPIC and CHANGE_TOPIC are defined as 'A topic shift', the other three as 'Not a topic shift', and Algorithm 1 then maps membership in the first two sets to YES. Therefore the 'deductive topic shift classification' is a direct read-off of the intent label, and the intent label is itself a topic-shift judgment. The claimed two-stage deduction reduces to a single LLM classification whose label space was constructed from the target variable.

full rationale

The paper does not fit parameters to gold topic-shift labels, and its comparisons against external baselines are real, falsifiable evidence, so this is not a fitted-input case. The main circularity is structural: the hand-crafted intent pool attaches 'A topic shift' or 'Not a topic shift' to each category, and the supposedly deductive final step is a deterministic lookup over those categories. This makes the multi-step reasoning partially equivalent to its input intent classification by construction, though the intent classification itself is a genuine LLM prediction. A separate correctness concern, not a circularity finding, is that the Dialseg711 SOTA claim omits the original turn-based S3-DST numbers reported in Table 15 (Pk=0.009, WD=0.008 vs. Def-DTS 0.015/0.018), and the paper's own Table 11 shows w/o all and w/o context outperform the full method on that dataset, which the paper attributes to the dataset having clear topic-shift signals.

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

The method has no learned parameters or fitted constants. Its design choices, especially the intent pool and context window, are hand-picked and unvalidated across alternative choices, but they are not fitted to the test labels. No new physical or formal entities are introduced.

free parameters (4)
  • bidirectional context window = preceding: 2 utterances; subsequent: 3 utterances
    Hand-chosen in Section 3.3; not fitted to data but affects all results.
  • utterance intent pool = TIAGE: 5 intents; SuperDialseg: 4 intents; Dialseg711: 4 intents (INTRODUCE_TOPIC removed)
    Hand-crafted categories (Table 1, Appendix A.2) that encode topic shift information; adapted per dataset.
  • few-shot intent examples = TIAGE: examples from Xie et al.; others: random train-split utterances (2-3 turns, <=100 chars)
    Selected by hand or simple heuristic (Appendix A.4); influence intent classification accuracy.
  • sampling temperature = 0
    Set to 0 for reproducibility (Appendix D.1).
assumptions (4)
  • ad hoc to paper The hand-crafted utterance intent list is exhaustive and mutually exclusive for open-domain dialogues.
    Stated as a design principle in Appendix A.5. The method depends on every utterance being mapped to one of the predefined intents.
  • ad hoc to paper A deterministic rule from intent labels to topic shift labels (INTRODUCE_TOPIC and CHANGE_TOPIC imply YES, others imply NO) is valid.
    Algorithm 1 and Section 3.5 enforce this rule without independent justification.
  • domain assumption A fixed context window of up to 2 preceding and 3 subsequent utterances is sufficient to determine topic shifts.
    Section 3.3 chooses this window to balance informativeness and token efficiency; no sensitivity analysis is provided.
  • domain assumption The LLM's intent classification is accurate enough to serve as the basis for deductive topic shift detection.
    The method's final output is a direct function of the intent label, so errors in intent classification propagate; Section 5.3 documents such errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Def-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation." pith.science (2026). https://pith.science/paper/YIEG72NG

@misc{pith2026250521033,
  author       = {Pith},
  title        = {Pith review of: Def-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YIEG72NG}},
  note         = {Machine review of arXiv:2505.21033}
}
read the original abstract

Dialogue Topic Segmentation (DTS) aims to divide dialogues into coherent segments. DTS plays a crucial role in various NLP downstream tasks, but suffers from chronic problems: data shortage, labeling ambiguity, and incremental complexity of recently proposed solutions. On the other hand, Despite advances in Large Language Models (LLMs) and reasoning strategies, these have rarely been applied to DTS. This paper introduces Def-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation, which utilizes LLM-based multi-step deductive reasoning to enhance DTS performance and enable case study using intermediate result. Our method employs a structured prompting approach for bidirectional context summarization, utterance intent classification, and deductive topic shift detection. In the intent classification process, we propose the generalizable intent list for domain-agnostic dialogue intent classification. Experiments in various dialogue settings demonstrate that Def-DTS consistently outperforms traditional and state-of-the-art approaches, with each subtask contributing to improved performance, particularly in reducing type 2 error. We also explore the potential for autolabeling, emphasizing the importance of LLM reasoning techniques in DTS.

Figures

Figures reproduced from arXiv: 2505.21033 by the authors.

Figure 1
Figure 1. An example of a topic shift in a conversation. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Prompt configuration and overall flow of our method: Def-DTS. (a) We utilize general intent list including [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) MATCHED INTENT indicates the accu￾racy of the other methodologies for grouped utterances by our intent classification process only in the true cases of our method. (b) MISMATCHED CASE indicates the co-error count of the other methodologies with our methods for only in the false cases of our method. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 22 canonical work pages

  1. [1]

    Bo Adler, Niket Agarwal, Ashwath Aithal, Dong H Anh, Pallab Bhattacharya, Annika Brundyn, Jared Casper, Bryan Catanzaro, Sharon Clay, Jonathan Cohen, et al. 2024. Nemotron-4 340b technical report. arXiv preprint arXiv:2406.11704

    arXiv 2024

  2. [2]

    Aleksei Artemiev, Daniil Parinov, Alexey Grishanov, Ivan Borisov, Alexey Vasilev, Daniil Muravetskii, Aleksey Rezvykh, Aleksei Goncharov, and Andrey Savchenko. 2024. https://doi.org/10.18653/v1/2024.findings-naacl.291 Leveraging summarization for unsupervised dialogue topic segmentation . In Findings of the Association for Computational Linguistics: NAACL...

    work page doi:10.18653/v1/2024.findings-naacl.291 2024

  3. [3]

    Doug Beeferman, Adam Berger, and John Lafferty. 1997. https://aclanthology.org/W97-0304 Text segmentation using exponential models . In Second Conference on Empirical Methods in Natural Language Processing

    work page 1997

  4. [4]

    Mohammad Hadi Bokaei, Hossein Sameti, and Yang Liu. 2016. https://doi.org/10.1017/S1351324914000199 Extractive summarization of multi-party meetings through discourse segmentation . Natural Language Engineering, 22(1):41–72

    work page doi:10.1017/s1351324914000199 2016

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    2020

  6. [6]

    Pawe Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, I \ n igo Casanueva, Ultes Stefan, Ramadan Osman, and Milica Ga s i\'c. 2018. Multiwoz - a large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2018

  7. [7]

    Jiaao Chen and Diyi Yang. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.336 Multi-view sequence-to-sequence models with conversational structure for abstractive dialogue summarization . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4106--4118, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.emnlp-main.336 2020

  8. [8]

    Yan Chen and Zhenghang Luo. 2023. https://doi.org/10.3390/s23052848 Pre-trained joint model for intent classification and slot filling with semantic feature fusion . Sensors, 23(5)

    work page doi:10.3390/s23052848 2023

Show all 49 references
  1. [9]

    Sarkar Snigdha Sarathi Das, Chirag Shah, Mengting Wan, Jennifer Neville, Longqi Yang, Reid Andersen, Georg Buscher, and Tara Safavi. 2024. https://doi.org/10.18653/v1/2024.findings-acl.891 S3- DST : Structured open-domain dialogue segmentation and state tracking in the era of ...

    2024 doi

  2. [10]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/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 Associat...

    2019 doi

  3. [11]

    Mihail Eric, Lakshmi Krishnan, Francois Charette, and Christopher D. Manning. 2017. https://doi.org/10.18653/v1/W17-5506 Key-value retrieval networks for task-oriented dialogue . In Proceedings of the 18th Annual SIG dial Meeting on Discourse and Dialogue , pages 37--49, Saarb...

    2017 doi

  4. [12]

    Song Feng, Siva Sankalp Patel, Hui Wan, and Sachindra Joshi. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.498 M ulti D oc2 D ial: Modeling dialogues grounded in multiple documents . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,...

    2021 doi

  5. [13]

    Song Feng, Hui Wan, Chulaka Gunasekara, Siva Patel, Sachindra Joshi, and Luis Lastras. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.652 doc2dial: A goal-oriented document-grounded dialogue dataset . In Proceedings of the 2020 Conference on Empirical Methods in Natural Lan...

    2020 doi

  6. [14]

    Hannah Fong and Ethel Ong. 2023. https://pcsc.dlsu.edu.ph/proceedings/main-conference/32.pdf Evaluating ChatGPT for joint intent detection and slot filling: Zero-shot vs. few-shot prompting . In Proceedings of the 23rd Philippine Computing Science Congress (PCSC)

    2023

  7. [15]

    McKeown, Eric Fosler-Lussier, and Hongyan Jing

    Michel Galley, Kathleen R. McKeown, Eric Fosler-Lussier, and Hongyan Jing. 2003. https://doi.org/10.3115/1075096.1075167 Discourse segmentation of multi-party conversation . In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pages 562--...

    2003

  8. [16]

    Haoyu Gao, Rui Wang, Ting-En Lin, Yuchuan Wu, Min Yang, Fei Huang, and Yongbin Li. 2023. https://doi.org/10.1145/3539618.3592081 Unsupervised dialogue topic segmentation with topic-aware contrastive learning . In Proceedings of the 46th International ACM SIGIR Conference on Re...

    2023

  9. [17]

    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, pages 6894--6910, Online ...

    2021 doi

  10. [18]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  11. [19]

    Wanwei He, Yinpei Dai, Yinhe Zheng, Yuchuan Wu, Zheng Cao, Dermot Liu, Peng Jiang, Min Yang, Fei Huang, Luo Si, Jian Sun, and Yongbin Li. 2022. https://doi.org/10.1609/aaai.v36i10.21320 Galaxy: A generative pre-trained model for task-oriented dialog with semi-supervised learni...

    2022 doi

  12. [20]

    Marti A. Hearst. 1997. https://aclanthology.org/J97-1003 Text tiling: Segmenting text into multi-paragraph subtopic passages . Computational Linguistics, 23(1):33--64

    1997

  13. [21]

    Xia Hou, Qifeng Li, and Tongliang Li. 2024. An unsupervised dialogue topic segmentation model based on utterance rewriting. arXiv preprint arXiv:2409.07672

    2024 arXiv

  14. [22]

    Shima Imani, Liang Du, and Harsh Shrivastava. 2023. https://doi.org/10.18653/v1/2023.acl-industry.4 M ath P rompter: Mathematical reasoning using large language models . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Indus...

    2023 doi

  15. [23]

    Junfeng Jiang, Chengzhang Dong, Sadao Kurohashi, and Akiko Aizawa. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.249 S uper D ialseg: A large-scale dataset for supervised dialogue segmentation . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language...

    2023 doi

  16. [24]

    Daniel Kahneman. 2011. Thinking, fast and slow. macmillan

    2011

  17. [25]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/6b493230205f78...

    2020

  18. [26]

    Jiwei Li, Will Monroe, Alan Ritter, Dan Jurafsky, Michel Galley, and Jianfeng Gao. 2016. https://doi.org/10.18653/v1/D16-1127 Deep reinforcement learning for dialogue generation . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages ...

    2016 doi

  19. [27]

    Ting-En Lin, Hua Xu, and Hanlei Zhang. 2020. Discovering new intents via constrained deep adaptive clustering with cluster refinement. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8360--8367

    2020

  20. [28]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    2024 arXiv

  21. [29]

    Bing Liu and Ian Lane. 2016. Attention-based recurrent neural network models for joint intent detection and slot filling. arXiv preprint arXiv:1609.01454

    2016 arXiv

  22. [30]

    Che Liu, Rui Wang, Junfeng Jiang, Yongbin Li, and Fei Huang. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.490 Dial2vec: Self-guided contrastive learning of unsupervised dialogue embeddings . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Pr...

    2022 doi

  23. [31]

    Yinhan Liu. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  24. [32]

    Xinbei Ma, Yi Xu, Hai Zhao, and Zhuosheng Zhang. 2024. https://doi.org/10.1016/j.neucom.2024.127385 Multi-turn dialogue comprehension from a topic-aware perspective . Neurocomputing, 578:127385

    2024

  25. [33]

    Passonneau and Diane J

    Rebecca J. Passonneau and Diane J. Litman. 1997. https://aclanthology.org/J97-1005/ Discourse segmentation by human and automated means . Computational Linguistics, 23(1):103--139

    1997

  26. [34]

    Lev Pevzner and Marti A. Hearst. 2002. https://doi.org/10.1162/089120102317341756 A critique and improvement of an evaluation metric for text segmentation . Computational Linguistics, 28(1):19--36

    2002 doi

  27. [35]

    MengNan Qi, Hao Liu, YuZhuo Fu, and Ting Liu. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.97 Improving abstractive dialogue summarization with hierarchical pretraining and topic segment . In Findings of the Association for Computational Linguistics: EMNLP 2021, pages...

    2021 doi

  28. [36]

    Yiping Song, Lili Mou, Rui Yan, Li Yi, Zinan Zhu, Xiaohua Hu, and Ming Zhang. 2016. Dialogue session segmentation by embedding-enhanced texttiling. Interspeech 2016, pages 2706--2710

    2016

  29. [37]

    Jinyuan Wang, Junlong Li, and Hai Zhao. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.179 Self-prompted chain-of-thought on large language models for open-domain multi-hop reasoning . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2717-...

    2023 doi

  30. [38]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf Chain-of-thought prompting elicits reasoning...

    2022

  31. [39]

    Huiyuan Xie, Zhenghao Liu, Chenyan Xiong, Zhiyuan Liu, and Ann Copestake. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.145 TIAGE : A benchmark for topic-shift aware dialog modeling . In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1684-...

    2021 doi

  32. [40]

    Linzi Xing and Giuseppe Carenini. 2021. https://doi.org/10.18653/v1/2021.sigdial-1.18 Improving unsupervised dialogue topic segmentation with utterance-pair coherence scoring . In Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue, p...

    2021 doi

  33. [41]

    Jun Xu, Zeyang Lei, Haifeng Wang, Zheng-Yu Niu, Hua Wu, and Wanxiang Che. 2021 a . https://doi.org/10.18653/v1/2021.acl-long.136 Discovering dialog structure graph for coherent dialog generation . In Proceedings of the 59th Annual Meeting of the Association for Computational L...

    2021 doi

  34. [42]

    Yi Xu, Hai Zhao, and Zhuosheng Zhang. 2021 b . Topic-aware multi-turn dialogue modeling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14176--14184

    2021

  35. [43]

    An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, Keming Lu, Mingfeng Xue, Runji Lin, Tianyu Liu, Xingzhang Ren, and Zhenru Zhang. 2024. Qwen2.5-math technical report: Toward mathematical expert model ...

    2024 arXiv

  36. [44]

    Seunghyun Yoon, Joongbo Shin, and Kyomin Jung. 2018. https://doi.org/10.18653/v1/N18-1142 Learning to rank question-answer pairs using hierarchical recurrent encoder with latent topic clustering . In Proceedings of the 2018 Conference of the North A merican Chapter of the Asso...

    2018 doi

  37. [45]

    Sai Zhang, Yuwei Hu, Yuchuan Wu, Jiaman Wu, Yongbin Li, Jian Sun, Caixia Yuan, and Xiaojie Wang. 2022. https://doi.org/10.18653/v1/2022.findings-acl.27 A slot is not built in one utterance: Spoken language dialogs with sub-slots . In Findings of the Association for Computation...

    2022 doi

  38. [46]

    Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. https://doi.org/10.18653/v1/P18-1205 Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational ...

    2018 doi

  39. [47]

    Ming Zhong, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. 2022. Dialoglm: Pre-trained model for long dialogue understanding and summarization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11765--11773

    2022

  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.