Apple launched the iPhone Duo — their first foldable — and told us what mattered: the A20 chip, Apple Intelligence, the hinge design. I wanted to know what 5,104 consumers actually talked about. Aspect-Based Sentiment Analysis across 12 product dimensions, two ML benchmarks (one automated, one human-validated), and 19 topic clusters discovered without supervision. The gap between F1=0.959 (auto-label) and F1=0.384 (human-validated) is itself the most important finding.
Apple's marketing pushed three things for the iPhone Duo: the A20 chip performance, the Apple Intelligence AI features, and the foldable design engineering. I wanted to know whether consumers were talking about the same things — and whether what they said was positive, negative, or somewhere more nuanced.
The core method is Aspect-Based Sentiment Analysis (ABSA) — rather than measuring overall sentiment ("this review is positive"), ABSA decomposes each comment into individual product aspects and classifies how the consumer feels about each one separately. A comment like "the foldable design is stunning but there's no way I'm paying $2,000" becomes: Foldable Design → Positive, Price → Negative. That's a fundamentally more useful signal than a blended sentiment score.
The launch window was September 9–13, 2026 (iPhone Duo launch week). 34+ YouTube videos, 5,104 raw comments, 4,470 after a cleaning pipeline that removed spam, near-duplicates, non-English comments, and anything under 4 words.
The gap between Apple's emphasis and consumer emphasis was real and specific:
| ASPECT | MENTIONS | SIGNAL | NOTE |
|---|---|---|---|
| AI / Siri | 1,000 | 🟡 Mixed | Most discussed, least resolved — skepticism dominated |
| Foldable Design | 920 | 🟡 Mixed | #1 topic for buyers, #1 concern for doubters |
| Camera | 693 | 🟢 Positive | Consistently praised across both phones |
| Price | 578 | 🟡 Mixed | Debated but not categorically rejected |
| Durability | 100 | 🔴 Negative | Only aspect with statistically negative bootstrap CI |
| Ecosystem | 219 | 🟢 Positive | Strong loyalty signal among existing Apple users |
Apple pushed: A20 chip, Apple Intelligence, Foldable engineering. Consumers discussed: AI/Siri (skeptically), Price (debating value), Durability (worried about hinges). The alignment between marketing emphasis and consumer attention was partial at best — and the durability concern, with the lowest mention count of any negative aspect, was also the clearest negative statistical signal in the entire dataset.
Two separate benchmarks were run. The number that matters is the second one.
| BENCHMARK | MODEL | MACRO F1 | NOTE |
|---|---|---|---|
| ① Auto-label | Majority Baseline | 0.396 | Reference |
| TF-IDF + LogReg | 0.959 | Reproduces pipeline labels well | |
| Sentence Emb. + LR | 0.729 | — | |
| ② Human-validated ← primary | Majority Baseline | 0.190 | 175 hand-labelled · 5-fold CV |
| TF-IDF + LogReg | 0.384 | Cohen's κ = 0.184 · beats baseline by +0.194 | |
| TF-IDF + Random Forest | 0.322 | Cohen's κ = 0.076 |
The gap between 0.959 and 0.384 is not a failure — it's a finding. YouTube comment sentiment involves heavy sarcasm, negation, code-switching, and implicit opinion that the automated pipeline is not designed to handle. The auto-label benchmark measures how consistently the pipeline assigns its own labels. The human benchmark measures how well the pipeline captures what a person actually means. F1 = 0.384 on genuine human judgment, against a baseline of 0.190, on one of the hardest NLP tasks, is honest progress. Labels are single-annotator; inter-annotator agreement was not measured — documented limitation.
All 4,470 comments were encoded with all-MiniLM-L6-v2 sentence embeddings (384 dimensions), reduced with UMAP (n_neighbors=15, cosine metric), and clustered with HDBSCAN (min_cluster_size=40). No cluster count was pre-specified — the algorithm found the structure. 19 interpretable clusters emerged, covering discussion themes from hinge durability concerns to camera comparisons to upgrade timing debates.
Regex patterns across 4 intent classes (Buy / Considering / Not buying / Conditional) were applied to all comments. Rare signal, limited statistical power — reported accurately rather than emphasised.
1,000-resample bootstrap confidence intervals on aspect sentiment means, plus chi-square tests with Cramér's V effect sizes comparing iPhone Duo vs iPhone 18 Pro sentiment distributions.
Two transformer sentiment models agreed on only ~43% of YouTube comments. Treated as a genuine finding about task difficulty, not averaged away or discarded.
Only people who chose to comment — not representative of all iPhone Duo buyers or observers. Early adopters and tech enthusiasts are overrepresented.
"This will bend in my pocket" → Durability concern, but the word "durability" doesn't appear. Keyword matching sets a precision floor at the cost of recall.
Sentiment after reviews, teardowns, and real-world durability reports may shift materially. This is a launch-week snapshot, not a stable consumer verdict.
The 175 human-validated labels were labelled by one person. Inter-annotator agreement (Cohen's κ between two independent annotators) was not measured.