Sentiment Analysis
Sentiment analysis uses AI to read reviews, surveys and comments and label the feeling in them as positive, negative or neutral, often per topic.
| Sentiment Analysis at a glance | |
|---|---|
| Also called | Opinion mining, review analysis, emotion AI (for emotion detection) |
| Field | Natural language processing (NLP), a branch of artificial intelligence |
| Input | Reviews, survey comments, social media posts, chats, emails, call transcripts |
| Output | Polarity (positive / negative / neutral), a score, and often topics and emotions |
| Main types | Document-level, sentence-level, aspect-based, emotion detection |
| Methods | Lexicon (rule-based), machine learning, deep learning and large language models |
| Key early research | Pang, Lee & Vaithyanathan (2002); Turney (2002); Hu & Liu (2004) |
| Common summary metric | Net sentiment = (% positive − % negative) |
| Best for | Finding patterns in large volumes of free-text feedback |
Key takeaways
- Sentiment analysis turns thousands of guest comments into data you can count, compare and track over time.
- Aspect-based sentiment analysis scores each topic in a comment separately, such as food, service, wait time and price, which is what makes it so useful for running restaurants.
- Modern AI models handle mixed opinions, slang and many languages far better than older keyword lists, although they still make occasional mistakes.
- The most valuable output is a trend by topic and location, and reading a sample of the underlying comments keeps your interpretation grounded.
- How often a topic comes up matters as much as how negative it is. A mild complaint mentioned by half your guests usually deserves more attention than a rare, angry one.
What is Sentiment Analysis?
Sentiment analysis, also known as opinion mining, uses natural language processing (NLP) to understand the feeling behind a piece of text. It reads a review, survey answer or message and determines whether the writer's view is positive, negative or neutral.
More advanced systems go a step further. They identify what the opinion is about, such as food, service or price, and how strongly the guest feels about it.
For your restaurant, the text is mostly guest feedback: reviews on Google, Yelp, TripAdvisor and delivery apps, comments from your NPS and CSAT surveys, social media posts and messages to your support team.
Sentiment analysis turns all of that unstructured text into data, which means you can count it, compare it between locations and follow it over time. That makes it a core part of restaurant intelligence.
How sentiment analysis works
A typical sentiment analysis process for guest feedback follows six stages:
- Collection. Text from every source (review sites, delivery apps, guest surveys, social media and messages) is gathered in one place, such as a unified inbox, together with the location, platform, date and rating.
- Preparation. The system detects the language, removes duplicates and spam, and splits long comments into sentences.
- Sentiment detection. Each comment, sentence or topic receives a label (positive, negative, neutral and sometimes mixed), usually with a score.
- Topic identification. The system recognizes the topics each comment mentions, such as food quality, staff, wait time, cleanliness, price or delivery, and attaches a sentiment to each.
- Aggregation. Results are combined by topic, location, channel and period, and summary measures such as net sentiment are calculated, so analytics dashboards can show them by location.
- Action. Dashboards, alerts for sudden changes and prioritized lists of comments help your team respond to feedback where it matters most and log incidents that need follow-up.
Types of sentiment analysis
| Type | What it does | Example output |
|---|---|---|
| Document-level | One label for a whole review or comment | "Negative" |
| Sentence-level | One label per sentence | Sentence 1 positive, sentence 2 negative |
| Aspect-based (ABSA) | A label for each topic mentioned | Food: positive, wait time: negative |
| Emotion detection | Identifies specific emotions | Frustration, delight, disappointment |
| Intent detection | Identifies what the writer wants or plans | Complaint, refund request, won't return |
| Fine-grained | A graded scale with more than three labels | Very negative … very positive (e.g. 1–5) |
For restaurant operations, aspect-based analysis is usually the most valuable type, because it connects how guests feel to the part of your business responsible for it. That connection is what turns feedback into operational excellence across locations.
Sentiment analysis example
Consider this review: "The ramen was incredible and the staff were lovely, but we waited 45 minutes for a table even though we had a booking."
| Method | Result | What you learn |
|---|---|---|
| Star rating only | 3 stars | Something was partly wrong |
| Document-level sentiment | Mixed / slightly positive | Little additional detail |
| Aspect-based sentiment | Food: positive · Staff: positive · Wait time: negative · Reservations: negative | Your kitchen and team are performing well, and the booking and seating process needs attention |
Repeated across hundreds of reviews, this breakdown shows which part of the guest experience drives your ratings at each location. Momos operational insights are built around this kind of topic-level view.
Sentiment analysis methods
| Method | How it works | Strengths | Limitations |
|---|---|---|---|
| Lexicon-based (rule-based) | Scores words from a dictionary of positive and negative terms, with rules for negation and intensity (e.g. VADER, 2014) | Simple, fast and transparent, with no training data required | Misses context, sarcasm and restaurant-specific meanings |
| Classical machine learning | Models such as Naive Bayes, logistic regression or support vector machines trained on labeled examples | Learns from real data and runs efficiently | Requires labeled data and struggles with new subjects and long sentences |
| Deep learning | Neural networks that learn the meaning of words and phrases | Handles context and word order better | Requires more data and computing power, and is harder to explain |
| Transformer models | Pre-trained models such as BERT, fine-tuned for sentiment | High accuracy, with good handling of context and negation | Fine-tuning and maintenance take effort |
| Large language models (LLMs) | General-purpose models asked to label sentiment, topics and emotions | Handle nuance, mixed opinions and many languages, discover topics without a fixed list, and can summarize | Need consistency checks, and costs grow with volume |
Many feedback platforms now combine several of these methods. Momos's insights agent, Rockie, applies AI to guest feedback.
Sentiment analysis topics in restaurant feedback
| Topic | Typical positive mentions | Typical negative mentions |
|---|---|---|
| Food quality | Fresh, tasty, well seasoned, generous | Cold, bland, overcooked, small portions |
| Service | Friendly, attentive, knowledgeable | Rude, slow, ignored, forgetful |
| Wait time | Quick, seated straight away | Long wait, slow food, late delivery |
| Order accuracy | Got exactly what we ordered | Missing items, wrong order |
| Cleanliness | Spotless, clean restrooms | Dirty tables, sticky floors |
| Atmosphere | Cozy, lively, good music | Too loud, cramped, cold |
| Value for money | Good value, fair prices | Overpriced, small for the price |
| Delivery and packaging | Arrived hot, well packed | Spilled, soggy, cold on arrival |
| Booking and seating | Easy to book, table ready | Booking lost, kept waiting despite a reservation |
| Digital experience | Easy to order online | App crashed, payment failed |
Sorting comments into topics like these is the starting point for discovering customer insights you can act on.
Sentiment analysis calculator
This calculator works out net sentiment, a common way to summarize sentiment for a topic, location or period: the percentage of positive mentions minus the percentage of negative ones. Like NPS, it ranges from -100 to +100.
Tools define net sentiment in slightly different ways, so comparisons are most reliable when every figure is calculated with the same method.
Some services report sentiment on a continuous scale instead. Google Cloud Natural Language, for example, gives each text a score from -1.0 to 1.0 and a separate magnitude that reflects the overall strength of emotion.
Net sentiment is most revealing when you calculate it per topic. Your brand might score +60 on food and −15 on wait time, which tells you far more than a single overall number would.
How restaurants use sentiment analysis
- Early warning. A rise in negative mentions of wait time at one location becomes visible weeks before the location's star rating begins to fall, an idea explored in predictive CX.
- Prioritized replies. The most negative or emotional comments can be surfaced first, so they receive a fast, personal response as part of your online reputation management, and an AI review manager can draft those replies.
- Location comparisons. A per-location view, such as an operator report, shows which locations excel at service while struggling with food, or the reverse, and share what your strongest locations do well.
- Measuring change. Sentiment shows whether a new menu, recipe, staffing change or delivery partner actually changed how guests talk about you, which complements restaurant data analytics.
- Explaining scores. A fall in NPS or CSAT can be traced to the topics guests mention most often in their comments.
- Menu insight. Tracking how often each dish is mentioned, and how guests feel about it, informs menu decisions and menu engineering.
- Guest recovery. Strongly negative comments flag guests worth contacting, and a customer recovery workflow helps win them back.
- Incident tracking. Recurring complaints about hygiene, safety or order errors can be logged as incidents. The Mr. Pickle's case study describes a brand that cut incidents by 46%.
- Competitor benchmarking. Public reviews of nearby competitors reveal where guests see them as stronger or weaker than you.
Accuracy and limitations of sentiment analysis
Sentiment analysis is never perfect, partly because people also disagree about the sentiment of many comments. These are the most common challenges:
- Sarcasm and irony. "Great, another 40-minute wait" uses a positive word to express a negative experience.
- Negation and contrast. A phrase such as "Not bad at all" expresses approval. A sentence that begins "The food was good, but…" usually leads into a complaint.
- Restaurant-specific meanings. "Hot" is a compliment for soup and a complaint about a dining room, and "spicy" can be either praise or criticism.
- Very short text. Comments such as "Ok", "fine" or a single emoji carry little reliable signal.
- Mixed opinions. Many reviews praise some aspects and criticize others, which a single label can hide.
- Language and dialect. Slang, local expressions and comments that mix languages are harder for some systems.
- Comparisons. A phrase such as "Better than the other branch" needs context to interpret.
- Fake and spam reviews. These distort results unless they are filtered out first.
How to evaluate a sentiment analysis tool
Evaluating a tool on your own data gives you a far clearer picture than any vendor demonstration.
- A test on your own comments. Labeling a few hundred real comments yourself and comparing them with the tool's results shows how well it understands your guests.
- Topic accuracy. For restaurants, correctly identifying topics such as "wait time" or "order accuracy" matters more than the overall label.
- Every language you operate in. Accuracy can differ between languages, so each market deserves a quick check.
- The nature of the mistakes. Errors on sarcasm or mixed reviews are expected, while errors on clear complaints are a warning sign.
- Traceability. Your managers should be able to click through from any chart to the original comments. Knowing how an AI agent sees your data also helps your team trust its conclusions.
- Consistency over time. The same comment should receive the same result next month, otherwise your trends become unreliable.
Best sentiment analysis tools for restaurants
For a restaurant brand, the most useful sentiment analysis tool is one that reads feedback from all your channels, understands restaurant-specific topics, and reports by location. It also helps when the analysis sits next to the tools your team uses to reply to guests.
| Tool | Best known for |
|---|---|
 MomosBest pick | Guest AI for multi-location restaurants. Reviews, surveys and messages are categorized by topic and sentiment in one Unified Inbox, with AI-drafted replies and recovery offers built in. |
 Medallia | Enterprise experience management with text analytics across surveys, reviews and contact center feedback |
 Qualtrics | Experience management platform whose Text iQ analytics categorizes open-text survey feedback |
 Chatmeter | Local brand reputation management with review and sentiment analytics for multi-location brands |
 ReviewTrackers | Review monitoring and analytics across review sites, now part of InMoment |
The strongest choice depends on your channels and the volume of feedback you receive.
Testing two or three tools on a sample of your own reviews is a worthwhile step before deciding.
Common sentiment analysis mistakes
- Treating the score as final. Sentiment scores show you where to look, and the comments themselves provide the full picture, which is how teams move from dashboard fatigue to clarity.
- Measuring only overall sentiment. Without topics, the analysis rarely shows which part of the business needs attention.
- Overlooking volume. How often a topic is mentioned matters as much as how negative it is.
- Mixing sources without labels. Delivery app reviews, Google reviews and survey comments come from different guests in different situations, so keeping them separable improves interpretation.
- Skipping spam and fake-review filtering. Unfiltered fake reviews can distort results for a location.
History of sentiment analysis
Interest in detecting opinions automatically grew rapidly in the early 2000s, as online reviews created large collections of opinionated writing for researchers to study.
Two papers from 2002 are often seen as the beginning of the field. Bo Pang, Lillian Lee and Shivakumar Vaithyanathan classified movie reviews with machine learning in "Thumbs up? Sentiment Classification using Machine Learning Techniques". Peter Turney classified reviews based on the meaning of their words and phrases in "Thumbs Up or Thumbs Down?".
In 2004, Minqing Hu and Bing Liu published "Mining and Summarizing Customer Reviews". Their method identified the product features customers mentioned and the opinion expressed about each one, which laid the groundwork for aspect-based analysis.
In 2012, Bing Liu published the book Sentiment Analysis and Opinion Mining, which gathered document-level classification, aspect-based analysis, sentiment lexicons and opinion spam detection into one reference for the field.
Restaurant reviews later became one of the field's standard test sets. The SemEval research competitions from 2014 to 2016 included aspect-based sentiment tasks built on restaurant reviews, starting with SemEval-2014 Task 4, with aspects such as food, service, price and ambience.
Methods have since moved from word lists to statistical machine learning, then to deep learning, and from 2018 onwards to transformer models such as BERT. Today's large language models (LLMs) can analyze sentiment and topics with little or no task-specific training.
Frequently asked questions
What is aspect-based sentiment analysis?
Aspect-based sentiment analysis (ABSA) identifies each topic mentioned in a piece of text and scores it separately. A comment such as "Great pasta, but we waited 40 minutes" becomes positive on food and negative on wait time, which is far more useful than a single "mixed" label.
How accurate is sentiment analysis?
Accuracy depends on the method, the subject matter and the language. Even people frequently disagree about the sentiment of a comment, so no system is perfect. Modern models handle most everyday reviews well, while sarcasm, very short comments and slang remain harder. Testing any tool on a sample of your own reviews is the most reliable way to judge it.
Why would a restaurant use sentiment analysis?
A multi-location brand can receive thousands of reviews, survey comments and messages each month, far more than any team can read. Sentiment analysis sorts them automatically, so you can see which issues are growing, where, and since when. The Condado Tacos case study shows how one brand built chain-wide guest intelligence from its feedback.
Isn't a star rating enough?
A star rating shows how a guest felt overall, and the text explains the reasons. Two 3-star reviews can describe completely different problems, and a 5-star review can still mention a complaint. Sentiment analysis reads the text so you can understand what is driving the rating.
What is a net sentiment score?
Net sentiment is a common summary score, calculated as the percentage of positive mentions minus the percentage of negative ones, giving a value from -100 to +100. It is usually calculated per topic, location or period. Tools define it slightly differently, so it is worth checking how yours works.
Can sentiment analysis detect emotions like anger?
Yes. Emotion detection goes beyond positive and negative to identify feelings such as anger, frustration or delight. It helps you prioritize replies, since an angry guest usually benefits from a faster response than a mildly disappointed one. A practical guide on how to respond to customer feedback covers what to say.
Does sentiment analysis work in languages other than English?
Modern multilingual models work across many languages, including comments that mix two languages. Accuracy can still vary between languages, so reviewing results in each market you operate in is a sensible precaution.