dimionx / semantic-text-matching

ML-model for solving the problem of searching semantically similar text documents

Semantic Text Matching

ML-model for solving the problem of searching semantically similar text documents.

The model processes a new incoming text question and returns a list of N similar questions from an existing dataset of 4,567 medical-related questions.

Model

Unsupervised learner for implementing neighbor searches: Nearest Neighbors

Metric to use for distance computation: cosine

Preprocessing

1. Tokenization: TfidfVectorizer
   • tokenizer: custom function with regular pattern r"(?u)\b\w\w+\b"
   • token_pattern: None
2. Stop words: nltk.corpus.stopwords
3. Minimum word length: 2
4. Remove punctuation
5. Decode special symbols: html.unescape
6. Lemmatization: spaCy en_core_web_sm model

Input

question - string input question

Output

questions - List[str] output similar questions (from Dataset)

Metric

Top-N accuracy (accuracy@n)

Accuracy@5 - 0.887

Run

1. Clone repo

git clone https://github.com/DimionX/semantic-text-matching.git

cd semantic-text-matching

2. Creation of virtual environment

python -m venv .venv
source .venv/bin/activate

Dev:

pip install -r requirements.dev.txt
pre-commit install

Production:

pip install -r requirements.txt

3. Download dataset

Dataset info - Medical Questions Pairs

python ./src/load_dataset.py

4. Model Training

Load spacy models

python -m spacy download en

Model training and save

python ./src/save_model.py

Jupyter Notebook

main.ipynb

Run Streamlit

streamlit run ./src/stream.py

Streamlit app in your browser: http://127.0.0.1:8501

Run CLI

python ./src/matcher.py

Run Docker Compose

docker compose up

Streamlit app in your browser: http://127.0.0.1:8501