Sentiment analysis dashboard
VerdeVita Cosmetics S.r.l., a cosmetics company, 85 employees, direct and e-commerce channels. 534 pieces of feedback gathered through AI interviews run with Tacita, average duration 3–5 minutes per interview. Analysis period: January–June 2025. Hybrid pipeline with a transformer classifier and aspect-based sentiment analysis.
Key metrics
Positive rate
54%
289 interviews
Negative rate
22%
117 interviews
Sarcasm detected
4,9%
26 of 534 interviews
Average subjectivity
0,56
scale 0 (factual) – 1 (opinion)
Monthly shift in the polarity of feedback gathered via Tacita
Histogram of the scores assigned by the pipeline (0–100)
The concentration in the 55–80 band indicates overall positive sentiment. The cluster below 25 accounts for 7.8% of the interviews.
Net promoter score derived from interview sentiment
Number of AI interviews completed through Tacita
Aspect-based sentiment analysis
The ABSA pipeline breaks each interview down into its constituent aspects and assigns an independent sentiment to each one. The model identifies 7 primary aspects for VerdeVita. Each interview touches on an average of 1.8 distinct aspects, for a total of 961 aspect mentions across the 534 interviews.
Score by aspect — click to expand
X axis = sentiment score. Y axis = number of mentions. Size = positive rate.
The aspects in the upper-left quadrant (low sentiment, high volume) require priority action. Price and returns are the two most critical areas.
Action priority
Key insight
68% of the interviews with a negative score contain at least two distinct emotions. Anger rarely appears on its own: in 64% of cases it comes with disappointment or frustration at an expectation that wasn't met.
Automated analysis — Plutchik emotion model, 534 interviews processed
Emotional and pragmatic analysis
The sentiment score captures polarity. Emotional analysis goes further, identifying specific emotions according to Plutchik's 8-dimension model. The sarcasm detection module corrects false positives. Subjectivity analysis distinguishes factual answers from personal opinions.
Intensity detected vs cosmetics sector benchmark
Transformer classifier fine-tuned on an Italian dataset
Interviews containing sarcasm
26
of 534 in total (4.9%)
Classifier accuracy
F1 = 0,84
on a balanced internal test set
False positive rate
7,1%
neutral answers wrongly classified as sarcastic
Sarcasm inverts the polarity of an answer. The classifier detects typical lexical patterns such as hyperbole, apparent antithesis and incongruous superlatives. Sarcastic answers receive an automatic correction to their sentiment score.
Distribution of interviews by degree of subjectivity
Average subjectivity
0,56
scale 0 (factual) – 1 (opinion)
Factual answers (low subjectivity) mainly contain reports of defects or technical problems. Highly subjective ones express personal preferences. Telling the two categories apart improves the prioritization of operational action.
Segments and time analysis
The 534 AI interviews with Tacita involved customers from different segments. Cross-analysis between segment and sentiment makes it possible to identify the groups with the most perceived problems.
Interview volume, average sentiment and distribution by customer type
| Segment | Interviews | Sentiment | Average duration |
|---|---|---|---|
| Repeat customers | 186 | 68% | 4.6 min |
| First purchase | 142 | 55% | 3.8 min |
| E-commerce customers | 118 | 48% | 4.1 min |
| Retail customers | 88 | 72% | 3.4 min |
Retail customers express the most positive sentiment (72%), probably linked to direct experience with the product. E-commerce customers show the most critical sentiment (48%), with complaints concentrated on shipping and returns.
Day of the week on which the interviews were run. Hover for details.
Comparison of the positive/neutral/negative split for each segment
Classified answers
Each interview goes through 4 phases after transcription. Tokenization and preprocessing. Sentiment classification with a transformer model. Aspect extraction with ABSA. Sarcasm detection and subjectivity calculation. The result is a composite score that brings all the dimensions together.
Process
After the AI interview run with Tacita is transcribed (average duration 3–5 minutes), the text enters the analysis pipeline. The 4 phases are executed in sequence on each transcript.
01
Preprocessing
~0.2s per transcriptTokenization, normalization, removal of conversational noise (interjections, false starts, repetitions). Handling of slang, abbreviations and Italian dialect variants gathered during the interviews.
02
Sentiment classification
~0.8s per transcriptTransformer model (fine-tuned RoBERTa-IT) with probabilistic output across 5 classes. Ensemble with weighted voting across 3 independent models to reduce the bias of any single classifier.
03
ABSA and emotional analysis
~1.2s per transcriptAspect extraction with sequence labeling. Assignment of independent sentiment to each identified aspect. Emotion detection according to Plutchik's 8 primary dimensions.
04
Post-processing
~0.3s per transcriptSarcasm detection with a dedicated classifier (F1 = 0.84). Subjectivity calculation. Automatic score correction for sarcastic answers. Generation of the final composite score.
Frequently asked questions
The 534 pieces of feedback were gathered through conversational interviews run by a Tacita AI agent. Each interview lasts an average of 3–5 minutes and follows a semi-structured flow that leaves the customer room to express their views naturally, with no format constraints.
The score is a weighted average of the results of three transformer models run in parallel. Each model produces a probability distribution across 5 classes (very negative, negative, neutral, positive, very positive). The final score is normalized on a 0–100 scale.
Traditional analysis assigns a single score to the whole interview. Aspect-based sentiment analysis breaks the transcript down into its constituent aspects and assigns a separate sentiment to each. An answer such as "the quality is excellent but the price is too high" produces two independent scores.
The classifier is a transformer model fine-tuned on a dataset of Italian texts annotated for sarcasm. It detects typical lexical patterns such as hyperbole, apparent antithesis and superlatives incongruous with the context. When an answer is classified as sarcastic, the sentiment score is inverted automatically.
The index measures how far an answer expresses a personal opinion (high value) rather than a factual observation (low value). Factual answers often contain reports of bugs, defects or logistics problems. Highly subjective ones express preferences. The distinction helps separate technical problems from perceptions.
Data handling and privacy. Tacita interview transcripts are anonymized before analysis. No personal data is retained in the training dataset. The infrastructure is GDPR compliant (art. 6.1.f — legitimate interest in improving the service). Anonymized transcripts are kept for 12 months and then deleted.
Tacita's AI interviews gather structured feedback in 3–5 minutes. The pipeline adapts to any sector.
Talk to the team