TopicAns aims at learning high-quality representations for the posts in Q&A sites with a neural topic model and a pre-trained model.

This involves three main steps: (1) generating topic-aware representations of Q&A posts with the neural topic model, (2) incorporating the corpus-level knowledge from the neural topic model to enhance the deep representations generated by the pre-trained language model, and (3) determining the most suitable answer for a given query based on the topic-aware representation and the deep representation.

Install all required python dependencies:

pip install -r requirements.txt

Our datasets are released at https://drive.google.com/drive/folders/1Pf8zyJGdufi_14qeAigPI1yyOkiL881J?usp=share_link.

Following is how we get them:

1. Download stackoverflow.com-Posts.7z from https://archive.org/details/stackexchange

2. Run dataset_scripts/split_qa.py to get:

   1. q_temp_output.csv: ["Id", "PostTypeId", "Score", "AnswerCount", "AcceptedAnswerId", "Title", "Body", "Tags"]
   2. a_temp_output.csv: ["Id", "PostTypeId", "Score", "ParentId", "Body"]
   3. python_question_info

3. Run dataset_scripts/extract_python_qa.ipynb to get python_qa (This file can also be used to get java_qa pairs!). This file does:

   1. Get python_question_info: (q_id, 1, accepted_a_id, other_a_ids...)
   2. Get python_q_id_content.csv, python_a_id_content.csv
   3. Then process the above two files
   4. delete some trash questions from python_question_info, and save the new one as python_qa_info.dataset. Please read this file by datasets (huggingface).

4. Then we need to construct candidate answer pools following Gao et al. by construct_candidate_pools.py.

5. Finally, we combine all together and get our datasets by combine_py_related_question.ipynb

python -u main.py -d so_python -n 0 \
    --model_class QATopicMemoryModel \
    --model_save_prefix top10_topic25_pooler_ \
    --pretrained_bert_path huggingface/CodeBERTa-small-v1 \
    --composition pooler \
    --train_batch_size 16 \
    --gradient_accumulation_steps 4 \
    --text_max_len 512 \
    --latent_dim 25 \
    --two_stage \
    --no_initial_test

The best model is saved at ./model/ and the lastest model is saved at ./last_model/. If training is broken, you can add --restore at the last without other changing. We have supported logging, so you can choose append | tee this_is_log.log to the command.

python -u main.py -d so_python -n 0 \
    --model_class QATopicMemoryModel  \
    --model_save_prefix top10_topic25_pooler_ \
    --pretrained_bert_path huggingface/CodeBERTa-small-v1 \
    --composition pooler \
    --train_batch_size 16 \
    --gradient_accumulation_steps 4 \
    --text_max_len 512 \
    --latent_dim 25 \
    --two_stage \
    --no_initial_test \
    --measure_time \
    --no_train

If you find this project useful in your research, please consider cite:

@article{10.1145/3607189,
    author = {Yang, Yuanhang and He, Wei and Gao, Cuiyun and Xu, Zenglin and Xia, Xin and Liu, Chuanyi},
    title = {TopicAns: Topic-Informed Architecture for Answer Recommendation on Technical Q&A Site},
    year = {2023},
    url = {https://doi.org/10.1145/3607189},
    journal = {ACM Trans. Softw. Eng. Methodol.},
}

This project is licensed under the MIT License.