My Voice — clustering open-ended survey answers by meaning

Hackathon solution for Rosatom's "My Voice" service, by team Users. The service collects free-text answers from employees; the problem is grouping thousands of differently-worded answers that mean the same thing.

We group them with multilingual BERT embeddings and a layered clustering pipeline, then present the result in a web interface. The approach is robust to typos and misspellings, and it keeps latent signal that naive matching throws away — emoji and punctuation.

Model comparison

Model	F1 macro	Time
Naive — every word is its own cluster	0.81	10 ms
Levenshtein similarity — answers >63% similar form one cluster	0.87	102 ms
Levenshtein + preprocessing (lemmatisation, punctuation removal)	0.89	1 s
SelfClusterModel #1 + sentiment transformer (BERT-multilingual + PCA + KMeans, TweetNLP + XLM-RoBERTa)	0.92
SelfClusterModel #2	0.94	6 s
SelfClusterModel #2 + sentiment transformer	0.97

How it works

• Preprocessing of raw answers: punctuation removal, profanity filtering, lemmatisation.
• Fine-tuned deep-learning models on the processed data: TweetNLP, XLM-RoBERTa multilingual sentiment classification, BERTopic.
• Multi-stage clustering: UMAP → HDBSCAN → CountVectorizer → TF-IDF.
• Sentiment scoring with XLM-RoBERTa trained on user comments.
• Clustering visualised through a Streamlit interface.

The combination of BERT-based text processing, Streamlit visualisation, and handling of both Russian and English gives accuracy without costing usability.

Stack

Python 3, git, GitHub — development HF Transformers, TweetNLP, BERTopic — deep learning Scikit-Learn, UMAP, KMeans — machine learning Plotly, Streamlit, AltChart — visualisation

Install and run

Requires Python 3.9 or newer.

git clone https://github.com/Baltsat/users-rosatom.git
pip install -r requirements.txt
streamlit run main.py

Notebooks

• research_models_visualization.ipynb — gradient-boosting model experiments
• data_preprocess.ipynb — data preprocessing

Django API

The project uses Django and Django REST Framework to expose the question-and-answer data.

/api/qaitems

• GET — list all QA items
• POST — create a new QA item

Data structure

• question — question text
• answer — answer text
• sentiment — sentiment of the item
• j — J value
• cluster_id — cluster identifier
• topic_name — topic name

Running the API

1. Clone the repository.
2. Create and activate a virtual environment.
3. pip install -r requirements.txt
4. python manage.py migrate
5. python manage.py runserver

Example request: http://localhost:8000/api/qaitems/

Team

Name	Role	Contact
Konstantin Baltsat	Data analysis	t.me/baltsat
Aleksandr Serov	Machine learning	t.me/thegoldian
Artem Tarasov	Full stack	t.me/tarasovxx
Sergey Vandanov	Machine learning	t.me/rapid76
Daniil Galimov	Data analysis	t.me/Dan_Gan