AIO2 TF-IDF Baseline

This is a very simple question answering system, which is developed as a lightweight baseline for AIO2 competition.

In the training stage, the model builds a sparse matrix of TF-IDF features from the questions in training dataset. In the inference stage, the model predicts answers of unseen questions by finding the most similar training question to the input by computing dot product scores of TF-IDF features.

Therefore, in principle, the model cannot predict answers unseen in the training data.

Steps to experiment with the model

Install requirements

$ pip install -r requirements.txt


$ python \
--train_file <data dir>/aio_02_train.jsonl \
--output_dir model \
--pos_list 名詞 \
--stop_words でしょ う \
--max_features 10000


$ python \
--model_dir model \
--test_file <data dir>/aio_02_dev_unlabeled_v1.0.jsonl \
--prediction_file <output dir>/predictions.jsonl

Building Docker image

$ docker build -t aio2-tfidf-baseline .

Test locally:

$ docker run --rm -v "<data dir absolute path>:/app/input" -v "<output dir absolute path>:/app/output" aio2-tfidf-baseline bash ./ input/aio_02_dev_unlabeled_v1.0.jsonl output/predictions.jsonl

Save the docker image to file:

$ docker save aio2-tfidf-baseline | gzip > aio2-tfidf-baseline.tar.gz


The codes in this repository are open-sourced under MIT License.

GitHub - singletongue/aio2-tfidf-baseline at
TFIDF-based QA system for AIO2 competition. Contribute to singletongue/aio2-tfidf-baseline development by creating an account on GitHub.