PaperlayerView exact revision on arXiv

Source-linked AI summary

Computational KJ-Ho: An Analyst-Bias-Free Insight Extraction Framework from Large-Scale Qualitative Data Using Domain-Specialized LLMs

Kasumi Ban

arXiv:2608.16467v1cs.HCcs.CLcs.CY

TL;DR

Qualitative insight generation is constrained by analyst cognitive capacity and variation across analysts. This concept paper proposes Computational KJ-Hō, a domain-specialized, two-stage CPT-plus-SFT framework with three analytical layers. Preliminary studies found agreement rose from 0.091 to 0.317 after a model-generation update, while semantic precision remained below domain-expert requirements.

  • Problem

    Qualitative analysis is limited by human processing capacity and substantial disagreement among analysts interpreting identical data.

  • Method

    Computational KJ-Hō combines CPT-based marketing specialization with SFT for analytical operations in a three-layer architecture.

  • Results

    Agreement rose from 0.091 to 0.317 after a model-generation update, but general-purpose models remained below domain-expert semantic precision.

  • Takeaways & Limitations

    The framework aims to extend data-emergent KJ analysis beyond individual analysts’ cognitive capacity while preserving its epistemological character.

  • Takeaways & Limitations

    This concept paper is presented before empirical validation, and validation through industry collaboration remains necessary.

Abstract

from arXiv · show

The qualitative research methodologies that underpin consumer-insight generation - the KJ method, Grounded Theory, and Thematic Analysis - share a structural constraint: the cognitive processing capacity of the human analyst. Replication research further shows that conclusions vary substantially across analysts analyzing identical data (analyst bias). This paper proposes Computational KJ-Ho (the Kawakita Jiro method), a theoretical framework that computationally realizes the KJ method's epistemology - letting structure emerge from the data itself without imposing the analyst's preconceptions - an orientation we term "analyst-bias-free." The framework employs a domain-specialized LLM built through continued pre-training (CPT) on a marketing-research corpus and supervised fine-tuning (SFT) on expert-curated insight pairs, organized as a three-layer architecture: data structuring, insight extraction, and strategy generation. Two preliminary studies in the Japanese marketing context support the necessity of CPT-based domain specialization. The paper makes five contributions: (1) a theoretical integration of the KJ method, Grounded Theory, and Peircean abduction into a single epistemological commitment of data-driven explanation generation; (2) a three-layer architecture leveraging domain-specialized embeddings for cross-interview analysis; (3) two novel evaluation metrics, InsightExtraction-F1 and MarketingQA; (4) explicit engagement with the WEIRD problem, centering a non-Western methodology; and (5) five practice-derived problem formulations from nearly three decades of marketing-research practice, translated into design requirements. The human analyst retains a supervisory role. This is a concept paper presented ahead of empirical validation.

1 Introduction

The introduction frames qualitative research as constrained by analyst bias, escalating data volume, limited reproducibility, and Western-centric assumptions, then proposes Computational KJ-Hō as a scalable, data-grounded framework for consumer-insight extraction. The concept integrates KJ, Grounded Theory, and abductive reasoning through domain-specialized computational layers, new evaluation metrics, cultural-validity requirements, and practice-derived design problems, while remaining prior to empirical validation.

  • Proposed framework: Computational KJ-Hō computationally realizes the KJ method by letting meaning structures emerge from the data rather than analysts’ pre-existing conceptual frameworks, while excluding conscious analyst preconceptions rather than all model premises.The framework is explicitly presented as a theoretical proposal intended to address the identified problems, not as a claim that models are premise-free.
  • Theoretical contribution: The framework integrates the KJ method, Grounded Theory, and abductive reasoning into a scalable paradigm for analyst-bias-free consumer-insight extraction, placing a non-Western methodology at its core.The KJ method is presented as the framework’s axis, while the integration addresses qualitative analysis’ need for data-driven explanation generation.
  • Architecture: Its three-layer architecture supports cross-dataset analysis with marketing-domain embeddings, including structuring, labeling, grouping, relational mapping, and insight generation by a domain-specialized LLM.The core computational layer converts text into structured data, assigns labels, clusters groups, adjusts hierarchy, maps relationships, and produces insight statements.
  • Evaluation: InsightExtraction-F1 and MarketingQA are proposed to evaluate domain-specific insight quality beyond standard NLP metrics, complementing Pi et al.’s [51] five codebook-quality metrics.The introduction argues that standard NLP metrics capture surface similarity rather than analytical value.
  • Cultural scope: The paper treats cultural validity as a design requirement by centering a Japan-originated methodology, addressing the WEIRD problem, and planning validation in non-WEIRD markets.It identifies Western, English-language development and validation as limitations of many existing systems and connects this concern to qualitative analysis.
  • Motivation and status: The framework responds to escalating qualitative-data volume, inconsistent analytical conclusions, and practice-derived constraints, but is presented before empirical validation and requires future industry collaboration.The introduction cites 55.1% replication across 164 papers and 34% complete agreement across 100 analyst re-analyses, while describing nearly three decades of marketing-research experience as the source of its five problem formulations.

2 Intellectual Foundations

The KJ method, Grounded Theory, and Peircean abduction share an epistemological commitment to generating explanations from data rather than imposing them externally. Computational KJ-Hō extends this commitment beyond individual analysts’ cognitive limits through domain-specialized computational representations.

  • KJ method: The KJ method organizes qualitative data by fragmenting information into labels, grouping similar items, arranging relationships spatially, rendering the structure into prose, and reaching consensus through collective revision.Its process uses flow, satellite, tree, and cycle structures while reflecting semantic proximity in spatial positioning.
  • KJ method: KJ’s defining epistemological principle is “letting the data speak”: group structures should emerge through iterative integration rather than being imposed by the analyst’s prior theoretical framework.The method’s ultimate purpose is idea generation and collective decision-making, not verification of what is already known.
  • Motivation: The conventional one- to two-hundred-card limit reflects the physical and cognitive capacity of a single analyst, motivating computational extension without abandoning data-emergent categorization.The framework questions the volume constraint, not the KJ method’s epistemological soundness.
  • Computational realization: Computationally realizing KJ’s “felt closeness” requires domain-specialized embeddings that encode contextual knowledge, while remaining distinct from mechanical Hierarchical Agglomerative Clustering.The proposed interpretation treats proximity as an expert-like, gestalt-oriented affinity rather than similarity-matrix grouping.
  • Common epistemological foundation: KJ, Grounded Theory, and abduction converge on generating explanatory elements from data: “letting the data speak”, “discovering theory”, and producing the “most plausible explanation”.The framework characterizes domain-specialized models’ generation of consumer insights absent from training data or predefined hypotheses as “functional abduction.”

3 Related Work: LLMs in Qualitative Research

Prior LLM-based qualitative-analysis systems automate coding, thematic analysis, grounded theory, or collaborative interpretation, but general-purpose models remain limited in domain knowledge and interpretive depth. Computational KJ-Hō therefore distinguishes itself through domain adaptation, two-stage training, and analysis of consumers’ actual statements rather than synthetic consumer simulation.

  • LLM-based qualitative analysis: LOGOS and Auto-TA automate broad qualitative-analysis pipelines, with LOGOS reporting up to 88% agreement with experts, but both rely on general-purpose LLMs without internalized domain knowledge [51] [66].LOGOS builds a hierarchical codebook through graph reasoning, while Auto-TA uses four role-conditioned agents and iterative refinement for Steps 1–3 of Reflexive Thematic Analysis.
  • LLM-based qualitative analysis: NGT achieved a 168-fold speedup over manual analysis, but Computational KJ-Hō differs by reshaping domain-specific embeddings through CPT rather than relying on general-purpose embeddings [65].NGT used OpenAI’s text-embedding-3-small, whereas this study follows the BioBERT paradigm for domain adaptation.
  • LLM-based qualitative analysis: SFT-TA is the closest prior work in training strategy, while Computational KJ-Hō combines SFT with CPT and domain-specialized embeddings for coding and insight extraction [67].SFT-TA embeds supervised-fine-tuning agents in both coding and theme-generation phases, whereas the proposed framework uses two-stage training.
  • Collaborative qualitative analysis: Collaborative systems such as CollabCoder, ThemeViz, and MindCoder augment human interpretation through interactive suggestions, visualization, or auditable reasoning, but users continue to position AI as support rather than an equal collaborator [25].ThemeViz found greater utility than ChatGPT with no observed hallucinations, while 27 of 28 participants still regarded the AI as a tool; MindCoder’s evaluation involved 12 users and 2 external reviewers.
  • LLMs in marketing research: Marketing-research LLM studies also simulate consumers, whereas Computational KJ-Hō analyzes what consumers actually said, addressing data-analysis depth and scale rather than data-collection cost and speed.Brand et al. found statistically comparable conjoint-survey responses and improved alignment after fine-tuning; this synthetic-consumer use is complementary rather than competing.

4 The KJ-Method Computational Structure · 4.1 Overview of the Computational Structure · Layer 2: Extraction and structuring of consumer insights from that data

Computational KJ-Hō uses a three-layer, sequential architecture whose outputs feed forward while each layer remains independently evaluable. A domain-specialized LLM drives the framework through marketing-domain continued pre-training followed by expert-curated supervised fine-tuning.

  • 4 The KJ-Method Computational Structure: The framework comprises three analytical layers: raw qualitative-data ingestion and structuring, consumer-insight extraction and structuring, and marketing-strategy conversion.The layers correspond to distinct analytical functions and connect structured insights to implementable outputs.
  • 4.1 Overview of the Computational Structure: Each layer feeds the next sequentially but has independent design decisions and evaluation criteria, enabling targeted quality control and improvement.This preserves the overall pipeline while keeping analytical processes separable.
  • Layer 2: Extraction and structuring of consumer insights from that data: A domain-specialized LLM drives the architecture through two-stage learning: continued pre-training on marketing-related corpora followed by supervised fine-tuning on expert-curated paired data.The paired format maps consumer-generated text to insights and then to marketing strategies and measures.
  • Layer 2: Extraction and structuring of consumer insights from that data: The CPT corpus incorporates consumer psychology, business, economics, law, and accumulated marketing-research knowledge to represent the marketing domain.The supplied passages specify these resource categories as minimum corpus requirements and emphasize skilled researchers’ practical knowledge.
  • Layer 2: Extraction and structuring of consumer insights from that data: The framework adds best-practice and benchmarking knowledge through company histories, reflecting practical learning from superior firms and Newton’s “standing on the shoulders of Giants”.Japanese-style shashi are presented as benchmark information on business models and marketing strategy.
  • Layer 2: Extraction and structuring of consumer insights from that data: CPT must precede SFT because domain knowledge is treated as a precondition for task adaptation and for representing domain-relevant distinctions, addressing Prescott et al..The paper argues that fine-tuning without prior domain knowledge would leave analysis in an inadequate semantic space.
  • Layer 2: Extraction and structuring of consumer insights from that data: Expert re-annotation illustrates the value of quality control: TACRED’s difficult data required over 50% relabeling, raising baseline average F1 from 62.1 to 70.1 [3].The example supports expert review of training or evaluation data rather than relying uncritically on crowdsourced labels.
  • Layer 2: Extraction and structuring of consumer insights from that data: The envisioned empirical model has 17 to 27 billion parameters and is designed for local operation without external APIs, supporting confidentiality and operational autonomy.The scale is framed as a balance between analytical capability and control over sensitive inputs, processes, and outputs.

4.2 The Three-Layer Architecture Layer 1: Ingestion and Structuring of Qualitative Data

Layer 1 ingests and structures large-scale qualitative data by integrating respondent attributes, extracting meaning from utterances, constructing multidimensional representations, and segmenting content into atomic data units.

  • Layer 1: Ingestion and Structuring of Qualitative Data: Layer 1 integrates speaker attributes, extracts intent, emotion, and behavioral motivation, links multidimensional representations to respondent attributes, and segments data into coherent atomic units.Segments may comprise individual utterances, exchanges, or thematically coherent fragments that serve as Computational KJ-Hō’s atomic data units.

Layer 2: Consumer-Insight Extraction · Step 1—Semantic Clustering (KJ method: group formation)

Layer 2 extracts marketing-significant consumer insights from structured qualitative data without hypotheses, implementing KJ grouping, heading assignment, and spatial arrangement. Step 1 uses CPT-trained, domain-specialized embeddings to form emergent cross-interview clusters that preserve consumer-motivation distinctions and support action design.

  • Layer 2: Consumer-Insight Extraction: Layer 2 analyzes qualitative data cross-sectionally to detect latent themes, inter-segment differences, consumer psychology, and purchase motivations or barriers, then verbalizes marketing-significant insights.The process does not require hypotheses and operates on data structured in the first layer.
  • Layer 2: Consumer-Insight Extraction: The Computational KJ-Hō core executes “grouping labels → assigning headings → spatial arrangement” as its computational realization of the KJ method.
  • Step 1—Semantic Clustering (KJ method: group formation): CPT-trained embeddings position utterances relative to consumer-behavior semantics rather than general language, producing groupings aligned with latent consumer motivations.The model internalizes marketing-research vocabulary and consumer-psychology concepts through continued pre-training.
  • Step 1—Semantic Clustering (KJ method: group formation): The model distinguishes opposite motivations behind conceptually similar terms, such as “high price” as dissatisfaction or overpricing versus “premium” as aspiration or value recognition.This finer semantic resolution makes grouping structures more directly tied to action design than general embeddings.
  • Step 1—Semantic Clustering (KJ method: group formation): Without predefined categories, the domain-specialized model generates semantically related clusters through methods such as hierarchical agglomerative clustering.The cluster count emerges from embedding structure through an SFT-tuned stopping criterion, preserving data-driven grouping.
  • Step 1—Semantic Clustering (KJ method: group formation): Clustering operates on the entire corpus simultaneously, allowing a single cluster to draw from multiple interviews and enabling cross-interview analysis.
  • Step 1—Semantic Clustering (KJ method: group formation): Domain-specialized vector similarity approximates experts’ grouping intuitions by encoding contextual marketing knowledge and placing expert-consistent data near one another.CPT structurally represents distinctions such as price sensitivity versus perceived value, providing the semantic foundation for clustering.

Step 2—Assigning Meaning to Groups (creating headings) · Step 3—Structuring the Relationships Among Groups (spatial arrangement and structuring)

Step 2 converts each cluster into a marketing-significant heading grounded in its constituent data, while Step 3 maps relationships among those groups to make their configuration explicit and actionable.

  • Step 2—Assigning Meaning to Groups (creating headings): The SFT-trained model generates a natural-language heading for each cluster.Each heading integrates the meaning of the cluster’s constituent data into a marketing-significant proposition.
  • Step 2—Assigning Meaning to Groups (creating headings): Each generated insight includes an explicit link to its supporting data.This linkage enables human reviewers to audit the reasoning and verify its empirical basis.
  • Step 2—Assigning Meaning to Groups (creating headings): The SFT-trained model is contrasted with general-purpose LLMs that depend on the prompt designer’s hypotheses and framing.The model is trained on expert-curated paired data, as described in the supplied passage.
  • Step 3—Structuring the Relationships Among Groups (spatial arrangement and structuring): Step 3 uncovers relationships among the groups produced in Step 2.The relationships may be causal, contradictory, hierarchical, complementary, or of other types.
  • Step 3—Structuring the Relationships Among Groups (spatial arrangement and structuring): The step produces a structured map that makes the groups’ relational configuration explicit and explorable.Examples include opposition between a purchase motivation and a barrier, and a latent need serving as a superordinate concept of a behavioral pattern.
  • Step 3—Structuring the Relationships Among Groups (spatial arrangement and structuring): Explicit relational structuring prepares an actionable input for strategy construction and planning.The map connects the organization of groups to the subsequent strategy-building process.

Layer 3: Strategy Generation

Layer 3 connects structured consumer insights to marketing strategy, producing implementable outputs for business application. It extends the framework beyond the KJ method’s core process into practical strategy generation.

  • Layer 3: Strategy Generation: Layer 3 connects structured consumer insights to marketing strategy and converts them into implementable outputs.It translates the second layer’s structured insights into practical business applications.
  • Layer 3: Strategy Generation: Outputs define the insight statement, target segment, business opportunities, and brand communication ideas.The communication output includes messages and creative ideas designed to appeal to the target.
  • Layer 3: Strategy Generation: The system automatically generates an evidence-cited analytical report with executive summary, detailed analysis, and proposals for additional research.Additional research addresses issues left unresolved in the present analysis.
  • Layer 3: Strategy Generation: This layer does not directly implement the KJ method’s core process, instead extending the framework from academic methodology to practical business application.It addresses the fourth problem identified in Section 1.2.

4.3 The Role of Domain Specialization: Why General-Purpose LLMs Are Insufficient

Preliminary studies show that general-purpose LLMs can execute the conceptual pipeline but miss the fine-grained semantic distinctions required by skilled marketing researchers. CPT-based specialization is therefore positioned as a white-box alternative to prompting and retrieval approaches, while supporting data sovereignty.

  • Preliminary evidence: Preliminary studies confirm that general-purpose LLMs support the analytical pipeline conceptually but fail to reach marketing researchers’ important semantic distinctions.These studies establish conceptual and methodological baselines for evaluating improvement from domain specialization in Section 5.
  • Preliminary evidence: 0.091 to 0.317 agreement with human labels improved across GPT-3.5-turbo and GPT-4o-mini, but remained below expert-level semantic precision despite processing 779 sake reviews in about 10 minutes versus 19 hours manually.Sentiment agreement reached 0.783, whereas distinctions such as drinking temperature and price range were poorly captured.
  • Why CPT: Domain-specific CPT reshapes the embedding space so domain-relevant distinctions between generally synonymous words become representable, unlike approaches that rely on prompts to specify the framework.Lee et al. attributed BioBERT’s gains over general-purpose BERT in biomedical named-entity recognition and relation extraction to domain adaptation through pre-training.
  • Data sovereignty: Data sovereignty is a primary reason for domain specialization because confidential markets, consumer data, and hypotheses cannot leave the company, limiting reliance on cloud-based tools.The section presents sovereignty alongside the continuing evolution of general-purpose models rather than denying that evolution.
  • Alternative approaches: Few-shot prompting, RAG, and prompt engineering offer alternatives without CPT, but they respectively rely on surface-level matching, external knowledge, or prompt-controlled analysis rather than changing the model’s embedding space.Ling et al.’s survey classifies these approaches as external augmentation or prompt crafting, whereas CPT+SFT is white-box model fine-tuning.

4.4 The Operational Definition of “Analyst-Bias-Free”

“Analyst-bias-free” denotes a methodology that structurally reduces analysts’ early intervention and rigid preconceptions, rather than claiming complete or statistical objectivity. The LLM proposes qualitative analyses from collective domain knowledge, while human researchers retain responsibility for final interpretation and accountability.

  • 4.4 The Operational Definition of “Analyst-Bias-Free”: The framework operationalizes “analyst-bias-free” as minimizing analysts’ prior hypotheses during early analysis, not eliminating bias completely.This extends the KJ method’s instruction to suspend subjective judgment and “let the data speak.”
  • 4.4 The Operational Definition of “Analyst-Bias-Free”: The CPT-refined model functions as a “Computational Mirror” of collective domain knowledge, mitigating premature thematic foreclosure rather than providing quantitative objectivity.This approach is framed as “Technological Reflexivity”.
  • 4.4 The Operational Definition of “Analyst-Bias-Free”: The LLM proposes qualitative analytical results as a rigorous starting point, while human expertise remains responsible for final interpretation and analytical accountability [21].The framework positions specialized-LLM rigor alongside the KJ method’s holistic integration.
  • 4.4 The Operational Definition of “Analyst-Bias-Free”: The framework excludes analysts’ rigid industry common sense and precedent while recognizing that hypothesis-driven thinking cannot keep pace with consumer diversity and change.It does not reject hypotheses outright; it rejects allowing analysts’ assumptions to constrain perception of qualitative data.
  • 4.4 The Operational Definition of “Analyst-Bias-Free”: The framework distinguishes necessary domain knowledge accumulated across projects from the preconceptions of individual analysts.This distinction addresses the apparent tension between training on skilled researchers’ curated data and excluding practitioners’ biases.

4.5 The Interdisciplinary Positioning of “Analyst-Bias-Free”

“Analyst-bias-free” means hypothesis-generating rather than hypothesis-free: Computational KJ-Hō uses abductive inference to derive insights, relationships, and strategic implications from observed patterns without pre-fixed hypotheses. The analyst primarily sets the analytical objective and evaluates or refines outputs toward implementation, while the framework combines traditions from marketing research, the KJ method, abduction, Grounded Theory, and LLMs.

  • 4.5 The Interdisciplinary Positioning of “Analyst-Bias-Free”: “Analyst-bias-free” is hypothesis-generating: it derives consumer insights, relationships, and strategic implications from observed data patterns without pre-fixed hypotheses.The analyst intervenes mainly by setting the analytical objective and evaluating or refining outputs toward implementation.
  • 4.5 The Interdisciplinary Positioning of “Analyst-Bias-Free”: The term “analyst-bias-free,” rather than “hypothesis-free,” reflects Peircean abduction, which generates probable explanations from observations.Because abduction is active inference, “hypothesis-free” would mischaracterize the framework’s explanation-generating process.
  • 4.5 The Interdisciplinary Positioning of “Analyst-Bias-Free”: Computational KJ-Hō treats the KJ method’s abductive labeling process as executable by an LLM, with “analyst-bias-free” defined as a structural design rather than a matter of degree.The KJ question—what label best explains a group of cards?—corresponds to abductive explanation generation.
  • 4.5 The Interdisciplinary Positioning of “Analyst-Bias-Free”: Computational KJ-Hō fuses marketing-research practice, the KJ method, abduction, Grounded Theory, and large language models into one interdisciplinary framework.These traditions originate across industry practice, cultural anthropology, logic, qualitative research, and artificial intelligence.

5 The Evaluation Structure

The evaluation structure combines intermediate analytical-structure assessment with final insight-quality metrics, supplemented by domain-specific benchmarking, ablation, and blinded expert evaluation. It uses InsightExtraction-F1 and MarketingQA while recognizing limitations of reference-based and automatic metrics, including the need to adjudicate potentially novel discoveries.

  • Human evaluation: Because surface similarity can miss context, strategic value, and novel insights, the framework also requires blinded human-expert evaluation of properties automatic metrics cannot capture.The study leaves the extent of human-undetected insight discovery for the empirical phase and sets no concrete novelty target.
  • Evaluation dimensions: The framework combines LOGOS’s intermediate codebook evaluation with final-output assessment through InsightExtraction-F1 and MarketingQA.LOGOS evaluates whether analytical categories are well-formed, consistent, and comprehensive, whereas the two proposed metrics assess the quality and usefulness of final consumer insights.
  • InsightExtraction-F1: InsightExtraction-F1 measures precision, recall, and their harmonic mean against consensus insights produced by skilled marketing researchers.Semantic matching uses an independent frozen general-purpose embedding model to avoid circular self-grading by the domain-specialized embeddings.
  • InsightExtraction-F1: Human adjudication prevents genuinely novel but unsupported-by-reference insights from being mechanically counted as errors.The metric retains reference-agreement limitations, so practitioners classify unmatched outputs as valid discoveries humans missed or unsupported outputs.
  • MarketingQA: MarketingQA is an original Japanese-language benchmark targeting marketing-research knowledge, with 0.75 accuracy set as a normative goal for the CPT-trained model.The pre-CPT general-purpose model serves as the improvement baseline, and the benchmark is intended for open release to support reproducibility.
  • Ablation design: The core ablation measures MarketingQA and InsightExtraction-F1 across general-purpose LLM, CPT-only, and CPT+SFT conditions.This design separates CPT’s domain-knowledge internalization from SFT’s acquisition of task-specific analytical operations; concrete values are deferred to the empirical-phase paper.

6 Discussion

The discussion frames Computational KJ-Hō as a culturally grounded, evidence-linked redesign of collaboration in marketing research rather than mere task automation. It also identifies substantial limitations, including unvalidated metrics, domain and data requirements, multimodal gaps, and the untested analyst-bias-free claim.

  • Interpretive objective: The framework’s practical objective is data-grounded, explanatory, and novel consumer insight, while the discussion cautions that novelty alone does not guarantee quality.The study leaves unresolved whether its theoretical question can be settled confidently because that resolution does not substantially affect the practical significance of consumer-insight generation.
  • Cultural and epistemological contribution: Computational KJ-Hō integrates CPT-based domain specialization with the KJ method to support culturally grounded, non-Western qualitative analysis and address the WEIRD problem.The Japanese-language implementation is presented as a step toward de-Westernizing AI epistemology, while acknowledging that non-Western perspectives are not homogeneous.
  • Human collaboration and oversight: The framework removes humans from routine coding but retains them in judgment, concentrating analyst expertise on final decisions rather than processing individual data units.This human-on-the-loop model requires process transparency to prevent blind approval and uses evidence links connecting each insight element to source utterances.
  • Integration and practical significance: Computational KJ-Hō redesigns collaborative marketing-research work by integrating previously separate components into a scalable pipeline, responding to the qualitative-research scalability crisis.The discussion situates this integration within CSCW and emphasizes the contribution of long-practice expertise, a non-Western methodology, and domain-specialized CPT.
  • Technical and domain limitations: The current system is mainly text-based, requires substantial domain data, and is not automatically transferable beyond marketing, motivating multimodal extensions and domain-specific CPT investment.Future work also targets weak-signal detection through advanced outlier methods and broader integration of non-text marketing information.
  • Validation requirements: InsightExtraction-F1 and MarketingQA remain theoretically derived metrics requiring validation against strategic usefulness and interpretive depth, while the analyst-bias-free claim requires empirical falsification.The study proposes testing the claim through embedding-space comparisons and blind evaluations of human, general-purpose LLM, and framework outputs on identical data.

7 Conclusion

Computational KJ-Hō proposes realizing the KJ method’s data-emergent epistemology at scale through a domain-specialized LLM, addressing qualitative research’s cognitive and scalability constraints. The concept paper integrates established traditions and envisions broader, more accessible insight extraction, pending validation.

  • Qualitative analysis faces a scalability constraint because analysts can hold only about 200 data points in productive analytical tension before cognitive overload shapes interpretation.
  • Computational KJ-Hō computationally realizes the KJ method’s commitment to letting meaning emerge from data rather than imposing external structure, using a domain-specialized LLM at scale.The proposed extension preserves qualitative inquiry’s epistemological attitude rather than treating computation as mere procedural automation.
  • The framework integrates the KJ method, Grounded Theory, and Peirce’s abductive reasoning around data-driven explanation generation and presents a three-layer computational architecture.
  • If validated, Computational KJ-Hō could extend beyond marketing research and democratize professional-quality consumer-insight extraction for organizations lacking large research teams.Its long-term significance depends on validation, which would enable broader access to large-scale, analyst-bias-free qualitative insight extraction.
Loading 2608.16467v1…