Papers
Topics
Authors
Recent
Search
2000 character limit reached

Pre-training via Paraphrasing

Published 26 Jun 2020 in cs.CL, cs.LG, and stat.ML | (2006.15020v1)

Abstract: We introduce MARGE, a pre-trained sequence-to-sequence model learned with an unsupervised multi-lingual multi-document paraphrasing objective. MARGE provides an alternative to the dominant masked language modeling paradigm, where we self-supervise the reconstruction of target text by retrieving a set of related texts (in many languages) and conditioning on them to maximize the likelihood of generating the original. We show it is possible to jointly learn to do retrieval and reconstruction, given only a random initialization. The objective noisily captures aspects of paraphrase, translation, multi-document summarization, and information retrieval, allowing for strong zero-shot performance on several tasks. For example, with no additional task-specific training we achieve BLEU scores of up to 35.8 for document translation. We further show that fine-tuning gives strong performance on a range of discriminative and generative tasks in many languages, making MARGE the most generally applicable pre-training method to date.

Citations (150)

Summary

  • The paper introduces MARGE, a retrieval-based pre-training model that achieves strong zero-shot results with BLEU scores up to 35.8 on document translation.
  • It employs a unique multi-document paraphrasing objective to jointly optimize document retrieval and text reconstruction, diverging from traditional MLM approaches.
  • The model's efficiency in multilingual tasks promises reduced computational costs and improved performance, especially for less-resourced languages.

An Examination of "Pre-training via Paraphrasing" and the Introduction of MARGE

The paper "Pre-training via Paraphrasing" presents a novel approach to natural language processing pre-training through the introduction of MARGE (Multilingual Autoencoder that Retrieves and Generates). This model offers an alternative to the widely adopted masked language modeling (MLM) techniques, by leveraging an unsupervised multi-lingual multi-document paraphrasing objective for pre-training. MARGE is designed to improve sequence-to-sequence tasks across languages by focusing on the retrieval and reconstruction of target texts through paraphrasing related documents.

Overview of MARGE's Approach

MARGE diverges from traditional LLMs by utilizing a retrieval-based method to gather a set of documents related to a target text, potentially across multiple languages. These documents are then used by a reconstruction model to generate the original text. The uniqueness of MARGE lies in its joint optimization of document retrieval and reconstruction, starting from random initialization.

The retrieval model calculates relevance scores based on cosine similarities between document embeddings, serving not only to improve retrieval accuracy but also influencing the reconstruction phase. This innovative method significantly departs from the expected denoising of inputs seen in MLMs, where the noise is introduced by the retrieval process rather than artificial masking.

Performance Analysis

One of the paper's notable claims is MARGE's ability to offer strong zero-shot learning results, achieving BLEU scores up to 35.8 for document translation without task-specific fine-tuning. The experiments span various NLP tasks, including cross-lingual sentence retrieval, document-level machine translation, summarization, and paraphrase detection, indicating MARGE's broad applicability.

When evaluated on cross-lingual retrieval tasks such as BUCC2018 and Tatoeba, MARGE showcases superior performance compared to pretrained LLMs like mBERT, XLM, and XLM-R. Its document translation capabilities align closely with mBART on well-resourced language pairs, further validating the model's cross-lingual generation capabilities.

MARGE's performance in summarization tasks such as MLSum exhibits improvements over existing methods. In addition, it demonstrates effective zero-shot performance in summarization across languages, pointing to its potential in performing translation and summarization concurrently. The paraphrasing abilities tested on PAWS-X affirm that MARGE holds competitive performance, suggesting that the pre-training strategy effectively captures semantic equivalences across paraphrase contexts.

Implications and Future Directions

The introduction of MARGE broadens the understanding of pre-training methodologies beyond MLMs, highlighting the efficacy of retrieval and paraphrasing-based pre-training. Given MARGE's already competitive performance and operational efficiency compared to other models such as XLM-R, it provides a compelling argument for adopting alternative pre-training objectives.

In practical terms, MARGE's adaptability to multilingual tasks without necessitating extensive fine-tuning could significantly lower computational costs and resource requirements, especially for less represented languages. This is particularly advantageous for applications involving multilingual text generation and translation.

Moving forward, several challenges present opportunities for enhancements. These include expanding MARGE's domain adaptability, improving cross-language retrieval mechanisms to cater equally to all languages, and scaling up model parameters without sacrificing efficiency. Future research could also explore hybrid models that integrate the strengths of both retrieval-based methods and traditional MLM approaches to refine the balance between generative performance and computational requirements.

In summary, "Pre-training via Paraphrasing" and the development of MARGE make a meaningful contribution to multilingual NLP, potentially reshaping the landscape of pre-trained models by challenging the dominance of masked LLMs with a retrieval and reconstruction paradigm.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Continue Learning

We haven't generated follow-up questions for this paper yet.

Collections

Sign up for free to add this paper to one or more collections.