LR2GS: New Intent Discovery from Local Relations to Global Structure¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/hustchenpeng/LR2GS
Area: LLM Reasoning (New Intent Discovery / Semantic Clustering / Semi-Supervised Learning)
Keywords: New Intent Discovery, Intent Representation Space, Local Relations, Global Structure, LLM-based Refinement, Contrastive Learning
TL;DR¶
Addressing the challenge in New Intent Discovery (NID) where absence of explicit supervision hinders the reliable formation of well-separated global intent clusters, LR2GS starts from the foundational building blocks—local relations. By integrating inter-query semantic neighbor alignment, intra-query descriptive/prototype variant consistency, and cluster-boundary ambiguity calibration, LR2GS progressively shapes an optimal global intent representation space with superior intra-cluster compactness and inter-cluster separability.
Background & Motivation¶
Traditional intent detection systems achieve remarkable classification accuracy in closed-world setups where intent categories are pre-annotated. However, real-world user queries are inherently dynamic, with novel, uncataloged intent categories continually emerging over time. The New Intent Discovery (NID) task aims to automatically uncover these latent, previously unseen intent classes from unlabeled query streams. Fundamentally, this requires training an intent representation encoder \(f_\theta\) that maps discrete text inputs into a continuous metric space \(\mathcal{R}\) exhibiting ideal topological clustering properties: queries sharing the same underlying intent must tightly coalesce into compact intra-cluster manifolds, while decision boundaries between distinct intent clusters remain sharply delineated.
Nevertheless, establishing an ideal global clustering topology in an unsupervised or label-scarce regime presents profound challenges. Naive baselines applying K-Means over frozen pretrained representations (such as vanilla BERT) falter because the unguided embedding space remains dominated by lexical overlap, syntax patterns, or surface entities rather than latent intent semantics. Iterative pseudo-labeling and self-training approaches frequently suffer from error accumulation and runaway confirmation bias near fuzzy cluster peripheries. Meanwhile, recent methods deploying large language models (LLMs) to synthesize proxy supervision often disregard fine-grained local relational structures between authentic user queries, leading to severe representation drift and overlapping clusters.
The central thesis of this paper is that an optimal global intent structure naturally emerges from the continuous accumulation of high-precision local relations. Instead of tackling noisy, ill-posed top-down global cluster estimation directly, model optimization should ground itself on verified, micro-level pairwise constraints—steadily pulling true semantic neighbors together, anchoring multi-view semantic variants of the same query, and actively pulling boundary-straddling instances toward verified cluster centers. Based on this perspective, the authors propose LR2GS (From Local Relations to Global Structure). Core idea: deconstruct global intent clustering into the accumulation of accessible local relational supervisions, combining inter-query semantic neighbor alignment, intra-query multi-view variant consistency, and boundary ambiguity calibration to progressively shape a compact and well-separated global intent representation space.
Method¶
Overall Architecture¶
LR2GS coordinates three interconnected modules to progressively cultivate the intent representation space: (1) Local Relation from Semantic Neighbors, which combines dense retriever retrieval with LLM-based consistency arbitration to inject reliable cross-query neighborhood alignment; (2) Local Relation from Query Variants, which prompts an LLM to generate descriptive and prototype intent variants to enforce intra-instance invariance against surface phrasing noise; and (3) Global Structure Disambiguation, which dynamically identifies boundary-straddling ambiguous queries and leverages cross-cluster comparisons to pull them toward calibrated positive exemplars.
flowchart TD
A["Raw Unlabeled User Query Set D"] --> B["Inter-Query Neighbor Mining & LLM Refinement (L_na)"]
A --> C["Intra-Query Multi-View Generation: Descriptive & Prototype Views (L_va)"]
B --> D["Initial Intent Representation Encoder f_θ"]
C --> D
D --> E["K-Means Cluster Boundary Ambiguity Detection (ADC Module)"]
E --> F["LLM-Guided Exemplar Calibration (L_ac)"]
F --> G["Final Highly Compact & Separable Global Intent Structure"]
Key Designs¶
1. Inter-Query Semantic Neighbor Mining with LLM Refinement: High-Fidelity Relational Scaffolding
In fully unsupervised regimes, nearest neighbors retrieved purely via cosine distance over dense text representations frequently contain hard false positives sharing superficial lexical structure but differing in operational intent. LR2GS decouples neighbor discovery into a dual-stage framework: a dense semantic retriever \(T\) (all-mpnet-base-v2) first identifies the Top-\(K_{\text{nn}}\) most semantically similar query candidates for query \(x_i\); subsequently, an LLM-based Neighbor Refinement (LNR) module analyzes the candidate pool via structured zero-shot prompts to filter out distracting negatives and select the true semantic positive \(x_i^+\). The intent encoder representations \(z_i = f_\theta(x_i)\) and \(z_i^p = f_\theta(x_i^+)\) are optimized via a normalized neighbor alignment objective:
$\(\mathcal{L}_{\text{na}} = -\frac{1}{N} \sum_{i=1}^N \log \frac{\exp(s_i / \tau)}{\sum_{j=1}^N \exp(s_j / \tau)}\)$
where \(s_i = \text{sim}(z_i, z_i^p)\). This steadily increases mutual similarity across authentic synonym pairs, forging continuous local connectivity throughout the latent manifold.
2. Intra-Query Multi-View Variant Alignment: Decoupling Core Intent from Surface Perturbations
Real-world queries exhibit extreme variance in form: some are overly laconic while others contain colloquial noise. To prevent the encoder from overfitting to surface lexical idiosyncrasies, LR2GS introduces multi-view instance-level augmentation. An LLM generates two complementary representations for each query \(x_i\): a concise Descriptive View \(x_i^{\text{des}}\) articulating the operational intent, and an abstracted Prototype View \(x_i^{\text{pro}}\) providing a categorical template label. The encoder projects both variants into representation space (\(z_i^{\text{des}}, z_i^{\text{pro}}\)) and minimizes bidirectional contrastive alignment distances:
$\(\mathcal{L}_{\text{va}} = \mathcal{L}_{\text{des}} + \mathcal{L}_{\text{pro}}\)$
This objective pulls raw instance embeddings toward their abstract intent counterparts, filtering linguistic noise and enforcing intra-instance semantic concentration.
3. Global Structure Disambiguation with Active Exemplar Calibration: Eliminating Boundary Noise
As the representation space organizes into preliminary clusters, queries situated near decision boundaries experience conflicting geometric attraction from competing cluster centroids, threatening to propagate confirmation errors. LR2GS introduces the Adaptive Disambiguation and Calibration (ADC) mechanism. It identifies ambiguous queries based on margin deficits between their distances to the closest and second-closest cluster centers. For each ambiguous sample, an LLM arbitrates across representative high-confidence exemplars from neighboring candidate clusters to identify the true calibrated positive \(x_i^{\text{cal}}\). An ambiguity correction objective \(\mathcal{L}_{\text{ac}}\) actively anchors these unstable samples to their ground-truth cluster core, resolving boundary overlaps.
Loss & Training¶
The total training objective unifies these three complementary local geometric constraints (augmented by cross-entropy classification \(\mathcal{L}_{\text{ce}}\) on labeled samples in semi-supervised settings): $\(\mathcal{L} = \mathcal{L}_{\text{na}} + \mathcal{L}_{\text{va}} + \mathcal{L}_{\text{ac}} \quad (+ \mathcal{L}_{\text{ce}})\)$ The backbone encoder is initialized with BERT-base-uncased. Alignment temperature \(\tau\) is set to 0.05. Optimization is conducted using AdamW with a batch size of 128 and an initial learning rate of \(2 \times 10^{-5}\). Training commences with \(\mathcal{L}_{\text{na}}\) and \(\mathcal{L}_{\text{va}}\) to establish coherent local submanifolds, while the ADC boundary disambiguation module activates mid-training (e.g., epoch 5) to dynamically inject \(\mathcal{L}_{\text{ac}}\), ensuring smooth and stable representation learning.
Key Experimental Results¶
Main Results¶
Evaluated on three standard NID benchmarks (BANKING, CLINC, StackOverflow) reporting clustering Accuracy (ACC), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI) under unsupervised (Known Intent Ratio KIR = 0%) and semi-supervised (KIR = 25%) configurations across 3 independent runs:
| Benchmark & Method | BANKING ACC | BANKING ARI | BANKING NMI | CLINC ACC | CLINC ARI | CLINC NMI | StackOverflow ACC | StackOverflow ARI | StackOverflow NMI |
|---|---|---|---|---|---|---|---|---|---|
| Unsupervised Setting (KIR = 0%) | |||||||||
| Raw BERT (Devlin et al., 2019) | 16.79 | 5.77 | 36.75 | 27.11 | 10.33 | 58.24 | 15.20 | 2.02 | 12.36 |
| Dense Retriever (all-mpnet-base) | 63.02 | 51.79 | 79.95 | 72.76 | 64.14 | 88.88 | 73.70 | 59.00 | 72.87 |
| MTP-CLNN (ACL 2022) | 65.90 | 55.75 | 81.80 | - | - | - | 81.43 | 67.63 | 78.71 |
| IDAS (NLP4ConvAI 2023) | 67.43 | 57.56 | 82.84 | 85.48 | 79.02 | 93.82 | 83.82 | 72.20 | 81.26 |
| INTENDD (EMNLP 2023) | 58.74 | 47.88 | 75.91 | 63.87 | 51.76 | 83.12 | 79.32 | 62.49 | 73.88 |
| SynCID (ACL 2024 SOTA) | 72.89 | 62.42 | 84.20 | 86.80 | 80.85 | 94.23 | 86.90 | 74.42 | 81.95 |
| SPILL (ACL 2025) | 71.48 | - | 85.12 | 84.43 | - | 93.63 | - | - | - |
| LR2GS (Ours) | 76.69 | 67.48 | 86.84 | 89.07 | 84.35 | 95.38 | 87.30 | 75.51 | 82.08 |
| (Relative Gain over Previous Best) | (+3.80) | (+5.06) | (+1.72) | (+2.27) | (+3.50) | (+1.03) | (+0.40) | (+1.09) | (+0.13) |
| Semi-Supervised Setting (KIR = 25%) | |||||||||
| BERT-CE Baseline | 33.51 | 23.35 | 62.45 | - | - | - | 29.70 | 14.22 | 31.17 |
| DeepAligned (AAAI 2021) | 49.08 | 37.62 | 70.50 | 74.07 | 64.63 | 88.97 | 54.50 | 37.96 | 50.86 |
| CsePL (EMNLP 2023) | 71.06 | 60.36 | 83.32 | 86.16 | 79.65 | 94.07 | 79.47 | 64.92 | 74.88 |
| GeoID (ACL 2024) | 68.75 | 57.92 | 82.54 | 88.13 | 81.24 | 94.37 | 81.80 | 66.35 | 77.62 |
| SynCID (ACL 2024 SOTA) | 75.41 | 65.40 | 85.39 | 87.85 | 82.39 | 94.85 | 87.86 | 76.11 | 82.46 |
| ALUP (NAACL 2024) | 74.61 | 62.64 | 84.06 | 88.40 | 82.44 | 94.84 | 82.20 | 64.54 | 76.58 |
| LR2GS (Ours) | 78.64 | 68.53 | 86.83 | 92.36 | 89.23 | 96.95 | 88.70 | 77.80 | 83.52 |
| (Relative Gain over Previous Best) | (+3.23) | (+3.13) | (+1.32) | (+3.96) | (+6.79) | (+2.10) | (+0.84) | (+1.69) | (+1.06) |
Ablation Study¶
Ablations verifying the incremental gains of component stacking (Table 2) and the severity of removing single components under semi-supervision (Table 3):
| Variant Configuration | BANKING ACC | BANKING ARI | BANKING NMI | CLINC ACC | CLINC ARI | CLINC NMI | StackOverflow ACC | StackOverflow ARI | StackOverflow NMI | Analysis |
|---|---|---|---|---|---|---|---|---|---|---|
| Unsupervised Incremental (Table 2) | ||||||||||
| Raw BERT Baseline | 16.79 | 5.77 | 36.75 | 27.11 | 10.33 | 58.24 | 15.20 | 2.02 | 12.36 | Unsupervised cold start lacks cluster topology |
| + Variant Alignment (\(\mathcal{L}_{\text{va}}\)) | 64.22 | 52.15 | 78.89 | 72.53 | 63.63 | 88.14 | 81.90 | 65.19 | 75.86 | Instance multi-view alignment yields massive leap |
| + Neighbor Alignment (\(\mathcal{L}_{\text{na}}\)) | 70.32 | 61.21 | 84.45 | 85.87 | 80.58 | 94.53 | 82.63 | 69.90 | 80.89 | Inter-query neighbors connect local manifolds |
| + Disambiguation (\(\mathcal{L}_{\text{ac}}\), Full) | 76.69 | 67.48 | 86.84 | 89.07 | 84.35 | 95.38 | 87.30 | 75.51 | 82.08 | Resolves overlapping margins for clean clustering |
| Semi-Supervised Ablation (Table 3) | ||||||||||
| LR2GS Full Framework | 78.64 | 68.53 | 86.83 | 92.36 | 89.23 | 96.95 | 88.70 | 77.80 | 83.52 | Unified local-to-global relational pipeline |
| w/o Neighbor Alignment (\(\mathcal{L}_{\text{na}}\)) | 72.37 | 63.45 | 84.69 | 90.04 | 85.43 | 95.63 | 85.90 | 73.34 | 80.44 | Loses cross-instance topological guidance |
| w/o Variant Alignment (\(\mathcal{L}_{\text{va}}\)) | 76.17 | 66.80 | 87.06 | 88.53 | 85.91 | 96.23 | 87.90 | 76.69 | 82.76 | Surface lexical noise degrades robust clustering |
| w/o Disambiguation (\(\mathcal{L}_{\text{ac}}\)) | 72.18 | 61.57 | 84.30 | 88.98 | 85.20 | 95.92 | 85.90 | 73.16 | 80.44 | Boundary queries suffer from conflicting centroids |
Key Findings¶
- Relation Consistency Accuracy (RCA) Predicts Clustering Performance: Quantifying pairwise relation purity using Relation Consistency Accuracy shows that naive BERT neighbors achieve poor consistency. Dense retrieval elevates RCA above 70%, and LLM refinement drives purity to 80%–92%, leading to proportional, linear leaps in downstream clustering metrics (ACC/ARI).
- Outperforming the Base Retriever Through Synergistic Reorganization: The dense retriever (all-mpnet) providing initial neighbor signals scores only 63.02% (BANKING) and 72.76% (CLINC) ACC when evaluated directly with K-Means. The fine-tuned LR2GS encoder achieves 76.69% and 89.07%, confirming that integrating multi-view local relations actively purifies and transcends the retriever's original feature space.
- Superior Handling of Highly Granular Intent Spaces: On BANKING (spanning 77 fine-grained, semantically overlapping financial intents), LR2GS achieves an unsupervised gain of +3.80% ACC and +5.06% ARI over the previous SOTA, demonstrating that fine-grained tasks depend decisively on localized relational constraints rather than coarse global partitioning.
Highlights & Insights¶
- Resolving Global Structure Discovery via Tractable Micro-Constraints: Rather than risking catastrophic drift by hallucinating global cluster centers directly, LR2GS demonstrates that accumulating reliable local relations naturally self-organizes the global manifold into a well-separated topology.
- Retriever-LLM Functional Complementarity: The dense retriever provides low-cost, high-recall candidate extraction, while the LLM conducts high-precision intent arbitration, maximizing supervision purity while minimizing inference overhead.
- Active Boundary Calibration Prevents Error Cascades: By intervening specifically on boundary-straddling samples at mid-training, LR2GS eliminates the self-reinforcing misallocations common in pseudo-labeling pipelines.
Limitations & Future Work¶
- Offline LLM Generation Overhead: Constructing multi-view variants and refining neighbor candidates requires pre-training LLM passes, incurring noticeable API latency and compute costs during large-scale industrial onboarding.
- Extreme Brevity and Context Deprivation: For extremely truncated or ambiguous queries consisting of single ungrounded acronyms, LLMs risk hallucinating prototype variants, occasionally introducing mild semantic drift into the graph.
Related Work & Insights¶
- vs. SynCID / GeoID: SynCID synthesizes pseudo-intent query distributions using LLMs, leaving it vulnerable to synthetic-to-real distribution shifts. LR2GS grounds itself strictly on authentic user queries, extracting self-supervised signals from their internal relations.
- vs. MTP-CLNN / DeepAligned: Prior self-supervised NID systems iteratively pseudo-label neighbors from the evolving encoder, often stagnating in confirmation bias loops. LR2GS leverages external retrieval models and LLMs to break out of representation lock-in.
Rating¶
- Novelty: ⭐⭐⭐⭐ Re-conceptualizes NID from a local-to-global perspective with elegant multi-view and boundary calibration mechanisms.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous unsupervised and semi-supervised benchmarks, detailed incremental/ablation analysis, RCA metrics, and t-SNE topological visualizations.
- Writing Quality: ⭐⭐⭐⭐⭐ Clear mathematical formulations, logically coherent transitions, and clean illustrative diagrams.
- Value: ⭐⭐⭐⭐ Highly applicable to dialogue intent mining, cold-start query taxonomy expansion, and unsupervised representation learning.