Edit model card

MQDD - Multimodal Question Duplicity Detection

This repository publishes trained models and other supporting materials for the paper MQDD – Pre-training of Multimodal Question Duplicity Detection for Software Engineering Domain. For more information, see the paper. The Stack Overflow Datasets (SOD) and Stack Overflow Duplicity Dataset (SODD) presented in the paper can be obtained from our Stack Overflow Dataset repository.

To acquire the pre-trained model only, see the UWB-AIR/MQDD-pretrained.

Fine-tuned MQDD

We release a fine-tuned version of our MQDD model for duplicate detection task. The model's architecture follows the architecture of a two-tower model as depicted in the figure below:

A self-standing encoder without a duplicate detection head can be loaded using the following source code snippet. Such a model can be used for building search systems based, for example, on Faiss library.

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("UWB-AIR/MQDD-duplicates")
model = AutoModel.from_pretrained("UWB-AIR/MQDD-duplicates")

A checkpoint of a full two-tower model can than be obtained from our GoogleDrive folder. To load the model, one needs to use the model's implementation from models/MQDD_model.py in our GitHub repository. To construct the model and load it's checkpoint, use the following source code:

from MQDD_model import ClsHeadModelMQDD

model = ClsHeadModelMQDD("UWB-AIR/MQDD-duplicates")
ckpt = torch.load("model.pt",  map_location="cpu")


This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. http://creativecommons.org/licenses/by-nc-sa/4.0/

How should I cite the MQDD?

For now, please cite the Arxiv paper:

  doi = {10.48550/ARXIV.2203.14093},
  url = {https://arxiv.org/abs/2203.14093},
  author = {Pašek, Jan and Sido, Jakub and Konopík, Miloslav and Pražák, Ondřej},
  title = {MQDD -- Pre-training of Multimodal Question Duplicity Detection for Software Engineering Domain},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Attribution Non Commercial Share Alike 4.0 International}
Downloads last month