Papers by Alexei Baevski

9 papers
Multilingual Speech Translation from Efficient Finetuning of Pretrained Models (2021.acl-long)

Copied to clipboard

Challenge: Recent advances in text pretraining and finetuning have improved multitasking applications significantly.
Approach: They propose a minimalistic LNA finetuning approach to build multilingual speech-to-text translation using a pretrained speech encoder and text decoder.
Outcome: The proposed approach surpasses the cascaded ST benchmark for 36 translation directions on the large-scale multilingual ST benchmark CoVoST 2.
fairseq: A Fast, Extensible Toolkit for Sequence Modeling (N19-4)

Copied to clipboard

Challenge: OpenNMT is a community-built toolkit written in multiple languages with an emphasis on extensibility.
Approach: They propose to use PyTorch to train custom sequence models for translation, summarization, language modeling, and other tasks.
Outcome: The proposed toolkit is fast, extensible, and useful for both research and production.
Pre-trained language model representations for language generation (N19-1)

Copied to clipboard

Challenge: Pre-trained language model representations have been successful in a wide range of language understanding tasks.
Approach: They propose to use pre-trained language model representations to integrate them into sequence to sequence models and apply it to machine translation and abstractive summarization.
Outcome: The proposed model is able to perform 5.3 BLEU in machine translation and 5.3 on the full text version of CNN/DailyMail.
Introducing Semantics into Speech Encoders (2023.acl-long)

Copied to clipboard

Challenge: Existing self-supervised speech encoders contain primarily acoustic rather than semantic information.
Approach: They propose a task-agnostic unsupervised way to incorporate semantic information from large language model (LLM) systems into self-supervised speech encoders without labeled audio transcriptions.
Outcome: The proposed approach improves spoken language understanding (SLU) performance by over 5% on intent classification (IC), with modest gains in named entity resolution (NER) and slot filling (SF), and spoken question answering (SQA) score by over 22%.
Reservoir Transformers (2021.acl-long)

Copied to clipboard

Challenge: Using random initialization, we show that some transformers obtain impressive performance even when some of the layers are frozen.
Approach: They propose to freeze transformer layers and use them to improve performance . they find that the transformers obtain impressive performance even when some of the layers are randomly initialized and never updated.
Outcome: The proposed model improves on translation and language modelling tasks even when some layers are frozen.
Cloze-driven Pretraining of Self-attention Networks (D19-1)

Copied to clipboard

Challenge: Existing work on pretraining language models has used unidirectional (left-to-right) or bi-directional (both left-to right and right-to left) LMs with loss function.
Approach: They propose a bi-directional transformer model that pretrains both directions of a large language-model-inspired self-attention cloze model and propose clozing to predict each word in the training data.
Outcome: The proposed model performs well on GLUE and state of the art benchmarks consistent with BERT.
On Generative Spoken Language Modeling from Raw Audio (2021.tacl-1)

Copied to clipboard

Challenge: Using a set of metrics to evaluate the learned representations, we aim to create a system that learns from natural interactions as infants learn their first language.
Approach: They propose a task of learning acoustic and linguistic characteristics from raw audio and a set of metrics to evaluate the learned representations at acustic, linguistic and encoding levels.
Outcome: The proposed models evaluate the learned representations at acoustic and linguistic levels for both encoding and generation.
Unified Speech-Text Pre-training for Speech Translation and Recognition (2022.acl-long)

Copied to clipboard

Challenge: Existing methods to pre-train speech and text use unlabeled data to learn universal feature representations.
Approach: They propose a method to jointly pre-train speech and text in an encoder-decoder modeling framework for speech translation and recognition.
Outcome: The proposed method achieves between 1.7 and 2.3 BLEU improvement above the state of the art on the MuST-C speech translation dataset and comparable WERs to wav2vec 2.0 on the Librispeech speech recognition task.
Toward Joint Language Modeling for Speech Units and Text (2023.findings-emnlp)

Copied to clipboard

Challenge: Speech and text are two major forms of human language and little effort has been made to model them together.
Approach: They propose to combine speech and text models to create mixed speech-text data by using different tokenizers and automatic metrics to evaluate how well the model mixes speech and texts.
Outcome: The proposed model improves over a speech-only baseline and shows zero-shot cross-modal transferability.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations