Conversational DNA: A Visual Language and Interactive Atlas of Human and AI Dialogue

Baihan Lin

cs.HCAug 11, 2025

Conversational DNA: A Visual Language and Interactive Atlas of Human and AI Dialogue

Authors: Baihan Lin

Organizations: Berkman Klein Center For Internet & Society Harvard University Cambridge, MA 02138

Abstract

What makes a conversation hold together when its participants speak across one another? Topic maps offer one view, but they leave the relationships between contributions difficult to inspect. We present Conversational DNA, a visual language and interactive atlas for exploring human and AI dialogue. Speaker strands preserve participation, communicative bases mark moves, and directed pairings connect responses to their targets. Adjustable helix geometry makes speaker switching, response distance, and contribution length visible. Across eight corpora containing 1.57 million source records, the atlas maps 151,489 indexed episodes and connects cohort comparison to source transcripts, local structural alignment, and recorded reply alternatives. On 189 held-out Molweni motif queries, adding target correspondence improves precision@5 from 58.8% to 77.2% for exact annotated structure. Case readings illustrate interleaved participation, delayed responses, and the influence of annotation coverage on apparent collection differences. The system supports a view of conversation as jointly organized activity, with visual patterns serving as starting points for examining evidence rather than substitutes for interpretation.

Figures & tables

Figure 1: The full episode atlas, captured from the running application. Every plotted point corresponds to one of 151,489 indexed episodes; clicking unfolds its conversation. Colors distinguish eight source corpora. Filters and region selection update the cohort panel, including participation, target span, and annotation coverage. The two principal components retain 44.4% of variance in 19 standardized descriptors. Proximity is structural, not semantic; corpus separation can reflect annotation availability. Overlapping windows remain in the map.

Figure 2: Two linked readings from actual interface captures. (A) The visual-grammar specimen shows nine turns and eight speaker strands in Molweni test-9001 . Selected turn 6 ( CLQ ) targets turn 4 across turn 5; turn 7 ( ANS ) targets turn 6. The text below the helix makes this annotation inspectable. (B) The Molweni target–response census contains 10,862 clarification-to-answer link occurrences across indexed windows. Selecting that cell opens an actual containing episode. These are overlapping-window counts, not unique pairs or verified instances of successful repair.

Source subset	Records	Utterances	Episodes
Molweni	10,000	88,303	13,033
Reddit discourse	9,483	115,827	19,157
CGA–CMV	6,842	42,964	6,466
DeliData	500	14,003	2,296
WildChat shard	36,821	217,432	32,454
PRISM	8,011	53,646	9,530

Table 1: Actual indexed subsets. Records follow each source’s conversation unit and retain source-level duplicates. Episodes are bounded windows drawn from 117,310 records; the atlas plots every indexed episode.

Figure 3: Implementation components. Corpus adapters preserve source evidence; the episode index stores structural tokens and projection descriptors; local alignment retrieves correspondences; the interface links the atlas, DNA specimen, and source text; exports record the selected evidence and analysis settings. FastAPI serves SQLite data to Canvas and SVG views. Arrows indicate the analysis workflow, not a strictly linear runtime dependency.

Ranking method	P@5	P@10	Success@10
TF–IDF (text)	1.0	1.0	9.5
MiniLM (text)	1.2	1.1	9.5
Sequence + speaker	58.8	52.9	94.7
+ target correspondence	77.2	65.9	98.4

Table 2: Exact annotated motif retrieval, percentages, over 189 family-separated queries. Success@10 indicates at least one relevant result. The bottom two rows share candidates and annotations. This evaluates structural retrieval, not semantic relevance or visual comprehension.

Visual channel	Linguistic reading and measured quantity
Axial order	Sequence: source position, not elapsed time.
Speaker strand	Participation: identity persists through others’ turns; beads mark speech.
Base code	Communicative move: source-derived label, with unknowns preserved.
Directed pairing	Attachment: annotated or platform response target; uptake is inspected in text.
Helix twist	Rhythm: local speaker-switch fraction.
Helix radius	Reach: source-order target distance.

Table 3: The visual language separates linguistic interpretation from an observable encoding. Color identifies speakers or moves; apparent depth is only a viewing cue. Numeric definitions appear in Appendix A .

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: A pinned full-population cohort compared with Molweni episodes containing a nonadjacent eligible response. The selected 11,533 episodes have 76.1% multiparty participation and 89.9% mean known-action coverage, respectively 44.5 and 60.2 percentage points above the full-map baseline. Showing annotation coverage beside participation makes a potential collection confound inspectable. This is a real application screenshot; no causal comparison is implied.

Figure 5: The visual-grammar workspace. A source-linked DNA specimen sits beside the seven declared encoding principles. Users can inspect a base, rotate the viewpoint, or disable geometric channels while retaining the same conversation. The selected clarification’s original text and native target annotation appear below the specimen. The encodings express observed structure; perceptual advantages over lanes and transcripts have not yet been measured.

Figure 6: Four recorded first-round response alternatives in PRISM conversation c14 . Replies at the same source prompt retain model, rating, and selection status. The selected path is observed; unchosen replies have no invented continuation. Ratings are individual participant judgments. This view operationalizes “variants” without equating observations with causal counterfactuals.

Figure 7: Inspecting a retrieved structural analog. The selected Molweni query, test-9001 , is aligned with train-1321 , an exchange about Firefox. The interface exposes the matched turn positions and original words below the ranking results. Similar interaction structure can connect different topics without making the conversations semantically equivalent. Ranking scores describe the application’s current query; they are separate from the held-out benchmark in Table 2 .

Explore similar work

Dial HEALTHDIAL for Advice: A Multilingual and Multi-Parallel Spoken Dialogue Dataset for Knowledge-Grounded Information Seeking
May 28, 2026Songbo Hu, Yinhong Liu, Ej Zhou +5Conversational Datasets Multilingual Healthcare Q&A
Language Technology Lab, University of Cambridge, UK · Independent Researcher · School of Computation, Information and Technology, Technical University of Munich, Germany
PersonaKit (PK): A Plug-and-Play Platform for User Testing Diverse Roles in Full-Duplex Dialogue
May 7, 2026Hyunbae Jeon, Jinho D. ChoiTurn-Taking Full-Duplex Speech Models
Department of Computer Science Emory University Atlanta, GA, USA
Toward Anthropomorphic Dialogue: A Closed-Loop Framework for Human-Like Chat Generation, Evaluation, and Preference Alignment
Jul 19, 2026Wentao Liu, Siyu Song, Xi Chen +4Dialogue Benchmarks Artificial Intelligence Personas
1Shanghai Institute of Innovation · 2East China Normal University · University of Science and Technology of China +3

May 28, 2026, cs.CLJ/K move · Enter open · S save

Dial HEALTHDIAL for Advice: A Multilingual and Multi-Parallel Spoken Dialogue Dataset for Knowledge-Grounded Information Seeking

Songbo Hu, Yinhong Liu, Ej Zhou, Evgeniia Razumovskaia+4

Language Technology Lab, University of Cambridge, UK · Independent Researcher · School of Computation, Information and Technology, Technical University of Munich, Germany

Conversational Datasets Multilingual Healthcare Q&A Multilingual Automatic Speech Recognition Dial Advice

Open paper AskPDF arXiv

Creating spoken dialogue datasets is methodologically challenging, and these challenges are amplified when the goal is to build multilingual, multi-parallel datasets at scale. This work introduces HEALTHDIAL, a large-scale, multilingual, and multi-parallel dataset for developing and evaluating retrieval-augmented generation (RAG)-based spoken dialogue systems. The dataset comprises 6,000 information-seeking dialogues (1,500 per language) grounded in trusted content from the World Health Organization (WHO) and 163 hours of user speech recorded from native speakers of diverse dialects across four official WHO languages: Arabic, Chinese, English, and Spanish. Each speaker is annotated with demographic (e.g., gender, age) and sociolinguistic (e.g., primary language, region of origin) variables. We report benchmark results across key dialogue tasks, which reveal consistent performance disparities across languages, even among high-resource ones. To support future research, we release the dataset, a prototype system, and a toolkit for data collection and system evaluation.

Abstract

Explore similar work

Dial HEALTHDIAL for Advice: A Multilingual and Multi-Parallel Spoken Dialogue Dataset for Knowledge-Grounded Information Seeking

PersonaKit (PK): A Plug-and-Play Platform for User Testing Diverse Roles in Full-Duplex Dialogue

Toward Anthropomorphic Dialogue: A Closed-Loop Framework for Human-Like Chat Generation, Evaluation, and Preference Alignment