QUEST: Quality-Aware Metropolis-Hastings Sampling for Machine Translation

Gonçalo Faria, Sweta Agrawal, António Farinhas, Ricardo Rei, José G. C. de Souza, Andre Martins

Paper: https://arxiv.org/abs/2406.00049

TL;DR: This paper presents a method to generate diverse and high-quality machine translations by sampling from a Gibbs distribution using the Metropolis-Hastings algorithm.

Abstract:

An important challenge in machine translation (MT) is to generate high-quality and diverse translations. Prior work has shown that the estimated likelihood from the MT model correlates poorly with translation quality. In contrast, quality evaluation metrics (such as COMET or BLEURT) exhibit high correlations with human judgments, which has motivated their use as rerankers (such as quality-aware and minimum Bayes risk decoding). However, relying on a single translation with high estimated quality increases the chances of "gaming the metric''. In this paper, we address the problem of sampling a set of high-quality and diverse translations. We provide a simple and effective way to avoid over-reliance on noisy quality estimates by using them as the energy function of a Gibbs distribution. Instead of looking for a mode in the distribution, we generate multiple samples from high-density areas through the Metropolis-Hastings algorithm, a simple Markov chain Monte Carlo approach. The results show that our proposed method leads to high-quality and diverse outputs across multiple language pairs (English$\leftrightarrow${German, Russian}) with two strong decoder-only LLMs (Alma-7b, Tower-7b).

---

Documentation

TBD

---

Quick Start Examples

Install

Install using pip (recommended):

pip install quest-decoding

Install using pip (from github):

pip install git+https://github.com/deep-spin/quest-decoding.git

Sentiment Steering

    from langchain.prompts import PromptTemplate
    from quest import RewardModel
    from quest import VLLM, RLHFSuffixProposal


    template =  PromptTemplate.from_template(
        "I received the following comment on a X: {tweet}. How should I respond?:\n"
    ) 
    # a prompt template you define - usefull for tasks like translation. 
    # also really important to have the chatformat for the majority of models in use now. 
    
    test_input_data = [{
        "tweet": "You should refrain from commenting on this matter."
    }]

 
    model = VLLM(
        model_path="meta-llama/Llama-3.2-1B",
    )


    def toformat(data):

        return {
            "prompt": model.tokenizer.apply_chat_template(
                [
                    {
                        "role": "user",
                        "content": template.format(**data),
                    }
                ],
                tokenize=False,
                add_generation_prompt=True,
            ),
        }

    test_input_data=map(toformat,test_input_data)

    reward = RewardModel("lvwerra/distilbert-imdb")  # sentiment model from HF. 
    
    chain = Quest(
        input_data=test_input_data,
        proposal=RLHFSuffixProposal(
            model=model,
        ),
        beta=0.1,
        reward=reward,
    )
    
    chain_outputs = chain.run(
        steps=10,
        use_tqdm=True,
    )
    
    print(chain_outputs.samples)
        

---

Contact

For bugs and feature requests please visit GitHub Issues. For business inquiries or professional support requests please send an e-mail.

---

Citation

@misc{faria2024quest,
      title={QUEST: Quality-Aware Metropolis-Hastings Sampling for Machine Translation}, 
      author={Gonçalo R. A. Faria and Sweta Agrawal and António Farinhas and Ricardo Rei and José G. C. de Souza and André F. T. Martins},
      year={2024},
      eprint={2406.00049},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}