Publications
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Sort By
1 - 15 of 19 publications
Preview abstract
We study the problem of model extraction in natural language processing, where an adversary with query access to a victim model attempts to reconstruct a local copy of the model. We show that when both the adversary and victim model fine-tune existing pretrained models such as BERT, the adversary does not need to have access to any training data to mount the attack. Indeed, we show that randomly sampled sequences of words, which do not satisfy grammar structures, make effective queries to extract textual models. This is true even for complex tasks such as natural language inference or question answering.
Our attacks can be mounted with a modest query budget of less than $400.The extraction's accuracy can be further improved using a large textual corpus like Wikipedia, or with intuitive heuristics we introduce. Finally, we measure the effectiveness of two potential defense strategies---membership classification and API watermarking. While these defenses mitigate certain adversaries and come at a low overhead because they do not require re-training of the victim model, fully coping with model extraction remains an open problem.
View details
Ensemble Distillation for BERT-Based Ranking Models
Shuguang Han
Mike Bendersky
Proceedings of the 2021 ACM SIGIR International Conference on the Theory of Information Retrieval (ICTIR ’21)
Preview abstract
Over the past two years, large pretrained language models such as BERT have been applied to text ranking problems and showed superior performance on multiple public benchmark data sets. Prior work demonstrated that an ensemble of multiple BERT-based ranking models can not only boost the performance, but also reduce the performance variance. However, an ensemble of models is more costly because it needs computing resource and/or inference time proportional to the number of models. In this paper, we study how to retain the performance of an ensemble of models at the inference cost of a single model by distilling the ensemble into a single BERT-based student ranking model. Specifically, we study different designs of teacher labels, various distillation strategies, as well as multiple distillation losses tailored for ranking problems. We conduct experiments on the MS MARCO passage ranking and the TREC-COVID data set. Our results show that even with these simple distillation techniques, the distilled model can effectively retain the performance gain of the ensemble of multiple models. More interestingly, the performances of distilled models are also more stable than models fine-tuned on original labeled data. The results reveal a promising direction to capitalize on the gains achieved by an ensemble of BERT-based ranking models.
View details
Preview abstract
This paper introduces a new encoder model for neural TTS. The proposed model, called PnG BERT, is augmented from the original BERT model, but taking both phoneme and grapheme representation of a text, as well as the word-level alignment between them, as its input. It can be pre-trained on a large text corpus in a self-supervised manner then fine-tuned in a TTS task. The experimental results suggest that PnG BERT can significantly further improve the performance of a state-of-the-art neural TTS model, by producing more appropriate prosody and more accurate pronunciation. Subjective side-by-side preference evaluation showed that raters had no statistically significant preference between the synthesized speech and the ground truth recordings from professional speakers.
View details
Preview abstract
The prosody of currently available speech synthesis systems can be unnatural due to the systems only having access to the text, possibly enriched by linguistic information such as part-of-speech tags and parse trees. We show that incorporating a BERT model in an RNN-based speech synthesis model - where the BERT model is pretrained on large amounts of unlabeled data, and fine-tuned to the speech domain - improves prosody. Additionally, we propose a way of handling arbitrarily long sequences with BERT. Our findings indicate that small BERT models work better than big ones, and that fine-tuning
the BERT part of the model is pivotal for getting good results.
View details
How multilingual is Multilingual BERT?
Telmo Pires
Eva Schlinger
Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics (2019)
Preview abstract
In this paper, we show that Multilingual BERT (M-BERT), released by Devlin et al. (2018) as a single language model pre-trained from monolingual corpora in 104 languages, is surprisingly good at zero-shot cross-lingual model transfer, in which task-specific annotations in one language are used to fine-tune the model for evaluation in another language. To understand why, we present a large number of probing experiments, showing that transfer is possible even to languages in different scripts, that transfer works best between typologically similar languages, that monolingual corpora can train models for code-switching, and that the model can find translation pairs. From these results, we can conclude that M-BERT does create multilingual representations, but that these representations exhibit systematic deficiencies affecting certain language pairs.
View details
Preview abstract
We propose a practical scheme to train a single multilingual sequence labeling model that yields state of the art results and is small and fast enough to run on a single CPU. Starting from a public multilingual BERT checkpoint, our final model is 34x smaller and 15x faster, and has higher accuracy than a state-of-the-art multilingual baseline. We show that our model especially outperforms on low-resource languages, and works on codemixed input text without being explicitly trained on codemixed examples. And we show the effectiveness of our method by reporting on part-of-speech tagging and morphological prediction on 70 treebanks and 47 languages.
View details
Preview abstract
While recent work on multilingual language models has demonstrated their capacity for cross-lingual zero-shot transfer on downstream tasks, there is a lack of consensus in the community as to what shared properties between languages enable such transfer. Analyses involving pairs of natural languages are often inconclusive and contradictory since languages simultaneously differ in many linguistic aspects. In this paper, we perform a large-scale empirical study to isolate the effects of various linguistic properties by measuring zero-shot transfer between four diverse natural languages and their counterparts constructed by modifying aspects such as the script, word order, and syntax. Among other things, our experiments show that the absence of sub-word overlap significantly affects zero-shot transfer when languages differ in their word order, and there is a strong correlation between transfer performance and word embedding alignment between languages (e.g., R=0.94 on the task of NLI). Our results call for focus in multilingual models on explicitly improving word embedding alignment between languages rather than relying on its implicit emergence.
View details
Sparse Mixers: Combining MoE and Mixing to build a more efficient BERT
James Patrick Lee-Thorp
Joshua Ainslie
Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 58–75
Preview abstract
We combine capacity of sparsely gated Mixture-of-Experts (MoE) with the speed and stability from mixing transformations to design the Sparse Mixer encoder model. The Sparse Mixer just (<1%) outperforms BERT on GLUE and SuperGLUE, but more importantly trains 65% faster and runs inference 61% faster. We also present a faster variant, Fast Sparse Mixer, that very slightly (<0.2%) under-performs BERT on SuperGLUE, but trains and runs nearly twice as fast: 89% faster training and 98% faster inference. We justify the design of these two models by carefully ablating through various mixing mechanisms, MoE configurations and model hyperparameters. The Sparse Mixer overcomes the speed and stability concerns of MoE models and shows that smaller sparse models may be served out of the box, without resorting to distilling them to dense student models.
View details
What Happens To BERT Embeddings During Fine-tuning?
Amil Merchant
Elahe Rahimtoroghi
Proceedings of the 2020 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, Association for Computational Linguistics (to appear)
Preview abstract
While there has been much recent work studying how linguistic information is encoded in pre-trained sentence representations, comparatively little is understood about how these models change when adapted to solve downstream tasks. Using a suite of analysis techniques (probing classifiers, Representational Similarity Analysis, and model ablations), we investigate how fine-tuning affects the representations of the BERT model. We find that while fine-tuning necessarily makes significant changes, it does not lead to catastrophic forgetting of linguistic phenomena. We instead find that fine-tuning primarily affects the top layers of BERT, but with noteworthy variation across tasks. In particular, dependency parsing reconfigures most of the model, whereas SQuAD and MNLI appear to involve much shallower processing. Finally, we also find that fine-tuning has a weaker effect on representations of out-of-domain sentences, suggesting room for improvement in model generalization.
View details
The MultiBERTs: BERT Reproductions for Robustness Analysis
Steve Yadlowsky
Jason Wei
Naomi Saphra
Iulia Raluca Turc
2022
Preview abstract
Experiments with pretrained models such as BERT are often based on a single checkpoint. While the conclusions drawn apply to the artifact (i.e., the particular instance of the model), it is not always clear whether they hold for the more general procedure (which includes the model architecture, training data, initialization scheme, and loss function). Recent work has shown that re-running pretraining can lead to substantially different conclusions about performance, suggesting that alternative evaluations are needed to make principled statements about procedures. To address this question, we introduce MultiBERTs: a set of 25 BERT-base checkpoints, trained with similar hyper-parameters as the original BERT model but differing in random initialization and data shuffling. The aim is to enable researchers to draw robust and statistically justified conclusions about pretraining procedures. The full release includes 25 fully trained checkpoints, as well as statistical guidelines and a code library implementing our recommended hypothesis testing methods. Finally, for five of these models we release a set of 28 intermediate checkpoints in order to support research on learning dynamics.
View details
Assessing ASR Model Quality on Disordered Speech using BERTScore
Qisheng Li
Katie Seaver
Richard Jonathan Noel Cave
Katrin Tomanek
Proc. 1st Workshop on Speech for Social Good (S4SG) (2022), pp. 26-30 (to appear)
Preview abstract
Word Error Rate (WER) is the primary metric used to assess automatic speech recognition (ASR) model quality. It has been shown that ASR models tend to have much higher WER on speakers with speech impairments than typical English speakers. It is hard to determine if models can be be useful at such high error rates. This study investigates the use of BERTScore, an evaluation metric for text generation, to provide a more informative measure of ASR model quality and usefulness. Both BERTScore and WER were compared to prediction errors manually annotated by Speech Language Pathologists for error type and assessment. BERTScore was found to be more correlated with human assessment of error type and assessment. BERTScore was specifically more robust to orthographic changes (contraction and normalization errors) where meaning was preserved. Furthermore, BERTScore was a better fit of error assessment than WER, as measured using an ordinal logistic regression and the Akaike's Information Criterion (AIC). Overall, our findings suggest that BERTScore can complement WER when assessing ASR model performance from a practical perspective, especially for accessibility applications where models are useful even at lower accuracy than for typical speech.
View details
Preview abstract
We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.
BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).
View details
Preview abstract
In this work, we study the large-scale pretraining of BERT-Large~\citep{devlin2018bert} with differentially private SGD (DP-SGD). We show that combined with a careful implementation, scaling up the batch size to millions (i.e., mega-batches) improves the utility of the DP-SGD step for BERT; we also enhance the training efficiency by using an increasing batch size schedule. Our implementation builds on the recent work of \citet{subramani20}, who demonstrated that the overhead of a DP-SGD step is minimized with effective use of JAX \cite{jax2018github, frostig2018compiling} primitives in conjunction with the XLA compiler \cite{xladocs}. Our implementation achieves a masked language model accuracy of 60.5\% at a batch size of 2M, for $\eps = 5$, which is a reasonable privacy setting. To put this number in perspective, non-private BERT models achieve an accuracy of $\sim$70\%.
View details
Jigsaw @ AMI and HaSpeeDe2: Fine-Tuning a Pre-TrainedComment-Domain BERT Model
Alyssa Whitlock Lees
Ian Kivlichan
Proceedings of Seventh Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2020), CEUR.org, Online (to appear)
Preview abstract
The Google Jigsaw team produced submissions for two of the
EVALITA 2020 shared asks, based in part on the
technology that powers the publicly available
PerspectiveAPI comment evaluation service.
We present a basic description of our submitted results and a review of
the types of errors that our system made in these shared tasks.
View details
Preview abstract
Reading comprehension models have been successfully applied to extractive text answers, but it is unclear how best to generalize these models to abstractive numerical answers. We enable a BERT-based reading comprehension model to perform lightweight numerical reasoning. We augment the model with a predefined set of executable 'programs' which encompass simple arithmetic as well as extraction. Rather than having to learn to manipulate numbers directly, the model can pick a program and execute it. On the recent Discrete Reasoning Over Passages (DROP) dataset, designed to challenge reading comprehension models, we show a 33% absolute improvement by adding shallow programs. The model can learn to predict new operations when appropriate in a math word problem setting (Roy and Roth, 2015) with very few training examples.
View details
