Intent Recognition and Out-of-Scope Detection using LLMs in Multi-party Conversations

Ga\"el de Chalendar; Galo Castillo-L\'opez; Nasredine Semmar

REVIEW 4 major objections 6 minor 18 references

Intent Recognition and Out-of-Scope Detection using LLMs in Multi-party Conversations

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Sharing BERT's probability estimates with an LLM, by shrinking the intent list in its prompt, improves intent recognition and out-of-scope detection on multi-party data while cutting LLM latency by about 40%.

desk verdict Clean, incremental hybrid-routing paper whose headline claim needs a same-size label-reduction control before it fully lands. read the letter →

arxiv 2507.22289 v1 pith:VNFAU53M submitted 2025-07-29 cs.CL cs.LG

Galo Castillo-López , Gaël de Chalendar , Nasredine Semmar This is my paper

classification cs.CLcs.LG

keywords intentrecognitionout-of-scopedetectionmulti-partyconversationlabelspacereductionBERTlargelanguagemodelsuncertainty-basedroutingfew-shotlearning

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

Task-oriented dialogue systems need to recognize user intents and flag out-of-scope (OOS) utterances, but classic fine-tuned classifiers require large amounts of labeled data, and running a large language model on every turn is slow. The paper proposes a hybrid: let a fine-tuned BERT model handle the easy turns, and route only uncertain ones to an LLM. Its distinctive move is to share BERT's softmax probabilities with the LLM by shrinking the intent list in the prompt to the smallest set whose cumulative probability reaches P=0.85. On two multi-party conversation datasets, MIntRec2.0 and MPGT, this label space reduction improves in-scope recognition and OOS detection compared with routing without shared information, and it lowers latency by roughly 40% relative to running the LLM alone. The paper's central claim is that sharing probability estimates between a cheap model and an LLM beats routing that keeps the two models' views separate.

What carries the argument

The central machinery is label space reduction (LSR). For each utterance $u_i$ routed to the LLM, the system orders the in-scope intents by BERT's softmax probabilities and forms $K_i = \{y_1,\dots,y_k\}$, the smallest subset satisfying $\sum_{j=1}^k p_i(y_j) \ge P$ with $P=0.85$. The prompt then asks the LLM to choose among $K_i$ or UNK, so the label list is cut by roughly 80% on MIntRec2.0 and 50% on MPGT while retaining the true intent in the subset about 90% of the time on validation. The same probability estimates also drive the uncertainty signal: the standard deviation across five fine-tuned BERT models decides which utterances bypass the LLM entirely. Thus a single cheap classifier supplies both the router and the label filter.

What would settle it

Run LSR at P=0.85 on the held-out test splits of both MIntRec2.0 and MPGT and report the fraction of routed utterances whose true intent actually lies in $K_i$, then re-run the MIntRec2.0 experiments with P swept as was done for MPGT; if the test hit rate is not above roughly 90% on both corpora and the performance curves do not follow the validation pattern, the claimed improvement is an artifact of threshold choice.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a small model's probability output can be used as a filter at inference time to make an LLM both more accurate and cheaper. After ten-shot fine-tuning of BERT on the in-scope intents, five runs give softmax probabilities $p_i(y_j)$ for each utterance, and the system routes a turn to the LLM when the standard deviation of those probabilities is high. For routed utterances, the prompt lists only $K_i$, the smallest subset of intents with cumulative probability at least $P=0.85$, instead of the full label set, while keeping UNK as a possible answer. The paper reports that this improves most in-scope metrics over BERT alone, LLM alone, and uncertainty routing without label reduction, and in particular materially raises in-scope weighted precision on MIntRec2.0 and in-scope plus out-of-scope accuracy on MPGT. It also reports latency reductions of about 40% when BERT is combined with Llama-3 70B or Mixtral 8x7B, because most turns are answered by BERT and the LLM sees a shorter list.

Load-bearing premise

The whole trick depends on BERT's ten-shot probabilities being good enough that the true intent is almost always inside the small label list; if BERT ranks the right intent poorly on new data, the LLM never sees it and every such example becomes an error.

Editorial extensions

If this is right

If the claim holds, production dialogue systems can get LLM-level generalization while paying LLM cost only on hard turns, because BERT handles the confident majority.
Prompt length directly controls latency, so label filtering makes LLM calls faster and cheaper at scale, not only in the laboratory setting.
The method requires only ten annotated examples per intent to fine-tune BERT, so it can be adopted quickly when new intents are added to a deployed system.
The same pipeline should transfer to dyadic, two-speaker task-oriented dialogues, since nothing in the mechanism depends on having more than two participants.
Because OOS utterances tend to produce high BERT uncertainty, routing them to an LLM with a reduced label list catches some OOS cases as UNK while keeping in-scope accuracy high.

Reading between the lines

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

A natural next step is to put the actual probability numbers into the prompt rather than only the filtered label names; the paper's own conclusion invites this, and it could help an LLM weigh borderline intents.
The method should work with any fast classifier in place of BERT, including sentence encoders, so the same information-sharing trick could be benchmarked in dyadic settings where those models are even cheaper.
Since P was tuned on the MPGT validation set and the validation hit rate is barely above 90%, test-set calibration of P per dataset is the most direct way to see whether the gains survive distribution shift.
One caution implied by the tables: on the larger, noisier MIntRec2.0 corpus, the reported F1-OOS with LSR is lower than for BERT alone, so the benefit is concentrated in in-scope precision and overall accuracy rather than in OOS recall.

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.

Desk Editor's Note

read the letter

Quick read for you: this is a small, clean paper from CEA on combining fine-tuned BERT with zero-shot LLMs for intent and OOS detection in multi-party dialogue. The genuinely new piece is label-space reduction (LSR): instead of routing uncertain BERT queries to an LLM with all intent labels, they sort BERT's softmax probabilities and include only the smallest intent subset whose cumulative probability reaches P=0.85. That is a real, if incremental, addition to Arora et al. (2024), and they evaluate it on two multi-party datasets rather than the usual dyadic ones.

What the paper does well: the setup is easy to follow, the datasets are public, they release their adapted MPGT multi-class version, and they report latency. The efficiency gain (roughly 40% lower latency versus running the LLM alone) is the hardest part of the claim and it is supported by their Table 3. I also think the discussion of MIntRec2.0's label overlap (Figure 2) is honest about why even LLMs struggle there.

The soft spots are real but not fatal. The main one is the one the stress-test flags: the Section 6 claim that sharing probability estimates 'outperforms methods without shared information' is not isolated from the benefit of a shorter label list. LSR always prompts with fewer intents, and there is no control that reduces the label set to the same size using random subsets or top-k by training frequency. Without that control, part of the gain could be prompt-size confound. The results themselves are also mixed: on MIntRec2.0, LSR drops Mixtral's F1-OOS (31.95 to 25.98), and Llama-3's OOS F1 is near zero in both settings. There are no error bars or significance tests, and P is tuned on validation with hit-rate behavior reported only there. The MPGT manual relabeling is a reasonable adaptation but it means the numbers are not directly comparable to the original corpus.

Overall, the idea is worth reporting and the paper is a legitimate engineering contribution. The conclusion just needs qualification, and a matched-size control would settle it. I would send it to review rather than desk-reject; it is clear, short, and the authors know the prior work. For my own work I would cite it if I were doing hybrid routing; it is a useful point of comparison.

Referee Report

4 major / 6 minor

Summary. The paper proposes Label Space Reduction (LSR), a hybrid intent-recognition and out-of-scope detection pipeline in which a ten-shot fine-tuned BERT first decides whether an utterance is uncertain, routes only uncertain utterances to an LLM, and also supplies a short list of likely intents (the smallest set whose cumulative softmax probability reaches P=0.85) to be used in the LLM prompt. The method is evaluated on two multi-party dialogue corpora, MIntRec2.0 and MPGT, with three open LLMs: Mixtral 8x7B, Llama-3 70B, and DeepSeek-R1 70B. The paper reports that LSR improves several in-scope and in-scope-plus-OOS metrics on MPGT, gives mixed results on MIntRec2.0, and reduces latency by about 40% relative to LLM-only inference. The central claim, stated in Section 6, is that sharing BERT probability estimates to reduce the label space outperforms methods without shared information.

Significance. If the central claim were fully supported, the paper would make a useful practical contribution: combining a cheap fine-tuned model's probability estimates with an LLM's zero-shot generalization is a natural and deployable design, and the multi-party evaluation extends prior dyadic work. The paper has concrete strengths: it reports the prompt template, gives hyperparameters for the BERT fine-tuning, uses five seeds for the small model, and releases the adapted multi-class MPGT corpus. However, the headline claim is only partially supported by the reported numbers, and the experimental design lacks a key control that would isolate the effect of sharing BERT probability estimates from the effect of shortening the label list. The mixed and inconsistent gains across models and metrics, together with the absence of variance or significance measures, mean that the paper's strongest conclusion should be substantially softened or re-evidenced.

major comments (4)

[Section 6, Tables 1 and 2] The central claim that 'sharing information among models such as probability estimates to reduce the label space outperforms methods without shared information' is not established by the reported comparisons. LSR is compared only against the full-label-set routing baseline of Section 3.2.3; no control reduces the label set to the same size without using BERT's probability ranking, such as random subsets, training-frequency top-k subsets, or shortest-label-list prompts. This control is essential because the gains are inconsistent: on MIntRec2.0, LSR lowers F1-OOS for Mixtral from 31.95 to 25.98 and for Llama-3 from 8.67 to 7.36, and lowers weighted precision for Mixtral from 47.83 to 47.03; on MPGT, Llama-3 in-scope accuracy drops from 90.23 to 89.47. Without a matched-size control, the observed improvements could be due to prompt-size reduction or noise rather than to the information content of BERT's ranking.
[Section 3.2.4 and Appendix C] The key assumption that BERT's softmax ranking places the true intent in the reduced label set is validated only on the validation splits (hit rates slightly above 90% for P=0.85), and the sensitivity analysis of P reported in Appendix C is run only on MPGT with BERT+DeepSeek. The main results in Tables 1 and 2 use P=0.85 on test splits without reporting the test-set hit rate or the distribution of subset sizes. Because a missing true intent converts an in-scope utterance into an UNK response, the paper should report the test-set hit-rate diagnostic and, if needed, select or calibrate P with that evidence rather than relying solely on validation behavior.
[Section 5, Table 3] The computational-efficiency analysis conflates uncertainty routing with label-space reduction. Table 3 compares BERT+LLM with LSR against the LLM-only baseline, but the bulk of the latency reduction comes from dispatching only uncertain queries to the LLM; the authors do not report the latency of BERT+LLM uncertainty routing without LSR. Consequently, the claimed approximately 40% reduction in computational cost cannot be attributed to the proposed label-space reduction component. In addition, the latency-ratio column is defined only in the caption and is confusing as printed: the BERTten rows are repeated with different ratios (0.034 and 0.016) depending on the paired LLM, and the ratio for the LSR rows appears to be relative to the corresponding LLM-only row rather than to the routing baseline.
[Section 4, Tables 1 and 2] No variance or significance measures are reported. The BERT models are fine-tuned over five seeds, but the tables present single point estimates without standard deviations, confidence intervals, or paired significance tests. Many of the LSR-versus-baseline differences are small in absolute terms (for example, roughly 0.2 to 1.5 accuracy points on MPGT), so it is unclear whether the reported improvements are stable across the five BERT runs or within the test-split variation. At minimum, standard errors over the five runs, or a paired test over the test utterances, should be provided for the headline comparisons.

minor comments (6)

[Section 3.3] The metric listed as 'F1-OSS' appears to be a typo for 'F1-OOS'; please fix this for consistency with the rest of the paper.
[Table 3] Consider restructuring Table 3 to include a BERT+LLM routing baseline without LSR, and provide a clearer column definition for 'latency ratio' so that the reader can see exactly which denominator is used for each row.
[Appendix C, Figure 5] The prose describes the effect of P on OOS precision and recall, but the figure itself is not accompanied by a caption that defines the plotted metrics; please add a self-contained caption and, ideally, include the test-set results for both datasets.
[Section 4] The sentence reporting an 'increase of approximately 3% on the weighted precision' refers to an absolute increase from 49.58 to 52.65; using 'percentage points' would avoid ambiguity.
[Figure 3] The prompt template lists intents as 'intent_1, intent_2, ... intent_N'; since LSR dynamically changes the intent list per query, the authors should state explicitly (or show with an example) how the reduced subset is inserted into the template.
[Abstract and Section 3.2] The abstract says 'zero and few-shot settings,' but the LLM component is zero-shot while BERT is fine-tuned with ten shots; clarifying this wording would make the contribution more precise.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the LSR pipeline is evaluated on held-out test labels; BERT probability estimates are model inputs, not the predicted outputs.

full rationale

The paper's central claim is an empirical comparison: BERT+LLM routing with label-space reduction (LSR) is compared against routing without shared information on held-out test splits of MIntRec2.0 and MPGT (Tables 1 and 2), with metrics computed against ground-truth labels. Hyperparameters P and sigma are tuned on validation sets (Sections 3.2.1 and 3.2.4, Appendix C), and the test splits are not used for this tuning, so the headline comparison is not statistically forced. The BERT softmax probabilities enter prompt construction as inputs; the LLM's final intent/OOS label is the output, and there is no equation under which the output is definitionally equal to the input. The only self-citation (Castillo-López et al., 2025) supports the motivational claim that multi-party dialogue is understudied; it is not load-bearing for any experimental result. The absence of a same-size label-set control (e.g., random or frequency-based k-intent subsets) is a possible attribution confound for the 'shared information' claim, but it is a methodological validity concern, not a circularity in the derivation chain.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method relies on BERT's softmax probabilities for both routing and label ranking, and on a manual conversion of the MPGT corpus from multi-label to single-label. The free parameters are validation-tuned thresholds and a context window choice; no invented entities are introduced.

free parameters (3)

P (label space reduction threshold) = 0.85
Cumulative softmax probability mass defining the smallest label subset sent to the LLM; chosen on validation sets; main results use P=0.85.
sigma (uncertainty routing threshold) = 0.10 (MIntRec2.0), 0.12 (MPGT)
Standard deviation of softmax probabilities across 5 BERT runs above which queries are routed to the LLM; tuned on validation sets.
Context window length = 3 preceding utterances
Each utterance is concatenated with its 3 prior utterances and <ts> tokens; no ablation is reported for this choice.

assumptions (3)

domain assumption Each utterance is assumed to express exactly one intent, either in-scope or out-of-scope.
Adopted in Appendix B.1, where MPGT's multi-label annotations are manually converted to multi-class labels; this shapes both training and evaluation.
domain assumption The standard deviation of softmax probabilities across 5 BERT seeds is a valid uncertainty signal for routing and label ranking.
Used in Sections 3.2.1 and 3.2.3 to decide which utterances go to the LLM and which labels to keep in the prompt; no calibration analysis is provided.
domain assumption LLMs follow the prompt and return one of the listed labels or UNK reliably in zero-shot settings.
Section 3.2.2 relies on instructed LLMs for classification; no parsing or validity checks are reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Intent Recognition and Out-of-Scope Detection using LLMs in Multi-party Conversations." pith.science (2026). https://pith.science/paper/VNFAU53M

@misc{pith2026250722289,
  author       = {Pith},
  title        = {Pith review of: Intent Recognition and Out-of-Scope Detection using LLMs in Multi-party Conversations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VNFAU53M}},
  note         = {Machine review of arXiv:2507.22289}
}

read the original abstract

Intent recognition is a fundamental component in task-oriented dialogue systems (TODS). Determining user intents and detecting whether an intent is Out-of-Scope (OOS) is crucial for TODS to provide reliable responses. However, traditional TODS require large amount of annotated data. In this work we propose a hybrid approach to combine BERT and LLMs in zero and few-shot settings to recognize intents and detect OOS utterances. Our approach leverages LLMs generalization power and BERT's computational efficiency in such scenarios. We evaluate our method on multi-party conversation corpora and observe that sharing information from BERT outputs to LLMs leads to system performance improvement.

Figures

Figures reproduced from arXiv: 2507.22289 by the authors.

**Figure 1.** Overview of our Label Space Reduction method. et al., 2024). These approaches consist in processing queries in two steps: first through the computationally efficient model, and then through LLMs, if necessary. In doing so, overall computational costs are reduced without compromising prediction quality. However, these methods do not share information among models, and hence LLMs miss potential relevant information… view at source ↗

**Figure 2.** Example of an instance difficult to classify by [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Prompt template used on all LLM experiments. Highlighted text in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Distribution of the intents in the adapted multi [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Performance metrics at distinct values of [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 13 canonical work pages

[4]

In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17554–17567

Dual-oriented disen- tangled network with counterfactual intervention for multimodal intent detection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17554–17567. Tim R Davidson, Luca Falorsi, Nicola De Cao, Thomas Kipf, and Jakub M Tomczak

work page 2024
[8]

In Proceedings of the 5th Workshop on NLP for Conver- sational AI (NLP4ConvAI 2023), pages 140–154

A survey of challenges and methods in the computational modeling of multi-party dialog. In Proceedings of the 5th Workshop on NLP for Conver- sational AI (NLP4ConvAI 2023), pages 140–154. Chandan Gautam, Sethupathy Parameswaran, Aditya Kane, Yuan Fang, Savitha Ramasamy, Suresh Sun- daram, Sunil Kumar Sahu, and Xiaoli Li

work page 2023
[9]

In Findings of the Association for Computational Lin- guistics: EMNLP 2024 , pages 9100–9112, Miami, Florida, USA

Class name guided out-of-scope intent classification. In Findings of the Association for Computational Lin- guistics: EMNLP 2024 , pages 9100–9112, Miami, Florida, USA. Association for Computational Lin- guistics. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelte...

work page 2024
[10]

arXiv preprint arXiv:2407.21783

The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bam- ford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, and 1 oth- ers

arXiv
[11]

arXiv preprint arXiv:2401.04088

Mixtral of experts. arXiv preprint arXiv:2401.04088. Stefan Larson, Anish Mahendran, Joseph J. Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K. Kummerfeld, Kevin Leach, Michael A. Laurenzano, Lingjia Tang, and Jason Mars

arXiv
[12]

An evaluation dataset for intent classification and out-of- scope prediction. In Proceedings of the 2019 Confer- ence on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1311–1316, Hong Kong, China. Association for Computational Linguistics. I-Fan Lin, Faeghe...

work page 2019
[13]

In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13138–13146, Miami, Florida, USA

Generate then refine: Data augmentation for zero- shot intent detection. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13138–13146, Miami, Florida, USA. Association for Computational Linguistics. Joongbo Shin, Youbin Ahn, Seungpil Won, and Stan- ley Jungkyu Choi

work page 2024
[15]

In Findings of the Association for Computational Lin- guistics: EMNLP 2023, pages 3926–3939, Singapore

APP: Adaptive prototypi- cal pseudo-labeling for few-shot OOD detection. In Findings of the Association for Computational Lin- guistics: EMNLP 2023, pages 3926–3939, Singapore. Association for Computational Linguistics. Pei Wang, Keqing He, Yejie Wang, Xiaoshuai Song, Yutao Mou, Jingang Wang, Yunsen Xian, Xunliang Cai, and Weiran Xu

work page 2023

Show all 18 references

[16]

In Proceedings of the 2024 Joint In- ternational Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING

Beyond the known: In- vestigating LLMs performance on out-of-domain in- tent detection. In Proceedings of the 2024 Joint In- ternational Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING

2024
[17]

In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38–45

Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38–45. Hanlei Zhang, Xin Wang, Hua Xu, Qianrui Zhou, Kai Gao, Jianhua Su, Wenrui Li, Yanting Chen...

2020 arXiv
[18]

i n t e n t

for sequence classification tasks. BERTBASE uncased is used in all the experiments. Table 4 shows the hyperparameter configuration we employ. hyperparameter value eval_monitor macro F1-score train_batch_size 16 eval_batch_size 16 test_batch_size 16 wait_patience 3 num_train_ep...

2024
[2018]

arXiv preprint arXiv:1804.00891

Hyperspher- ical variational auto-encoders. arXiv preprint arXiv:1804.00891. DeepSeek-AI

arXiv
[2019]

Bert: Pre-training of deep bidirectional transformers for language understand- ing. In Proceedings of the 2019 conference of the North American chapter of the association for com- putational linguistics: human language technologies, volume 1 (long and short papers), pages 4171...

2019
[2020]

arXiv preprint arXiv:2003.04807

Efficient intent detection with dual sentence encoders. arXiv preprint arXiv:2003.04807. Galo Castillo-López, Gaël de Chalendar, and Nasre- dine Semmar

2003 arXiv
[2022]

In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages 188–197, Abu Dhabi, UAE

Multi-tenant optimization for few-shot task-oriented FAQ retrieval. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages 188–197, Abu Dhabi, UAE. Association for Computational Linguistics. Pei Wang, Keqing He, Yutao ...

2022
[2023]

In Proceedings of the 2023 8th international conference on machine learning technologies, pages 151–156

Chat- bot recommender systems in tourism: A systematic review and a benefit-cost analysis. In Proceedings of the 2023 8th international conference on machine learning technologies, pages 151–156. Iñigo Casanueva, Tadas Tem ˇcinas, Daniela Gerz, Matthew Henderson, and Ivan Vuli´c

2023
[2024]

In Proceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages 1559–1570, Miami, Florida, US

Intent detection in the age of LLMs. In Proceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages 1559–1570, Miami, Florida, US. Association for Computational Linguistics. Mark Anthony Camilleri and Ciro Troise

2024
[2025]

Preprint, arXiv:2501.12948

Deepseek-r1: Incentivizing rea- soning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [4]

In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17554–17567

Dual-oriented disen- tangled network with counterfactual intervention for multimodal intent detection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17554–17567. Tim R Davidson, Luca Falorsi, Nicola De Cao, Thomas Kipf, and Jakub M Tomczak

work page 2024

[2] [8]

In Proceedings of the 5th Workshop on NLP for Conver- sational AI (NLP4ConvAI 2023), pages 140–154

A survey of challenges and methods in the computational modeling of multi-party dialog. In Proceedings of the 5th Workshop on NLP for Conver- sational AI (NLP4ConvAI 2023), pages 140–154. Chandan Gautam, Sethupathy Parameswaran, Aditya Kane, Yuan Fang, Savitha Ramasamy, Suresh Sun- daram, Sunil Kumar Sahu, and Xiaoli Li

work page 2023

[3] [9]

In Findings of the Association for Computational Lin- guistics: EMNLP 2024 , pages 9100–9112, Miami, Florida, USA

Class name guided out-of-scope intent classification. In Findings of the Association for Computational Lin- guistics: EMNLP 2024 , pages 9100–9112, Miami, Florida, USA. Association for Computational Lin- guistics. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelte...

work page 2024

[4] [10]

arXiv preprint arXiv:2407.21783

The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bam- ford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, and 1 oth- ers

arXiv

[5] [11]

arXiv preprint arXiv:2401.04088

Mixtral of experts. arXiv preprint arXiv:2401.04088. Stefan Larson, Anish Mahendran, Joseph J. Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K. Kummerfeld, Kevin Leach, Michael A. Laurenzano, Lingjia Tang, and Jason Mars

arXiv

[6] [12]

An evaluation dataset for intent classification and out-of- scope prediction. In Proceedings of the 2019 Confer- ence on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1311–1316, Hong Kong, China. Association for Computational Linguistics. I-Fan Lin, Faeghe...

work page 2019

[7] [13]

In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13138–13146, Miami, Florida, USA

Generate then refine: Data augmentation for zero- shot intent detection. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13138–13146, Miami, Florida, USA. Association for Computational Linguistics. Joongbo Shin, Youbin Ahn, Seungpil Won, and Stan- ley Jungkyu Choi

work page 2024

[8] [15]

In Findings of the Association for Computational Lin- guistics: EMNLP 2023, pages 3926–3939, Singapore

APP: Adaptive prototypi- cal pseudo-labeling for few-shot OOD detection. In Findings of the Association for Computational Lin- guistics: EMNLP 2023, pages 3926–3939, Singapore. Association for Computational Linguistics. Pei Wang, Keqing He, Yejie Wang, Xiaoshuai Song, Yutao Mou, Jingang Wang, Yunsen Xian, Xunliang Cai, and Weiran Xu

work page 2023