ThemeClouds: Concept Visualization for Interviews

• ThemeClouds is an open-source system that creates participant-weighted, concept-level word clouds from qualitative interview transcripts.
• It uses LLM-based concept elicitation and semantic mapping to overcome limitations of traditional frequency-based word clouds.
• The tool empowers researchers to audit and refine thematic summaries for early-stage qualitative analysis and data triage.

ThemeClouds is an open-source visualization system for qualitative interview analysis that produces concept-level, participant-weighted word clouds from dialogue transcripts. Instead of sizing words by raw token frequency, it elicits short thematic phrases with a LLM, maps those concepts back onto individual transcripts through semantic judgment, and weights each concept by the number of unique participants who mentioned it. The result is a cloud intended to summarize breadth of mention rather than repetition of surface forms, especially in small, noisy, conversational corpora where filler, paraphrase, and multi-word concepts make classical frequency clouds misleading (Colonel et al., 11 Aug 2025).

1. Conceptual basis and analytical target

ThemeClouds is motivated by a specific failure mode of conventional word clouds in qualitative interviews. Spoken transcripts contain disfluencies, discourse markers, response scaffolding, interviewer language, and lexical variation, so token-frequency displays often foreground how people speak rather than what they mean. In this setting, semantically shared concerns are fragmented across paraphrases such as “in the way,” “kind of distracting,” “I kept noticing the device,” “it made me self-conscious,” and “felt watched,” even when the analytic question is simply how many participants raised a given issue (Colonel et al., 11 Aug 2025).

The system therefore shifts the representational unit from tokens to concepts and the weighting rule from frequency to breadth. A ThemeClouds item is typically a short phrase such as “image quality,” “felt watched,” “Not distracting,” or “Blends into the desk,” rather than a unigram or stem. A concept contributes at most one count per transcript or participant, so repeated mention by a single talkative participant does not inflate its visual prominence. This makes the cloud closer to qualitative claims of the form “many participants raised this” than to corpus-linguistic claims of the form “this word occurred often” (Colonel et al., 11 Aug 2025).

The intended use is early-stage sensemaking, communication, and triage rather than full thematic analysis. The method is explicitly framed as a lightweight, pragmatic first-pass summary for interview corpora, particularly for dialogue-based data and automatic speech recognition transcripts. In the reported evaluation, the transcripts were produced with Whisper, and the authors retain punctuation and stop words for the LLM-assisted pipeline rather than applying aggressive lexical normalization before theme extraction (Colonel et al., 11 Aug 2025).

2. Formal model and computational pipeline

ThemeClouds formalizes a corpus as a set of transcripts

$\mathcal{T}=\{t_1,\dots,t_M\}$

and a concept inventory as

$\mathcal{C}=\{c_1,\dots,c_N\}.$

For each transcript–concept pair, the default representation is a binary presence assignment

$y(t,c)\in\{0,1\},$

with an optional soft-score mode

$\hat{p}(t,c)\in[0,1]$

and threshold $\tau$ for binarization. The central statistic is participant breadth,

$b(c)=\sum_{t\in\mathcal{T}} y(t,c),$

and the visual weight is

$w(c)=g(b(c)),$

where $g(\cdot)$ is a monotone scaling function. The reported options are linear, logarithmic, and square-root, with linear as the default. For comparative displays across two conditions, ThemeClouds defines

$\Delta b(c)=b_{\text{A}}(c)-b_{\text{B}}(c),$

which underlies condition-wise “diff clouds” (Colonel et al., 11 Aug 2025).

The pipeline has three stages. First, concept elicitation is performed at corpus or condition level. An LLM is prompted to propose a compact list of short, semantically meaningful concept phrases while ignoring filler words, repeated question prompts, and generic study language. In the reported evaluation, the model is LLaMa-3.3-70B-Instruct, and it is asked to identify exactly 20 meaningful and distinctive words or short phrases for each webcam setup (Colonel et al., 11 Aug 2025).

Second, concept mapping is performed transcript by transcript. Given the fixed concept list and the full text of one transcript, the LLM decides which concepts are meaningfully reflected in that transcript, including cases where the participant implies or clearly expresses the concept without using the same wording. This is the semantic normalization mechanism: paraphrase grouping is not implemented through stemming, synonym dictionaries, or a separate clustering algorithm, but through prompt-driven concept elicitation followed by fixed-vocabulary semantic matching (Colonel et al., 11 Aug 2025).

Third, participant-weighted aggregation and visualization are applied. The system counts how many distinct transcripts contain each concept and renders a word cloud whose items are concept phrases and whose sizes reflect participant breadth. The cloud thus preserves the familiar visual form of a word cloud while changing the semantics of both the displayed units and their weights (Colonel et al., 11 Aug 2025).

Stage	Input	Output
Concept elicitation	Corpus or condition-specific transcripts	Short concept list
Concept mapping	One transcript + fixed concept list	Binary or thresholded concept assignments
Aggregation and visualization	Assignment matrix	Participant-weighted cloud or diff cloud

3. Research workflow, transparency, and analyst control

A defining feature of ThemeClouds is that it is not presented as autonomous qualitative analysis. The workflow is explicitly researcher-in-the-loop. Researchers can load a transcript corpus, run LLM-based concept elicitation, inspect the concept list, run transcript-level mapping, review assignment tables, generate participant-weighted clouds, revise prompts, change granularity, and export outputs for downstream coding or reporting (Colonel et al., 11 Aug 2025).

The system’s transparency mechanisms are central to that stance. It exposes prompts, concept lists, assignment tables, scaling choices, and per-concept participant counts. The persistent assignment table has rows as transcripts and columns as concepts, so any cloud can be reconstructed from auditable intermediate data. This table also supports inspection of why a concept appears large, export into downstream thematic workflows, and correction of questionable assignments (Colonel et al., 11 Aug 2025).

Control is provided at several levels. Researchers can adjust the number of concepts $N$, change word cloud layout, choose font scaling, edit prompts, use prompt variants, seed or pin concepts of interest, rerun elicitation to split broad concepts, and audit or correct transcript-level mappings. This design treats LLM assistance as a semantic grouping and assignment layer inside a configurable workflow rather than as a substitute for analytic judgment (Colonel et al., 11 Aug 2025).

This suggests a specific epistemic role for the tool. ThemeClouds is strongest when the task is orientation: identifying what many participants noticed, producing a compact summary for collaborators or stakeholders, and seeding later coding. It is not framed as a method for close reading, theory-building, or final interpretive synthesis. The paper repeatedly positions the output as a first-pass overview and a scaffold for subsequent qualitative work rather than a terminal analysis product (Colonel et al., 11 Aug 2025).

4. Empirical evaluation and comparative behavior

The reported evaluation uses a clinical psychology research study on hardware setups for psychiatric outpatient clinical assessments. The corpus contains 31 participants and 155 interviews or transcripts, reflecting five device conditions evaluated by each participant: insta, single iphone, dual iphones, logitech, and obsbot. The interviews were one-hour, in-person sessions conducted as naturalistic dialogue between participants and a clinical research coordinator, and audio was transcribed with Whisper (Colonel et al., 11 Aug 2025).

ThemeClouds is compared against three baseline families. The first is a standard frequency-based word cloud built from raw word frequency with NLTK stop-word removal. The second is Latent Dirichlet Allocation trained on the full 155-document corpus. The third is BERTopic trained on the same corpus. The evaluation is qualitative rather than a controlled user study, with assessment centered on legibility of surfaced themes, alignment with analyst notes, relevance to device-specific concerns, actionability, robustness on small conversational transcripts, suppression of filler, and merging of semantically related expressions (Colonel et al., 11 Aug 2025).

The reported outputs illustrate the difference in representational quality. ThemeClouds surfaces phrase-level descriptors such as “Small and compact,” “Not distracting,” “Easy to ignore,” “Less noticeable,” “Not too visible,” “Fades into the background,” “Simple and straightforward,” “Convenient,” and “Reminds me of a Polaroid.” Elsewhere, examples include “Discreet,” “Blends into the desk,” “felt watched,” “image quality,” and “in the way.” By contrast, BERTopic outputs are dominated by conversational filler and generic tokens such as “yeah, like, maybe, okay, whatever,” while LDA topics similarly concentrate on high-frequency surface words such as “like,” “okay,” “think,” “would,” and “little” (Colonel et al., 11 Aug 2025).

The comparative interpretation is straightforward. ThemeClouds suppresses filler and interviewer scaffolding because the elicitation prompt asks the model to ignore them; it preserves phrase-level meaning rather than breaking multi-word concepts into fragments; it merges paraphrases through semantic matching; and it weights by breadth across participants instead of repetition within transcripts. The paper further reports that these participant-weighted concept lists aligned more closely with analyst field notes and recorded per-device concerns than either frequency clouds or the topic-modeling baselines (Colonel et al., 11 Aug 2025).

5. Relation to adjacent visualization and cloud-atlas traditions

ThemeClouds belongs to a family of cloud-like summarization methods, but its methodological commitments differ sharply from both classical tag clouds and pattern-discovery atlases in other domains. The closest software-engineering analogue in the provided literature is SoftCloud, which builds tag clouds from multiple kinds of software artifacts. SoftCloud extracts words, splits compound identifiers using camel-case rules, stems with WordNet, and weights tags by raw repetition count; font size encodes frequency, color is decorative, and there is no semantic grouping, theme extraction, or thematic layout. In that sense, SoftCloud is artifact-generic but still fundamentally a frequency cloud, whereas ThemeClouds uses LLM-mediated concept elicitation and participant-weighted aggregation (Al-Msie'deen, 2021).

The contrast with topic modeling is equally important. ThemeClouds is not presented as LDA or BERTopic with a different display layer. Its concept vocabulary is produced directly as short, human-readable concept phrases, and transcript-level mapping is performed against that fixed vocabulary. A plausible implication is that ThemeClouds trades the latent-variable formalism of topic models for direct semantic prompting and auditability, which is why its outputs remain phrase-level and stakeholder-readable even on small spoken corpora (Colonel et al., 11 Aug 2025).

The term also differs from “theme” usage in cloud-imagery science. AICCA, for example, uses a rotation-invariant autoencoder with hierarchical agglomerative clustering to discover 42 AI-generated cloud classes from MODIS image patches, yielding an atlas of recurring cloud regimes defined by morphology, texture, and multispectral structure rather than by transcript concepts. Those classes are physically interpretable and geographically coherent, but they are image-pattern themes in a climatological atlas, not participant-weighted semantic themes in qualitative interviews (Kurihana et al., 2022). This distinction matters because ThemeClouds addresses thematic summarization of language, whereas AICCA addresses unsupervised discovery of cloud scene types from remote-sensing imagery.

These comparisons clarify ThemeClouds’ niche. It is neither a conventional token-frequency cloud nor a generic topic model, and it should not be conflated with scientific “cloud classification atlases” whose objects are atmospheric patterns. Its defining combination is concept-level elicitation, transcript-level semantic mapping, and participant-weighted visual encoding (Colonel et al., 11 Aug 2025).

6. Limitations, trade-offs, and likely development trajectory

ThemeClouds inherits several limitations from its reliance on LLM judgment. Concept elicitation is prompt-sensitive, so outputs can vary with wording, requested concept count, and study framing. Concept mapping may overgeneralize, miss nuance, or misclassify presence. The system mitigates these risks by fixing the concept vocabulary once elicited, using binary mapping by default, exposing assignment tables for audit, allowing user correction, and supporting analyst-seeded concepts, but it does not claim perfect reproducibility across prompts, runs, or model versions (Colonel et al., 11 Aug 2025).

A second limitation is the granularity trade-off. A concept can become too broad, with phrases such as “camera concerns” potentially obscuring distinctions like “felt watched” versus “image quality.” The proposed remedy is iterative rather than automatic: split concepts, merge concepts, rerun mapping, and compare the resulting clouds. This means the method preserves analyst agency but also requires active curation when thematic granularity is substantively important (Colonel et al., 11 Aug 2025).

A third limitation concerns validity claims. The evaluation is observational and qualitative rather than a controlled user study. The paper does not report task-completion experiments, inter-annotator reliability studies, or benchmark-style semantic ground truth. Its principal empirical claim is narrower: on the reported interview corpus, ThemeClouds produced more legible and more actionable first-pass summaries than frequency clouds and the LDA and BERTopic baselines (Colonel et al., 11 Aug 2025).

Privacy and deployment are also treated as nontrivial. Because interview data may be highly sensitive, the paper assumes de-identification and emphasizes local or compliant deployment support together with caution in consequential settings. This is particularly salient in the reported clinical domain, where prompt framing, model priors, and semantic abstraction could all affect how participant speech is summarized (Colonel et al., 11 Aug 2025).

The forward-looking feature set is modest but concrete. The paper discusses interactive analyses such as per-condition contrasts through diff clouds, along with researcher controls for prompt variants, concept seeding, reruns, and revised granularity. A plausible implication is that ThemeClouds is best understood not as a fixed visualization algorithm but as an auditable interface between transcript corpora, LLM-assisted semantic normalization, and analyst-guided comparative summarization (Colonel et al., 11 Aug 2025).