Skip to content

Preference-Aligned Data Synthesis for Fine-Grained Intent Detection

Conference: EMNLP 2026
Paper: ACL Anthology
Area: Dialogue Systems
Keywords: Fine-grained Intent Detection, Data Synthesis, Multi-Centroid Prototypes, Direct Preference Optimization, Iterative Preference Refresh

TL;DR

Addressing inter-class ambiguity and intra-class homogeneity in LLM-synthesized data for fine-grained intent detection, this paper introduces a parameter-level generator alignment framework that mines multi-centroid prototype-aligned diverse positives and confusability-weighted cross-intent hard negatives for Direct Preference Optimization (DPO), delivering significant downstream gains in zero-shot and few-shot regimes while improving generator reliability.

Background & Motivation

Intent detection serves as an indispensable perceptual foundation for task-oriented dialogue systems and autonomous LLM agents, underpinning critical capabilities such as conversational planning, action dispatching, and tool execution. In real-world enterprise deployments, conversational label taxonomies constantly expand and become increasingly fine-grained. The semantic boundaries separating neighboring intents become exceedingly razor-thin (e.g., distinguishing "order physical card" from "card arrival inquiry" in banking domains). Meanwhile, collecting high-quality labeled utterances for emerging long-tail intents remains prohibitively expensive, making generalized zero-shot and few-shot intent recognition a vital necessity.

While prompting Large Language Models (LLMs) to synthesize pseudo-labeled training samples has become a standard recipe to alleviate label scarcity, existing synthesis pipelines hit a severe ceiling in fine-grained intent spaces due to two foundational roadblocks. First, inter-class semantic overlap causes LLMs to overlook subtle pragmatic boundaries, frequently producing ambiguous pseudo-samples that blur decision boundaries across sibling classes. Second, intra-class homogeneity drives generators to over-sample safe, high-probability linguistic templates, resulting in repetitive lexical patterns that cripple the downstream classifier's generalization. Crucially, prevailing remedies rely heavily on inference-time steeringโ€”such as prompt engineering, iterative self-refinement, or heuristic filteringโ€”leaving the generator's underlying parameters completely frozen. Consequently, they cannot eliminate structural generative bias, nor can they gauge the generator's intrinsic reliability in true zero-shot settings devoid of labeled validation data.

To move beyond inference-only heuristics, this paper introduces parameter-level preference optimization for intent data synthesis. Core Idea: represent intent semantics via multi-centroid prototypes, construct block-level preference pairs comprising prototype-consistent diverse positives and cross-intent confusability-weighted hard negatives via Maximal Marginal Relevance (MMR), and iteratively optimize the generator via Direct Preference Optimization (DPO) to synthesize boundary-sharp, lexically diverse pseudo-data prior to downstream classifier training.

Method

Overall Architecture

The proposed framework decouples generator preference alignment from downstream classifier training into a clean, two-phase paradigm. During the alignment phase, an intent-specific anchor set is initialized and mapped into an embedding space to form multi-centroid prototypes that capture multimodal semantic expressions within each intent. At each refresh round, the generator produces candidate utterance blocks conditioned on intent instructions. An MMR-based selection operator then mines prototype-aligned, diverse positives and confusability-weighted cross-intent hard negatives from the pooled candidates to assemble paired preferred and rejected blocks. These structured block pairs directly supervise the generator via Direct Preference Optimization (DPO). As the generator improves across iterative refresh rounds, anchor prototypes and hard negatives are dynamically updated until generator-level reliability signals saturate. Finally, the aligned generator synthesizes pseudo-labeled data to pretrain the downstream discriminative classifier, with real labeled examples reserved exclusively for downstream post-training.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}%%
flowchart TD
    A["Intent Descriptors & Anchor Set Initialization"] --> B["Multi-Centroid Prototype Construction & Dynamic Refresh"]
    B --> C["Block-Conditioned Candidate Generation<br/>Atomic Utterance Blocks"]
    C --> D["Positive & Negative Preference Block Mining<br/>MMR Diverse Positives + Confusability-Weighted Hard Negatives"]
    D --> E["Generator Direct Preference Optimization<br/>LoRA Parameter Update & Iterative Refresh"]
    E -->|Generator Reliability Converged| F["Downstream Classifier Pseudo-Data Pretraining<br/>Real Few-Shot Examples for Post-Training Only"]

Key Designs

1. Anchor Set and Multi-Centroid Prototypes: Capturing Multimodal Intent Manifolds and Mitigating Drift

Single label-name embeddings fail to capture the wide pragmatic dispersion of real-world user queries. The framework maintains a text-level anchor set for each class, mapped via a frozen sentence encoder into an embedding set. In the zero-shot cold start (\(r=0\)), the anchor is initialized solely from the label's descriptive text. Across subsequent refresh rounds (\(r \ge 1\)), high-confidence synthesized utterances are clustered into \(k=5\) normalized centroids:

\[s_c(x) = \max_{j \in \{1,\dots,k\}} \langle e(x), P_c^{(j)} \rangle\]

Here \(s_c(x)\) represents the utterance's prototype similarity. To prevent runaway "bootstrapping drift"โ€”where spuriously frequent pseudo-patterns hijack the semantic centroid in zero-shot regimesโ€”the anchor set is capped at \(K_a=5\) anchors per intent and dynamically pruned using an MMR diversity filter in embedding space, ensuring that the anchor set retains broad semantic coverage rather than collapsing into a narrow mode.

2. Block-Conditioned Generation and Diversity-Aware Positive Mining: Combating Intra-Class Homogeneity

Generating utterances independently one-by-one frequently yields repetitive local artifacts. The framework treats an \(\ell\)-utterance block (\(\ell=4\)) as the atomic unit for both generation and optimization. Repeating block generation \(a\) times constructs an intent candidate pool. To assemble \(M=20\) preferred blocks without replacement from this pool, greedy top-\(\ell\) selection by \(s_c(x)\) would inevitably induce severe mode collapse around near-duplicate phrases. Instead, the method deploys an MMR operator with prototype similarity as the relevance score and pairwise cosine similarity as the redundancy penalty:

\[\hat{x}_u = \arg\max_{x \in \Omega \setminus S} \left[ \lambda \phi^+(x) - (1-\lambda) \max_{x' \in S} \langle e(x), e(x') \rangle \right]\]

with trade-off \(\lambda=0.7\). Sequential MMR packing produces non-redundant, prototype-faithful positive blocks that compel the generator to synthesize varied syntax and lexical realizations.

3. Confusability-Weighted Hard Negative Reservoir: Sharpening Fine-Grained Decision Boundaries

To separate highly adjacent intents, the framework builds a cross-intent hard negative reservoir. For target intent \(c\), the confusability weight of another source intent \(c' \neq c\) is computed using the cosine similarity between their mean prototype vectors under a softmax temperature \(\gamma=0.6\):

\[w(c' \to c) = \frac{\exp(\gamma \langle \bar{p}_{c'}, \bar{p}_c \rangle)}{\sum_{\tilde{c} \neq c} \exp(\gamma \langle \bar{p}_{\tilde{c}}, \bar{p}_c \rangle)}\]

The total budget of \(M\ell=80\) negative utterances per class is allocated across source intents proportionally to \(w(c' \to c)\). The most confusable candidates from each source intent's candidate shortlist are retrieved and packed via MMR into \(M=20\) rejected blocks. Pairing these boundary-challenging negatives against the diverse in-class positives presents a sharp discriminative contrast for preference learning.

4. Format Compliance Rate (FCR) as an Early-Warning Signal for Iterative Refresh Termination

Iterative preference refresh (iDPO) mines progressively harder negatives as the generator evolves, but unconstrained iterations lead to severe over-refresh and degraded controllability. In true zero-shot settings lacking labeled validation splits, the framework tracks three intrinsic generator-side signals: Intent Fidelity Rate (IFR), Neighbor Confusion Rate (NCR@k), and Format Compliance Rate (FCR). FCR measures the exact proportion of generated blocks strictly obeying the requested \(\ell\)-line format without empty lines, extraneous reasoning, or label leakage. Empirically, FCR remains perfect (1.00) through rounds \(R=0\) to \(R=2\), but drops to 0.96 at \(R=3\), which precisely coincides with a steep crash in downstream accuracy (from 83.86% to 76.16%). FCR serves as an effective, self-supervised early-warning indicator that informs stopping at \(R=2\).

Loss & Training

At each refresh round, \(M=20\) block-level preference tuples \((I(c), Y_{c,b}^+, Y_{c,b}^-)\) are constructed per class. The generator is optimized via Direct Preference Optimization (DPO) combined with LoRA (\(r=8\), \(\alpha=32\) on all linear projection layers):

\[\mathcal{L}_{\text{DPO}}(\theta) = -\mathbb{E}_{(I, Y^+, Y^-) \sim \mathcal{D}_{\text{pref}}} \left[ \log \sigma \left( \beta \left( \log \frac{\pi_\theta(Y^+ \mid I)}{\pi_{\text{ref}}(Y^+ \mid I)} - \log \frac{\pi_\theta(Y^- \mid I)}{\pi_{\text{ref}}(Y^- \mid I)} \right) \right) \right]\]

with DPO temperature \(\beta=0.1\). Real labeled few-shot samples are strictly withheld from generator alignment and anchor derivation; they are utilized exclusively to fine-tune the downstream classifier (RoBERTa-base) via standard cross-entropy post-training after synthetic pretraining.

Key Experimental Results

Main Results

Downstream intent classification was benchmarked on BANKING77 (77 intents), HWU64 (64 intents), and CLINC150 (150 intents) across zero-shot and few-shot regimes, standardized with RoBERTa-base and a budget cap of \(N=100\) synthetic samples per class.

Dataset Setting Ours (Qwen3-8B) Prev. SOTA (IDG-CF) Gain
BANKING77 0-shot 78.66 75.26 +3.40
BANKING77 5-shot 84.92 84.41 +0.51
BANKING77 10-shot 90.39 89.67 +0.72
HWU64 0-shot 80.48 75.05 +5.43
HWU64 5-shot 86.25 82.54 +3.71
HWU64 10-shot 88.75 87.51 +1.24
CLINC150 0-shot 83.64 75.63 +8.01
CLINC150 5-shot 94.18 93.20 +0.98
CLINC150 10-shot 95.36 94.75 +0.61

Ablation Study

Ablation experiments evaluated the impact of removing individual components from the full framework (\(R=2\)) on BANKING77 and CLINC150.

Config BANKING77 (0-shot) BANKING77 (5-shot) CLINC150 (0-shot) CLINC150 (5-shot) Note
Full Framework (\(R=2\)) 78.66 84.92 83.64 94.18 Complete framework
w/o Iterative Refresh (\(R=1\)) 73.50 84.22 78.98 92.78 Single round; drops ~5% in zero-shot
w/o DPO (SFT Only) 71.98 83.31 76.04 91.31 Trained on positives only; loses contrastive boundary signal
w/o Hard Negatives (Random) 70.84 82.52 81.36 93.84 Random inter-class negatives fail to separate close neighbors
w/o Confusability Weighting (Uniform) 76.72 84.77 82.00 93.87 Uniform negative quota allocation ignores density differences
w/o MMR (Top-\(\ell\) Substitute) 69.61 81.53 72.98 91.05 Largest drop (-9.05% / -10.66%); triggers severe mode collapse

Key Findings

  • MMR Diversity Control is the Primary Defense Against Mode Collapse: Disabling MMR selection produced the most catastrophic drop across all ablations (falling by 9.05% on BANKING77 0-shot and 10.66% on CLINC150 0-shot), demonstrating that naive top-ranked selection breeds highly repetitive phrases that ruin downstream feature diversity.
  • Parameter Optimization Fundamentally Outperforms Inference Prompting: Compared against the strongest prompting and feedback baseline IDG-CF, parameter-level DPO tuning yields massive improvements (+3.40% to +8.01% in zero-shot) while bypassing multi-turn inference latency during data synthesis.
  • Intrinsic Reliability Correlates Closely with Downstream Accuracy: Preference alignment increases the Intent Fidelity Rate (IFR) from 0.7922 to 0.9804 and slashes the Neighbor Confusion Rate NCR@3 from 0.1469 down to 0.0065 on BANKING77.
  • Robust Transfer Across Diverse LLM Backbones: Consistent gains replicated seamlessly on Llama3-8B (78.83% on BANKING77 0-shot) and Gemma2-9B (76.59%), confirming that the methodology is model-agnostic.

Highlights & Insights

  • Adapting DPO from General Alignment to Domain Data Generation: Instead of aligning LLMs for general chat helpfulness or safety, DPO is reframed as a synthesis mechanism where boundary-confusable negative utterances and diverse in-class positive utterances form preference pairs, directly shaping generative decision boundaries.
  • Block-Conditioned Generation with Decoupled Pooled MMR Re-ranking: Generating in atomic blocks forces grammatical coherence while pooling and sequentially re-ranking via MMR decouples the final preference tuples from single-block generation noise.
  • Format Compliance as an Unsupervised Stopping Criterion: The empirical observation that structural format violations precede semantic degradation provides an elegant, zero-cost early stopping rule for zero-shot iterative fine-tuning.

Limitations & Future Work

  • Cold-Start Sensitivity to Underspecified Label Descriptors: In generalized zero-shot scenarios where class names are terse (e.g., banking action labels omitting specific constraints), initial candidate generation may skew toward incorrect distributions, potentially amplifying drift during anchor updates. Incorporating retrieval-augmented descriptor enrichment could alleviate this issue.
  • Coupled Hyperparameter Complexity: The pipeline involves multiple tunable parameters (number of centroids \(k\), MMR balance \(\lambda\), confusability temperature \(\gamma\), and DPO regularization \(\beta\)), which currently rely on manual calibration.
  • Computational Footprint of Iterative Fine-Tuning: While executing fully on a single consumer GPU (NVIDIA RTX 4090), the multi-round generate-mine-train loop naturally requires more compute than frozen, prompt-only generation.
  • vs IDG-CF (Zhang et al., 2025): IDG-CF guides frozen LLMs via iterative contrastive feedback prompts at inference time; this work directly updates generator parameters via DPO, achieving substantially higher intent fidelity and diversity without multi-step inference chains.
  • vs PromptMix / CoDa (Sahu et al., 2023; Evuru et al., 2024): Traditional LLM augmenters synthesize interpolated or constrained sentences without explicitly modeling cross-intent neighbor confusability; this paperโ€™s confusability-weighted reservoir precisely targets the most confusable boundary failure modes.

Rating

  • Novelty: โญโญโญโญโ˜† Repurposing DPO with multi-centroid prototypes and confusability weighting provides an original, principled approach to intent data synthesis.
  • Experimental Thoroughness: โญโญโญโญโญ Comprehensive evaluation across three classic benchmarks, multiple data regimes, three LLM backbones, and detailed generator-side reliability audits.
  • Writing Quality: โญโญโญโญโญ Clear exposition, disciplined mathematical formulation, and well-structured empirical analyses.
  • Value: โญโญโญโญโญ Highly actionable paradigm for production dialogue systems facing expanding intent taxonomies under low-resource constraints.