Three meetings couldn't decide who we were selling to, or whether our own sentiment model could survive scrutiny. What actually broke the deadlock?

Context

I was part of a five-person undergraduate subgroup inside NTNU's HR-AI Lab, working alongside daytime and in-service master's students under Dr. Hung-Yue Suen on an NSTC-funded research project. The lab split the research question three ways: how a speaker's tone, facial expression, and wording each shape a listener's emotional response. My subgroup owned the wording channel, turning a speaker's speech into a sentiment score. The underlying research data was 30 recorded university lectures, built to help instructors locate moments where student sentiment dipped, for after-the-fact teaching review.

That research framing was never in question. What put the team in a room arguing for multiple meetings was a different problem: in 2023 we decided to enter the same work into the USPACE National Youth Future Action Competition, and a classroom teaching tool is a small market to pitch a venture around. Repositioning for a bigger market (we landed on monitoring enterprise meetings) meant our sentiment score would face a tougher, more commercially-minded audience than an academic one. And the team wasn't unanimously confident that a Chinese sentiment lexicon alone could hold up under that kind of scrutiny.

Those two questions, who do we sell this to and can we actually defend the tech to them, got tangled together, and tangled is what stalled the room.

The question

The main question: with a market too small to pitch and doubts about whether our own technical approach could survive tougher scrutiny, how do you get five people who'd been arguing past each other for multiple meetings to resolve both?

Sub-questions:

1. Is the blocking issue really about the market, or about the team's confidence in the technology, or are the two entangled in a way that's stalling both?
2. If they're entangled, which question has to be resolved first for the other to become answerable?
3. What would it take to get people who each argue from a fixed instinct (one teammate reflexively finds the risk in everything, another, me, reflexively manages process and structure) to genuinely reason from a different premise, rather than just restate their starting position louder?

What would change the decision: the competition deadline meant the team had to arrive at both a customer segment and a defensible technical story within weeks, or have nothing coherent to pitch. That constraint is what turned an open-ended debate into a problem that had to resolve.

Data

• Team and scope: a five-person undergraduate subgroup (the "wording" channel of a three-way research split: tone / facial expression / wording), deciding how to reposition existing classroom-research work for a startup pitch.
• Positioning method: internal team discussion and competitor analysis; no formal user interviews or market-sizing study were conducted for either the teaching or enterprise market. "Teaching is too small a market" was the team's own judgment call, not an independently validated finding.
• Deadlock duration: roughly 2–3 team meetings stuck on the positioning question before the facilitation intervention 🟡 (recalled, not logged).
• Resolution: a single Six Thinking Hats session, after which the team converged quickly 🟡 (recalled).
• Supporting technical validation (see Approach): a weighted lexicon-plus-GPT scoring design, spot-checked qualitatively against roughly 30–50 transcript excerpts 🟡, plus a lecture-level regression against student satisfaction surveys that came back mixed and inconclusive.
• Known limits:
  • No transcript or notes exist from the Six Hats session itself. The account below is Hugo's recollection of what each reassigned hat contributed.
  • The market decision was never tested commercially: the product was never actually sold into the enterprise-meeting market it was pitched at.
  • The competition result reflects pitch/business-plan-weighted judging criteria, not an independent evaluation of the underlying technology.

Approach

Step 1: Name that two different questions were tangled together.

The team had spent 2–3 meetings debating, at the same time, which market to position the product for and whether the underlying sentiment-scoring approach could survive scrutiny from a tougher audience. Debating both simultaneously meant every market argument imported someone's unresolved doubt about the tech, and every tech argument imported someone's unstated market assumption. Nobody had said out loud that these were two separate decisions.

Step 2: Break the deadlock with a facilitation move, not another round of debate.

My role in the team's discussions was usually the same one: stepping back from the content of the argument and pushing people to trade perspectives rather than defend their own. Here, I proposed Edward de Bono's Six Thinking Hats, with a deliberate twist. Each of the five of us first named which hat matched our own natural instinct, then took the opposite one. I identified as Blue (process- and structure-focused) and deliberately argued from Green (generative, unconstrained). One consistently critical teammate, who identified as Black (risk- and flaw-focused), deliberately argued from Yellow (the optimistic case).

The point wasn't performing a framework. It was that a teammate who always finds the risk had to build the strongest optimistic case for once, and I had to argue for a possibility rather than manage the process around it. That's what got people reasoning from a premise they didn't already hold, instead of restating the one they walked in with.

Step 3: The market call came out of that session, and it came first.

We deliberately rejected the teaching-context positioning (the original research framing, and the technically safer story to defend) because the team judged, through internal discussion and competitor analysis rather than formal market research, that the education market was too small to pitch as a venture. We repositioned around enterprise meetings instead: monitoring a live or recorded meeting for emotionally significant segments, then surfacing moderation suggestions to the meeting's facilitator and wording-correction suggestions to other participants.

Step 4: That market decision is what forced the technical decision, not the reverse.

Pitching to a business-oriented competition audience meant our sentiment score would get a harder question than an academic audience would ask: why should anyone trust this number? Once the positioning was set, that question wasn't optional anymore. The team needed an honestly defensible answer, and "a Chinese sentiment lexicon" alone wasn't going to survive it, since a word-level lexicon is structurally blind to negation, double negation, metaphor, and sarcasm. That requirement is what led directly to weighting the lexicon score more heavily for interpretability while layering in a prompt-engineered GPT-4 adjustment for exactly the context-dependent cases the lexicon couldn't resolve: a design built to answer a scrutiny question the market call had just created, not a fix for a bug discovered afterward.

We validated that design decision, not a trained model, the only way the timeline allowed: a qualitative spot-check of roughly 30–50 transcript excerpts chosen to represent the failure categories the hybrid was meant to handle, plus a coarser lecture-level regression against student satisfaction surveys that came back mixed. Both were honest, bounded checks, not a claim to an accuracy figure we never had the data to earn.

What happened after. When the same work was later shown at NTNU's own internal exhibition, the team reversed course and pitched the teaching angle instead, because a teacher-training university was the stronger audience fit for that framing. The underlying product was genuinely adjusted for both contexts, not just relabeled. The same instinct that resolved the original deadlock (figure out who's actually in the room before deciding how to talk to them) showed up again, on its own, for a different audience.

The lab kept going past that point, too. Dr. Suen later carried the same lexicon-plus-GPT approach forward into a full product, Echo, an emotional-expression AI assistant that launched publicly in June 2024. 🟡 (I'd already left the team by then; reported to me secondhand.)

Findings

The Six Hats swap resolved a 2–3-meeting deadlock in a single session. The clearest signal was the critical teammate's forced turn at the optimistic case, which is what ended up carrying the market decision, alongside my own forced turn at generative thinking, which helped move the room off the safer, more familiar teaching framing.

The market decision reached that way (enterprise meeting monitoring) went to competition and won 1st of 18 teams at the 2023 USPACE National Youth Future Action Competition. That result reflects judging criteria weighted toward business plan and pitch quality, with the technical build serving as supporting evidence. It's a signal that the positioning and packaging worked, not an independent measurement of the sentiment model's accuracy.

The technical design that resulted (the lexicon-weighted, GPT-adjusted hybrid) held up under the qualitative spot-check it was built for, but that check was a small, deliberately non-random sample, not a counted pass rate, and the lecture-level regression against student surveys stayed inconclusive. Both are honest limitations of a validation built to answer a specific scrutiny question, not a broader claim about model performance.

Recommendation

When a market decision and a team's confidence in its own technical work are tangled together, untangle them before trying to resolve either. Arguing both at once is what stalled the room for 2–3 meetings; separating "who do we sell to" from "can we defend the tech" is what let each get answered.

A structured device that forces people out of their default reasoning style is worth the setup time. Six Thinking Hats, with a deliberate swap to each person's opposite hat, produced a genuinely new argument from the teammate most likely to only restate a fixed position. That's a different outcome than another open debate would have produced.

Match the pitch to the audience in the room, not to the product itself. The same underlying work was honestly repositioned twice (enterprise for a venture competition, education for a teacher-training university's own exhibition) because each audience needed a different case made to it, not because the product changed to fit either one.

Key learning

My most useful contribution in a stuck room wasn't a better argument. It was a structure that changed what people were arguing from. The Six Hats swap worked because it moved a habitually critical teammate into building the optimistic case and moved me into generative thinking instead of managing the process around it. Neither of us won the argument; the argument changed shape.

A market decision can force a technical decision, just as easily as the reverse. We didn't choose the lexicon-plus-GPT hybrid because it tested best in isolation. We chose it because the positioning call had already decided we'd need to defend our score's credibility to a tougher audience, and that requirement came out of a meeting about market fit, not a lab result.

The right packaging depends on who's in the room, not on the product. The same work was legitimately pitched two different ways to two different audiences (a competition panel and a teacher-training university's own exhibition) because recognizing what each audience actually needed to believe isn't spin. It's the same instinct as finding a stakeholder's real incentive before asking them for anything.