AI Text Classifier

About AI Text Classifier

How Does Browser-Based Sentiment Analysis Work?

Traditional sentiment analysis requires sending text to a cloud API, which raises privacy concerns and costs money per request. This tool takes a different approach by running the entire model locally. On first use, the browser downloads approximately 67 MB of model weights from a CDN and caches them. Every analysis after that runs entirely on your CPU through WebAssembly, with no network requests.

Step	What Happens
1. Model loading	On first use, ~67 MB of DistilBERT weights download and cache in your browser via ONNX Runtime
2. Tokenization	Your text is split into subword tokens using the BERT WordPiece tokenizer (vocabulary of 30,522 tokens)
3. Encoding	Tokens are converted to numerical IDs with positional embeddings added
4. Inference	The model processes the token sequence through 6 transformer encoder layers via WebAssembly
5. Classification	A linear classification head maps the [CLS] token output to two logits (positive and negative)
6. Softmax	Logits are converted to probabilities that sum to 100%
7. Results	The label with the higher probability is shown alongside confidence bars

The inference engine is ONNX Runtime compiled to WebAssembly, which runs at near-native speed in modern browsers. Transformers.js v3 added WebGPU support as well, though this tool uses the WASM backend for maximum browser compatibility. A single sentence typically classifies in under 200 milliseconds once the model is loaded.

What Is DistilBERT?

DistilBERT was introduced in 2019 by Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf at Hugging Face. The paper, presented at the EMC2 workshop at NeurIPS 2019, demonstrated that knowledge distillation could compress BERT into a model with 40% fewer parameters while retaining 97% of its language understanding ability. The distilled model runs 60% faster than BERT-base on comparable hardware.

Knowledge distillation works by training a smaller "student" model to mimic the outputs of a larger "teacher" model. The DistilBERT authors used a triple loss function combining language modelling loss, distillation loss (soft target probabilities from the teacher), and a cosine distance loss between hidden state representations. This approach preserved far more of BERT's capability than simply training a smaller model from scratch.

Property	DistilBERT	BERT-base
Parameters	66 million	110 million
Transformer layers	6	12
Hidden size	768	768
SST-2 accuracy	91.3%	92.7%
Inference speed	~60% faster	Baseline
Model file size	~67 MB (ONNX)	~440 MB

The version used here was fine-tuned on SST-2, a widely used English sentiment benchmark. SST-2 is a binary classification subset of the Stanford Sentiment Treebank, originally published by Socher et al. in 2013. The dataset contains 67,349 phrases extracted from 11,855 movie review sentences sourced from Rotten Tomatoes, each labelled as positive or negative by human annotators.

Understanding Confidence Scores

The model outputs a probability distribution across two classes. A result of "92% positive" means the model assigns a 0.92 probability to the positive class and 0.08 to negative. These probabilities always sum to 1.0 (100%). Higher confidence generally means the text contains clearer sentiment signals.

Confidence Range	What It Means	Example Text
95-100%	Strong, unambiguous sentiment - the model is very certain	"This product is absolutely amazing, best purchase I've ever made"
80-95%	Clear sentiment with minor ambiguity	"Really good experience overall, would recommend"
65-80%	Moderate confidence - text may contain mixed signals	"It was okay, nothing special but not bad either"
50-65%	Low confidence - the model is nearly guessing. Treat results with caution	"Great food but awful service and a long wait"

Scores below 65% often indicate genuinely mixed sentiment rather than a model failure. A review that praises a restaurant's food but criticises the service contains real contradictory signals. For these cases, the raw confidence percentages are more informative than the positive/negative label alone.

Batch Analysis Mode

For analysing multiple texts at once, put each text on a separate line. The tool runs inference on each line individually and displays all results together with colour-coded labels. This is practical for scanning a list of product reviews, survey responses, or social media comments without copying them one at a time.

Batch Size	Typical Processing Time	Best For
1-10 texts	1-3 seconds	Quick spot checks on a handful of reviews or comments
10-50 texts	3-15 seconds	Scanning a page of survey responses or support tickets
50-100 texts	15-45 seconds	Bulk analysis of an export from a review platform

Processing time depends on text length and your device's CPU speed. Shorter sentences classify faster because the tokenizer produces fewer tokens for the model to process. The 512-token input limit is a BERT architecture constraint - texts longer than roughly 400 words get truncated before classification.

Practical Use Cases

The global sentiment analytics market was valued at roughly $5.7 billion in 2025, according to Precedence Research, reflecting how central opinion mining has become across industries. While enterprise teams typically use API-based services, this browser tool provides a quick, private way to run the same underlying NLP technique on smaller datasets.

Use Case	How It Helps
Customer review analysis	Sort reviews by sentiment before reading them - spot negative outliers in a batch of mostly positive feedback
Social media monitoring	Paste mentions or comments about a brand and check the overall tone without a paid tool
Support ticket triage	Identify angry or frustrated messages that need priority response
Survey response screening	Flag negative free-text responses for follow-up before reading every answer
Content tone checking	Verify that marketing copy, emails, or announcements read as positive rather than inadvertently negative
NLP education	See a real transformer model classify text without writing code or setting up a Python environment

For understanding how tokenizers break text into subwords before classification, the AI token counter shows exact token counts and boundaries. To estimate API costs if you need sentiment analysis at scale, the AI pricing calculator compares per-token pricing across providers. And for checking overall text quality alongside sentiment, the readability score tool measures reading level and complexity.

Known Limitations

No model is perfect, and DistilBERT has well-documented blind spots. Understanding these helps interpret results more accurately.

Limitation	Details
English only	Trained exclusively on English text. Other languages produce unreliable results. A multilingual model like XLM-RoBERTa would be needed for non-English text.
Binary classification	Only positive vs negative. There is no neutral class, no emotion detection (anger, joy, sadness), and no intensity scale beyond the confidence percentage.
Short text bias	SST-2 training data consists of single movie review sentences. Very long texts (paragraphs or full articles) get truncated at 512 tokens, so only the opening portion influences the result.
Sarcasm and irony	Sarcastic statements like "Oh great, another meeting" tend to be classified as positive because the surface-level words are positive. Detecting sarcasm requires pragmatic reasoning that this model architecture does not support.
Domain-specific language	Medical, legal, or highly technical text uses vocabulary that may not appear in the model's training data. Clinical terms, legal jargon, or code-related text may produce low-confidence or incorrect results.
Negation handling	Complex negation chains ("I wouldn't say it wasn't good") can sometimes confuse the model. Simple negations ("not good") are generally handled correctly.

Tips for Better Results

The model was trained on single sentences from movie reviews, so input that resembles short opinion statements tends to classify most accurately. A few practical tips can improve the quality of results.

Keep inputs to one idea per line. A sentence like "The camera is excellent but the battery life is terrible" contains both positive and negative signals, and the model has to pick one label for the whole input. Splitting it into two lines ("The camera is excellent" and "The battery life is terrible") produces more accurate per-topic results.

Remove boilerplate and greetings. Text like "Hi team, hope you're well" at the start of an email adds neutral noise that dilutes the sentiment signal from the actual message content. Trim down to the opinion-bearing sentences before analysing.

Watch for domain mismatch. The model was trained on movie reviews, which use colloquial English. Formal business language, medical terminology, or legal text may produce low-confidence scores not because the sentiment is ambiguous, but because the vocabulary is unfamiliar to the model. Scores below 65% on professional text should be treated as "uncertain" rather than taken at face value.

Use batch mode for comparison. When analysing a set of customer reviews, the relative ranking of confidence scores is often more useful than the absolute values. A review scored at 88% positive is probably more positive than one scored at 62% positive, even if both carry the same "POSITIVE" label.

How Does This Compare to API-Based Sentiment Analysis?

Cloud APIs from providers like Google Cloud Natural Language, AWS Comprehend, and Azure Text Analytics offer more advanced features - multi-language support, entity-level sentiment, and fine-grained emotion categories. They also charge per request, which adds up when processing thousands of texts. This browser tool trades those advanced features for complete privacy and zero cost. No API key is needed, no text leaves the browser, and there are no usage limits.

Feature	This Tool (Browser)	Cloud APIs
Privacy	Text never leaves your device	Text sent to provider servers
Cost	Free, unlimited	$1-3 per 1,000 requests typical
Setup	None - works instantly	API key, SDK, billing account
Languages	English only	20-100+ languages
Classification	Binary (positive/negative)	Multi-class, entity-level, aspect-based
Throughput	Limited by CPU speed	Thousands of requests per second
Offline use	Works after first model download	Requires internet connection

For quick checks on English text where privacy matters, the browser approach is ideal. For production-scale multilingual analysis, a cloud API or self-hosted model is the better choice.

Sources

• Sanh et al. - DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter (2019)
• Hugging Face - distilbert-base-uncased-finetuned-sst-2-english model card
• Hugging Face - Transformers.js documentation
• Stanford NLP - SST-2 dataset
• Precedence Research - Sentiment Analytics Market Size (2025)