Controllable Conversational Theme Detection

• Controllable conversational theme detection is a method that identifies and categorizes core topics in dialogues while allowing user-driven control over theme granularity.
• It employs both global similarity and local density approaches to robustly detect diffuse, interleaved themes in noisy conversational data.
• Recent frameworks integrate user constraints and dynamic weighting to produce semantically coherent labels that enhance downstream analysis and application performance.

Searching arXiv for recent and foundational papers on controllable conversational theme detection. Controllable conversational theme detection is the task of automatically identifying and categorizing the underlying topics or “themes” within conversations while allowing the resulting theme inventory, granularity, or labels to be shaped by explicit control signals such as user preferences, clinician priorities, or instruction prompts. In recent work, the task is distinguished from traditional intent detection because themes are intended as a direct, user-facing summary of the conversation’s core inquiry rather than a fixed schema for downstream logic, and because a single conversation may contain multiple themes whose mentions are interleaved, localized only loosely, or obscured by ASR errors and conversational variation (Shalyminov et al., 26 Aug 2025, Bost et al., 2018).

1. Task definition and conceptual boundaries

Theme detection research treats conversation as a thematically structured object rather than a sequence of isolated utterances. In customer-care speech analytics, themes are defined by application documentation topics, and a conversation may include irrelevant mentions together with multiple themes whose mentions may be interleaved portions of a conversation that cannot be well defined (Bost et al., 2018). In the DSTC 12 formulation, the goal is joint clustering and theme labeling of dialog utterances, with controllability of the resulting theme clusters’ granularity achieved via provided user preference data; the corresponding outputs are concise natural-language labels intended for analytical review (Shalyminov et al., 26 Aug 2025).

A central boundary in the literature is the distinction between theme detection and intent induction or intent classification. Theme detection does not rely on a predefined schema in the sense used by conventional intent systems, and it requires cross-dialogue consistency together with alignment to user-specific preferences or analyst needs (Ke et al., 25 Dec 2025). This makes the task simultaneously a representation problem, a clustering problem, and a label-generation problem.

Another recurring boundary concerns the unit of analysis. Some systems treat the entire dialogue as a bag of features, whereas others localize theme evidence at word positions, segment boundaries, sentence units, or utterance clusters. This division reflects a factual property of conversational data: in some conversations theme mentions are localized in specific segments, while in others mentions cannot be localized cleanly (Bost et al., 2018). A plausible implication is that controllability depends not only on the label space but also on the scale at which thematic evidence is modeled.

2. Foundational approaches: global evidence, local density, and multi-theme speech analytics

Early work on telephone conversations established two complementary paradigms. The first is a global similarity approach in which a conversation $d$ and a theme $c$ are represented by weighted feature vectors over automatically selected words, bigrams, and distance bigrams, and theme assignment is based on cosine similarity (Bost et al., 2018). The similarity score is defined as

$$sc(d, c) = \frac{\sum_{t \in d \cap c} w_d(t) w_c(t)}{\sqrt{\sum_t w_d(t)^2 w_c(t)^2}}$$

with feature selection driven by purity and coverage, using the Gini purity criterion

$$G(t) = \sum_{c \in \mathbb{C}} \left( \frac{df_c(t)}{df_\mathbb{T}(t)} \right)^2$$

and feature weighting

$w_c(t) = df_c(t) \cdot idf^2(t) \cdot G^2(t).$

The second paradigm is local thematic-density modeling. For each word position $p_i$, the contribution of theme $c$ is accumulated from the local feature set and normalized by the theme-vector norm, and the thematic density at position $p_i$ diffuses these contributions across neighboring positions through a decay parameter $\lambda \geq 1$ (Bost et al., 2018). The density is

$$d_c(p_i) = \frac{\sum_{j=1}^n \frac{w_c(p_j)}{\lambda^{|i-j|}}}{\sum_{j=1}^n \frac{1}{\lambda^{|i-j|}}}.$$

The sequence of these densities constitutes a theme “skeleton” of the dialogue. This representation supports soft, context-sensitive localization of dominant themes without requiring hard discourse boundaries.

These two views—global aggregation and local density—were motivated by the same empirical difficulty: real conversations are noisy, theme mentions may be diffuse or localized, and ASR outputs can have very high word error rates due to background noise, accent, and disfluencies (Bost et al., 2018). In the 2018 telephone-conversation study, both the cosine-similarity method and the thematic-density method outperformed a linear kernel SVM trained on the same features. On the test set with automatic transcriptions, both proposed methods achieved accuracy of 0.71 versus 0.64 for SVM, and F-score of 0.80–0.81 versus 0.75 (Bost et al., 2018).

A related line of work in human/human customer-care conversations combined global and local systems in a sequential decision strategy. The strategy first accepted conversations where all four systems agreed, then expanded coverage through lower-consensus stages and a fifth recovery system using ASR-lattice features under semantic preconditions (Bost et al., 2018). On the test set, 4-system consensus yielded coverage 0.54 with accuracy 0.89, 3-system consensus yielded coverage 0.80 with accuracy 0.86, and 2-system consensus yielded coverage 0.98 with accuracy 0.80 (Bost et al., 2018). This established an early form of controllability as an operating-point trade-off between coverage and annotation accuracy.

3. Contextualization, segmentation, and sentence-level clustering

Subsequent work shifted from document-style topic inference toward context-aware conversational modeling. In open-domain human-chatbot dialogues, incorporating conversational context and dialog act features improved topic classification accuracy and unsupervised keyword detection for interactions where topics frequently span multiple utterances (Khatri et al., 2018). The paper evaluated DAN, ADAN, and BiLSTM variants, with context integrated either by concatenating averaged previous turns or by sequential encoding, and dialog acts appended as auxiliary features. The strongest reported topic-classification result was 74% accuracy for BiLSTM-Seq with context and dialog act, compared with a 55% baseline, while CADAN plus dialog act improved keyword recall from 36% to 40% (Khatri et al., 2018).

This line also linked thematic modeling to dialogue quality. Topical depth—the number of sustained topic-specific turns within a conversation—showed Pearson correlation 0.80 with coherence and 0.77 with engagement, both statistically significant with $c$0 (Khatri et al., 2018). This suggests that theme detection is not merely descriptive; it can function as a control-relevant proxy for conversational quality.

A parallel development in personal finance questions argued that multi-theme detection should operate at the sentence rather than question level. The proposed pipeline segmented customer problem statements into sentences, encoded them with alternatives including USE and SBERT-family models, clustered them with KMeans or HDBSCAN, and aggregated sentence-level cluster distributions back to the question (Qiu et al., 2021). The highest-performing configuration was original full sentences encoded by Universal Sentence Encoder and clustered with KMeans, achieving Micro-F1 of 0.46; this outperformed semantic-role-labeling variants, SBERT-family models, HDBSCAN, and TF-IDF baselines (Qiu et al., 2021). The result is notable because it shows that multi-theme conversational analysis can benefit from relatively simple sentence-level partitioning even when more sophisticated encoders are available.

More recent controllable frameworks integrate segmentation directly into representation learning. CATCH uses dual-branch topic segmentation inspired by DialSTART, combining a topic encoder and a coherence encoder to compute a boundary score $c$1, then applies TextTiling to obtain topical segments whose context is fed back into utterance representations (Ke et al., 25 Dec 2025). This design explicitly targets the short, sparse utterance problem that earlier context-aware classifiers and sentence-clustering systems had to address indirectly.

4. Mechanisms of controllability

The defining control signal in the DSTC 12 task is the set of user-provided should-link and cannot-link utterance pairs, which express desired granularity. A should-link requires two utterances to belong to the same theme; a cannot-link requires them to be separated (Shalyminov et al., 26 Aug 2025). The track formalized controllability as the ability to merge or split themes in accordance with these constraints while still producing semantically meaningful clusters and concise labels.

CATCH operationalizes this idea through preference-guided topic clustering. Its semantic-preference distance kernel is

$c$2

where $c$3 is Euclidean distance in embedding space and $c$4 is a preference scalar predicted by a Preference Reward Model trained from a small number of should-link/cannot-link pairs (Ke et al., 25 Dec 2025). The framework performs initial clustering on semantic similarity, replaces the distance metric with the SP kernel, extracts preference conflicts, reclusters conflictive topics, and reassigns pending sub-clusters to anchor clusters. It then uses a hierarchical theme generation procedure—subgroup labeling, noisy-label cleaning by majority voting and semantic filtering, and theme consolidation—to produce coherent labels (Ke et al., 25 Dec 2025).

Clinical dialogue introduces a different control regime. PDIMC first uses in-context learning to extract five explicit themes—family, work, mental, medical, and overall—from clinical interviews, then models intra-theme and inter-theme correlations, and finally applies interactive theme adjustment through external feedback weights (Zhao et al., 16 Feb 2025). The adjustment step is

$c$5

so theme importance can be interactively increased or decreased before fusion. This framework was motivated by the observation that previous end-to-end depression models did not allow clinicians to intervene and focus on themes of interest (Zhao et al., 16 Feb 2025).

Other systems place controllability in the interaction substrate itself. In full-duplex speech, F-Actor encodes conversation topic as a narrative instruction in a prompt prefix together with controls for dialogue initiation, interruptions, backchanneling, and speaker voice (Züfle et al., 16 Jan 2026). In conversational recommendation, the CSHI simulator uses a plugin manager across profile initialization, preference initialization, and message handling, with parameters $c$6 and $c$7 controlling the mixture of known and unknown thematic preferences and an intent-understanding plugin extracting both intent and the relevant attribute or theme (Zhu et al., 2024). Across these systems, controllability is not limited to cluster boundaries; it includes preference priors, label style, and dynamic weighting of thematic evidence.

Control regime	Mechanism	Representative work
Granularity control	Should-link / cannot-link constraints	(Shalyminov et al., 26 Aug 2025)
Preference-aware clustering	Semantic-preference distance kernel and PRM	(Ke et al., 25 Dec 2025)
Interactive weighting	Feedback-adjusted theme representations	(Zhao et al., 16 Feb 2025)
Prompt-level topic control	Narrative instruction prefix	(Züfle et al., 16 Jan 2026)
Preference simulation	Plugin-based known/unknown preference routing	(Zhu et al., 2024)

5. Label generation, evaluation, and benchmark design

Evaluation in controllable conversational theme detection is intrinsically multi-objective. The DSTC 12 track used clustering metrics—Normalized Mutual Information and Accuracy—as well as labeling metrics including ROUGE, cosine similarity of SentenceBERT embeddings, BERTScore, and LLM-as-a-Judge evaluation for adherence to label guidelines (Shalyminov et al., 26 Aug 2025). Human evaluation further decomposed quality into structural criteria such as conciseness and grammatical structure, and functional criteria such as semantic relevance, analytical utility, granularity, actionability, domain relevance, and thematic distinctiveness (Shalyminov et al., 26 Aug 2025). This evaluation design reflects the fact that a controllable theme system must do more than recover clusters: it must produce labels that are useful to analysts and aligned with requested specificity.

The reported benchmark results illustrate the interaction between clustering quality and label quality. In DSTC 12, Team C achieved clustering accuracy 67.97% and NMI 70.39%, together with 100% on label guidelines, outperforming the baseline and other submissions on major metrics (Shalyminov et al., 26 Aug 2025). In CATCH’s offline evaluation on DSTC-12 Banking, the framework achieved 56.7% clustering accuracy, NMI 65.4%, ROUGE-1 35.3%, and LLM evaluation 95.9%, and also led cross-domain evaluation on Finance/Insurance with Acc around 55–56%, NMI above 60, and ROUGE-L above 41.8 (Ke et al., 25 Dec 2025). Ablation results were especially diagnostic: removing preference-guided clustering reduced label alignment CosSim by 7.8%, removing context modeling reduced Acc by 14.2% and NMI by 13.8%, and removing hierarchical generation caused the largest drops in ROUGE and LLM scores (Ke et al., 25 Dec 2025).

Older conversational theme models also exposed an important evaluation principle: multi-theme evaluation is sensitive to the unit and noise model. In call-center speech, accuracy falls on noisy ASR output, but the relative advantage of density-based and global-similarity methods remains (Bost et al., 2018). In personal finance questions, cluster quality remained useful even when some clusters were noisy, because sentences in the same cluster were still topically consistent with the associated label (Qiu et al., 2021). A plausible implication is that controllability metrics should be interpreted jointly with robustness-to-noise and with the semantic usefulness of imperfect clusters.

6. Relation to controllable generation, applications, and open problems

Theme detection increasingly overlaps with controllable generation. In grounded response generation, the CGRG framework injects lexical control phrases—either user-supplied or predicted from dialogue context and grounding knowledge—into a GPT-2-based generator with an inductive attention mask that prevents irrelevant interactions among control phrases and grounding sentences (Wu et al., 2020). The best controlled-and-grounded setting, $c$8 with inductive attention, achieved BLEU 1.80, NIST 3.26, and Div-2 11.6, outperforming comparison settings and demonstrating that explicit topic-bearing control variables can steer response content while maintaining grounding (Wu et al., 2020).

Adapter-based and instruction-following dialogue systems extend this logic to theme or skill selection. Adapter-Bot inserts skill-specific residual adapters into a frozen backbone such as DialoGPT and uses a BERT-based dialogue manager to choose the adapter automatically or allow manual selection (Madotto et al., 2020). On the Plug-and-Play styles dataset, its style/topic accuracy was 70.01%, and the dialogue manager achieved 95.35% accuracy in multi-turn and 92.92% in single-turn selection (Madotto et al., 2020). F-Actor applies the same principle to duplex speech: with word-level alignment and audio delay, it achieved narrative adherence 2.39 on a 1–5 scale versus a random baseline of 1.26, while ground truth scored 4.04; removing the instruction prompt caused theme control to collapse (Züfle et al., 16 Jan 2026).

Applications span customer support, finance, recommendation, and clinical interviewing. In open-domain dialog, topical metrics correlate strongly with coherence and engagement, implying that conversational topic models can predict user satisfaction (Khatri et al., 2018). In recommendation, CSHI uses controllable, human-involved simulation to model explicit and latent thematic preferences, enabling realistic elicitation and evaluation scenarios (Zhu et al., 2024). In depression screening, PDIMC shows that explicit and interactively controllable theme modeling can improve downstream clinical prediction, with absolute improvements of 35% and 12% compared to prior state of the art on DAIC-WOZ (Zhao et al., 16 Feb 2025).

Several open problems are explicit in the literature. Granularity remains hard: DSTC 12 reports persistent over- and under-clustering, especially beyond the small set of explicitly constrained preference pairs (Shalyminov et al., 26 Aug 2025). Generalization across domains is still a stress point, as shown by held-out Travel evaluation in DSTC 12 and cross-domain analysis in CATCH (Shalyminov et al., 26 Aug 2025, Ke et al., 25 Dec 2025). Label generation is also nontrivial; systems often default to noun phrases or excessive specificity unless prompted or structured carefully (Shalyminov et al., 26 Aug 2025). Finally, there is a conceptual misconception that stronger topical relatedness alone guarantees better conversation. In controllable chitchat, response-relatedness control did not yield clear improvement in human scores, whereas repetition control and calibrated question-asking had larger effects on perceived quality (See et al., 2019). This suggests that theme detection is a necessary but not sufficient condition for high-quality conversational behavior.