Mfaq

🚀 MFAQ

MFAQ is a multilingual FAQ retrieval model. Trained on the MFAQ dataset, it ranks candidate answers according to a given question, offering a practical solution for multilingual FAQ retrieval scenarios.

🚀 Quick Start

MFAQ is a multilingual FAQ retrieval model. It ranks candidate answers based on a given question, trained on the MFAQ dataset.

✨ Features

• Multilingual Support: Supports multiple languages including Czech, Danish, German, English, etc.
• Based on Popular Frameworks: Can be used with sentence-transformers and HuggingFace Transformers.

📦 Installation

pip install sentence-transformers transformers

💻 Usage Examples

Basic Usage

You can use MFAQ with sentence-transformers or directly with a HuggingFace model. In both cases, questions need to be prepended with <Q>, and answers with <A>.

Sentence Transformers

from sentence_transformers import SentenceTransformer

question = "<Q>How many models can I host on HuggingFace?"
answer_1 = "<A>All plans come with unlimited private models and datasets."
answer_2 = "<A>AutoNLP is an automatic way to train and deploy state-of-the-art NLP models, seamlessly integrated with the Hugging Face ecosystem."
answer_3 = "<A>Based on how much training data and model variants are created, we send you a compute cost and payment link - as low as $10 per job."

model = SentenceTransformer('clips/mfaq')
embeddings = model.encode([question, answer_1, answer_3, answer_3])
print(embeddings)

HuggingFace Transformers

from transformers import AutoTokenizer, AutoModel
import torch

def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

question = "<Q>How many models can I host on HuggingFace?"
answer_1 = "<A>All plans come with unlimited private models and datasets."
answer_2 = "<A>AutoNLP is an automatic way to train and deploy state-of-the-art NLP models, seamlessly integrated with the Hugging Face ecosystem."
answer_3 = "<A>Based on how much training data and model variants are created, we send you a compute cost and payment link - as low as $10 per job."

tokenizer = AutoTokenizer.from_pretrained('clips/mfaq')
model = AutoModel.from_pretrained('clips/mfaq')

# Tokenize sentences
encoded_input = tokenizer([question, answer_1, answer_3, answer_3], padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling. In this case, max pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

📚 Documentation

Training

You can find the training script for the model here.

People

This model was developed by Maxime De Bruyn, Ehsan Lotfi, Jeska Buhmann and Walter Daelemans.

Citation information

@misc{debruyn2021mfaq,
      title={MFAQ: a Multilingual FAQ Dataset}, 
      author={Maxime De Bruyn and Ehsan Lotfi and Jeska Buhmann and Walter Daelemans},
      year={2021},
      eprint={2109.12870},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

📄 License

This project is licensed under the Apache 2.0 license.