| --- |
| base_model: google/gemma-3-270m |
| pipeline_tag: text-classification |
| language: |
| - en |
| tags: |
| - sentiment-analysis |
| - text-classification |
| - seo |
| license: other |
| license_name: link-attribution |
| license_link: https://dejan.ai/link-attribution/ |
| --- |
| |
| # DEJAN AI Sentiment Classifier |
|
|
| This is a text sentiment classifier trained by [DEJAN AI](https://dejan.ai), small enough to run over thousands of pages on a laptop. Read how we built it, with training data, charts and results: [DEJAN releases a free sentiment model](https://dejan.ai/blog/sentiment-model/). Try it in the [public demo](https://dejan.ai/tools/sentiment/). |
|
|
| | Score | Label | |
| |---|---| |
| | 0 | very positive | |
| | 1 | positive | |
| | 2 | somewhat positive | |
| | 3 | neutral | |
| | 4 | somewhat negative | |
| | 5 | negative | |
| | 6 | very negative | |
|
|
| > [!IMPORTANT] |
| > **Feed it one sentence at a time, not a whole page.** |
| > The model reads up to about 90 words and ignores the rest. To score a page, split it into sentences and score each one. The example below shows how. |
|
|
| ## What you can use it for |
|
|
| - **Reviews:** score your own and competitors' reviews sentence by sentence, and find exactly which product features people praise or complain about. |
| - **Brand mentions:** measure how your brand is talked about in articles, forums, social posts, and AI answers. |
| - **Content audits:** find negative or off-tone sentences on landing pages, product pages, and service pages before users do. |
| - **Competitor pages:** compare how positive your comparison and product pages are against the pages that outrank you. |
| - **SERP research:** see whether the top-ranking content for a query is enthusiastic, balanced, or critical. |
|
|
| It understands the formats you work with: product pages, reviews, comparisons, FAQs, blog posts, press releases, forum threads, tweets, transcripts, and tables. |
|
|
| ## How accurate is it |
|
|
| | | | |
| |---|---| |
| | Exact score | **84%** | |
| | Exact or one point off | **95%** | |
|
|
| When it misses, it is almost always by one point, for example "negative" instead of "somewhat negative". |
|
|
|  |
|
|
|  |
|
|
|  |
|
|
| ## Quick start |
|
|
| ```python |
| from transformers import pipeline |
| |
| clf = pipeline("text-classification", model="dejanai/sentiment-classifier") |
| clf("Delivery took three weeks and nobody answered my emails.") |
| # [{'label': 'negative', 'score': 0.88}] |
| ``` |
|
|
| ## Scoring a whole page |
|
|
| ```python |
| import re |
| from transformers import pipeline |
| |
| clf = pipeline("text-classification", model="dejanai/sentiment-classifier") |
| |
| def sentences(text): |
| for line in text.splitlines(): |
| line = re.sub(r"^\s*(#+|[-*+]|\d+[.)])\s*", "", line).replace("**", "").strip() |
| if line: |
| yield from (s for s in re.split(r"(?<=[.!?])\s+(?=[A-Z0-9\"'])", line) if s) |
| |
| page = """# Six months with the X200 |
| The battery is superb. The camera, sadly, is a step backwards. |
| - Price: $499 |
| - Software updates have been slow.""" |
| |
| parts = list(sentences(page)) |
| for part, result in zip(parts, clf(parts)): |
| print(result["label"], "|", part) |
| ``` |
|
|
| ## Good to know |
|
|
| - English only. |
| - Each sentence is scored on its own, so a sentence like "And yet it works" may be misread without the sentence before it. |
| - Trained on about 83,000 sentences written and scored by AI across 26 content types. It has not yet been tested on real-world review datasets. |
|
|
| ## License |
|
|
| Released under the [DEJAN link attribution license](https://dejan.ai/link-attribution/). Built on Google's Gemma, so use is also subject to the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). |
|
|
