Ciprian Chelba

Chris Alberti, Kuzman Ganchev, Michael Collins et al. · deepmind

The paper presents an approach to semantic grounding of language models (LMs) that conceptualizes the LM as a conditional model generating text given a desired semantic message formalized as a set of entity-relationship triples. It embeds the LM in an auto-encoder by feeding its output to a semantic parser whose output is in the same representation domain as the input message. Compared to a baseline that generates text using greedy search, we demonstrate two techniques that improve the fluency and semantic accuracy of the generated text: The first technique samples multiple candidate text sequences from which the semantic parser chooses. The second trains the language model while keeping the semantic parser frozen to improve the semantic accuracy of the auto-encoder. We carry out experiments on the English WebNLG 3.0 data set, using BLEU to measure the fluency of generated text and standard parsing metrics to measure semantic accuracy. We show that our proposed approaches significantly improve on the greedy search baseline. Human evaluation corroborates the results of the automatic evaluation experiments.

Rami Botros, Rohit Prabhavalkar, Johan Schalkwyk et al.

In end-to-end (E2E) speech recognition models, a representational tight-coupling inevitably emerges between the encoder and the decoder. We build upon recent work that has begun to explore building encoders with modular encoded representations, such that encoders and decoders from different models can be stitched together in a zero-shot manner without further fine-tuning. While previous research only addresses full-context speech models, we explore the problem in a streaming setting as well. Our framework builds on top of existing encoded representations, converting them to modular features, dubbed as Lego-Features, without modifying the pre-trained model. The features remain interchangeable when the model is retrained with distinct initializations. Though sparse, we show that the Lego-Features are powerful when tested with RNN-T or LAS decoders, maintaining high-quality downstream performance. They are also rich enough to represent the first-pass prediction during two-pass deliberation. In this scenario, they outperform the N-best hypotheses, since they do not need to be supplemented with acoustic features to deliver the best results. Moreover, generating the Lego-Features does not require beam search or auto-regressive computation. Overall, they present a modular, powerful and cheap alternative to the standard encoder output, as well as the N-best hypotheses.

Ciprian Chelba, Johan Schalkwyk

We present a modular approach to building cascade speech translation (AST) models that guarantees that the resulting model performs no worse than the 1-best cascade baseline while preserving state-of-the-art speech recognition (ASR) and text translation (MT) performance for a given task. Our novel contribution is the use of an ``exporter'' layer that is trained under L2-loss to ensure a strong match between ASR embeddings and the MT token embeddings for the 1-best sequence. The ``exporter'' output embeddings are fed directly to the MT model in lieu of 1-best token embeddings, thus guaranteeing that the resulting model performs no worse than the 1-best cascade baseline, while allowing back-propagation gradient to flow from the MT model into the ASR components. The matched-embeddings cascade architecture provide a significant improvement over its 1-best counterpart in scenarios where incremental training of the MT model is not an option and yet we seek to improve quality by leveraging (speech, transcription, translated transcription) data provided with the AST task. The gain disappears when the MT model is incrementally trained on the parallel text data available with the AST task. The approach holds promise for other scenarios that seek to couple ASR encoders and immutable text models, such at large language models (LLM).

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Behrooz Ghorbani, Orhan Firat, Markus Freitag et al.

We present an empirical study of scaling properties of encoder-decoder Transformer models used in neural machine translation (NMT). We show that cross-entropy loss as a function of model size follows a certain scaling law. Specifically (i) We propose a formula which describes the scaling behavior of cross-entropy loss as a bivariate function of encoder and decoder size, and show that it gives accurate predictions under a variety of scaling approaches and languages; we show that the total number of parameters alone is not sufficient for such purposes. (ii) We observe different power law exponents when scaling the decoder vs scaling the encoder, and provide recommendations for optimal allocation of encoder/decoder capacity based on this observation. (iii) We also report that the scaling behavior of the model is acutely influenced by composition bias of the train/test sets, which we define as any deviation from naturally generated text (either via machine generated or human translated text). We observe that natural text on the target side enjoys scaling, which manifests as successful reduction of the cross-entropy loss. (iv) Finally, we investigate the relationship between the cross-entropy loss and the quality of the generated translations. We find two different behaviors, depending on the nature of the test data. For test sets which were originally translated from target language to source language, both loss and BLEU score improve as model size increases. In contrast, for test sets originally translated from source language to target language, the loss improves, but the BLEU score stops improving after a certain threshold. We release generated text from all models used in this study.

Ciprian Chelba, Junpei Zhou, Yuezhang et al.

The paper investigates the feasibility of confidence estimation for neural machine translation models operating at the high end of the performance spectrum. As a side product of the data annotation process necessary for building such models we propose sentence level accuracy $SACC$ as a simple, self-explanatory evaluation metric for quality of translation. Experiments on two different annotator pools, one comprised of non-expert (crowd-sourced) and one of expert (professional) translators show that $SACC$ can vary greatly depending on the translation proficiency of the annotators, despite the fact that both pools are about equally reliable according to Krippendorff's alpha metric; the relatively low values of inter-annotator agreement confirm the expectation that sentence-level binary labeling $good$ / $needs\ work$ for translation out of context is very hard. For an English-Spanish translation model operating at $SACC = 0.89$ according to a non-expert annotator pool we can derive a confidence estimate that labels 0.5-0.6 of the $good$ translations in an "in-domain" test set with 0.95 Precision. Switching to an expert annotator pool decreases $SACC$ dramatically: $0.61$ for English-Spanish, measured on the exact same data as above. This forces us to lower the CE model operating point to 0.9 Precision while labeling correctly about 0.20-0.25 of the $good$ translations in the data. We find surprising the extent to which CE depends on the level of proficiency of the annotator pool used for labeling the data. This leads to an important recommendation we wish to make when tackling CE modeling in practice: it is critical to match the end-user expectation for translation quality in the desired domain with the demands of annotators assigning binary quality labels to CE training data.

Hongge Chen, Si Si, Yang Li et al.

Multi-stage training and knowledge transfer, from a large-scale pretraining task to various finetuning tasks, have revolutionized natural language processing and computer vision resulting in state-of-the-art performance improvements. In this paper, we develop a multi-stage influence function score to track predictions from a finetuned model all the way back to the pretraining data. With this score, we can identify the pretraining examples in the pretraining task that contribute most to a prediction in the finetuning task. The proposed multi-stage influence function generalizes the original influence function for a single model in (Koh & Liang, 2017), thereby enabling influence computation through both pretrained and finetuned models. We study two different scenarios with the pretrained embeddings fixed or updated in the finetuning tasks. We test our proposed method in various experiments to show its effectiveness and potential applications.

Junpei Zhou, Ciprian Chelba, Yuezhang et al.

Sentence level quality estimation (QE) for machine translation (MT) attempts to predict the translation edit rate (TER) cost of post-editing work required to correct MT output. We describe our view on sentence-level QE as dictated by several practical setups encountered in the industry. We find consumers of MT output---whether human or algorithmic ones---to be primarily interested in a binary quality metric: is the translated sentence adequate as-is or does it need post-editing? Motivated by this we propose a quality classification (QC) view on sentence-level QE whereby we focus on maximizing recall at precision above a given threshold. We demonstrate that, while classical QE regression models fare poorly on this task, they can be re-purposed by replacing the output regression layer with a binary classification one, achieving 50-60\% recall at 90\% precision. For a high-quality MT system producing 75-80\% correct translations, this promises a significant reduction in post-editing work indeed.

Ciprian Chelba, Mia Chen, Ankur Bapna et al.

Motivated by the fact that most of the information relevant to the prediction of target tokens is drawn from the source sentence $S=s_1, \ldots, s_S$, we propose truncating the target-side window used for computing self-attention by making an $N$-gram assumption. Experiments on WMT EnDe and EnFr data sets show that the $N$-gram masked self-attention model loses very little in BLEU score for $N$ values in the range $4, \ldots, 8$, depending on the task.

Isaac Caswell, Ciprian Chelba, David Grangier

Recent work in Neural Machine Translation (NMT) has shown significant quality gains from noised-beam decoding during back-translation, a method to generate synthetic parallel data. We show that the main role of such synthetic noise is not to diversify the source side, as previously suggested, but simply to indicate to the model that the given source is synthetic. We propose a simpler alternative to noising techniques, consisting of tagging back-translated source sentences with an extra token. Our results on WMT outperform noised back-translation in English-Romanian and match performance on English-German, re-defining state-of-the-art in the former.

Wei Wang, Isaac Caswell, Ciprian Chelba

Noise and domain are important aspects of data quality for neural machine translation. Existing research focus separately on domain-data selection, clean-data selection, or their static combination, leaving the dynamic interaction across them not explicitly examined. This paper introduces a "co-curricular learning" method to compose dynamic domain-data selection with dynamic clean-data selection, for transfer learning across both capabilities. We apply an EM-style optimization procedure to further refine the "co-curriculum". Experiment results and analysis with two domains demonstrate the effectiveness of the method and the properties of data scheduled by the co-curriculum.

Jonathan Shen, Patrick Nguyen, Yonghui Wu et al.

Lingvo is a Tensorflow framework offering a complete solution for collaborative deep learning research, with a particular focus towards sequence-to-sequence models. Lingvo models are composed of modular building blocks that are flexible and easily extensible, and experiment configurations are centralized and highly customizable. Distributed training and quantized inference are supported directly within the framework, and it contains existing implementations of a large number of utilities, helper functions, and the newest research ideas. Lingvo has been used in collaboration by dozens of researchers in more than 20 papers over the last two years. This document outlines the underlying design of Lingvo and serves as an introduction to the various pieces of the framework, while also offering examples of advanced features that showcase the capabilities of the framework.

Wei Wang, Taro Watanabe, Macduff Hughes et al.

Measuring domain relevance of data and identifying or selecting well-fit domain data for machine translation (MT) is a well-studied topic, but denoising is not yet. Denoising is concerned with a different type of data quality and tries to reduce the negative impact of data noise on MT training, in particular, neural MT (NMT) training. This paper generalizes methods for measuring and selecting data for domain MT and applies them to denoising NMT training. The proposed approach uses trusted data and a denoising curriculum realized by online data selection. Intrinsic and extrinsic evaluations of the approach show its significant effectiveness for NMT to train on data with severe noise.

Patrick H. Chen, Si Si, Yang Li et al.

Model compression is essential for serving large deep neural nets on devices with limited resources or applications that require real-time responses. As a case study, a state-of-the-art neural language model usually consists of one or more recurrent layers sandwiched between an embedding layer used for representing input tokens and a softmax layer for generating output tokens. For problems with a very large vocabulary size, the embedding and the softmax matrices can account for more than half of the model size. For instance, the bigLSTM model achieves state-of- the-art performance on the One-Billion-Word (OBW) dataset with around 800k vocabulary, and its word embedding and softmax matrices use more than 6GBytes space, and are responsible for over 90% of the model parameters. In this paper, we propose GroupReduce, a novel compression method for neural language models, based on vocabulary-partition (block) based low-rank matrix approximation and the inherent frequency distribution of tokens (the power-law distribution of words). The experimental results show our method can significantly outperform traditional compression methods such as low-rank approximation and pruning. On the OBW dataset, our method achieved 6.6 times compression rate for the embedding and softmax matrices, and when combined with quantization, our method can achieve 26 times compression rate, which translates to a factor of 12.8 times compression for the entire model with very little degradation in perplexity.

Ciprian Chelba, Mohammad Norouzi, Samy Bengio

We investigate the effective memory depth of RNN models by using them for $n$-gram language model (LM) smoothing. Experiments on a small corpus (UPenn Treebank, one million words of training data and 10k vocabulary) have found the LSTM cell with dropout to be the best model for encoding the $n$-gram state when compared with feed-forward and vanilla RNN models. When preserving the sentence independence assumption the LSTM $n$-gram matches the LSTM LM performance for $n=9$ and slightly outperforms it for $n=13$. When allowing dependencies across sentence boundaries, the LSTM $13$-gram almost matches the perplexity of the unlimited history LSTM LM. LSTM $n$-gram smoothing also has the desirable property of improving with increasing $n$-gram order, unlike the Katz or Kneser-Ney back-off estimators. Using multinomial distributions as targets in training instead of the usual one-hot target is only slightly beneficial for low $n$-gram orders. Experiments on the One Billion Words benchmark show that the results hold at larger scale: while LSTM smoothing for short $n$-gram contexts does not provide significant advantages over classic N-gram models, it becomes effective with long contexts ($n > 5$); depending on the task and amount of data it can match fully recurrent LSTM models at about $n=13$. This may have implications when modeling short-format text, e.g. voice search/query LMs. Building LSTM $n$-gram LMs may be appealing for some practical situations: the state in a $n$-gram LM can be succinctly represented with $(n-1)*4$ bytes storing the identity of the words in the context and batches of $n$-gram contexts can be processed in parallel. On the downside, the $n$-gram context encoding computed by the LSTM is discarded, making the model more expensive than a regular recurrent LSTM LM.

Ciprian Chelba, Fernando Pereira

We describe Sparse Non-negative Matrix (SNM) language model estimation using multinomial loss on held-out data. Being able to train on held-out data is important in practical situations where the training data is usually mismatched from the held-out/test data. It is also less constrained than the previous training algorithm using leave-one-out on training data: it allows the use of richer meta-features in the adjustment model, e.g. the diversity counts used by Kneser-Ney smoothing which would be difficult to deal with correctly in leave-one-out training. In experiments on the one billion words language modeling benchmark, we are able to slightly improve on our previous results which use a different loss function, and employ leave-one-out training on a subset of the main training set. Surprisingly, an adjustment model with meta-features that discard all lexical information can perform as well as lexicalized meta-features. We find that fairly small amounts of held-out data (on the order of 30-70 thousand words) are sufficient for training the adjustment model. In a real-life scenario where the training data is a mix of data sources that are imbalanced in size, and of different degrees of relevance to the held-out and test data, taking into account the data source for a given skip-/n-gram feature and combining them for best performance on held-out/test data improves over skip-/n-gram SNM models trained on pooled data by about 8% in the SMT setup, or as much as 15% in the ASR/IME setup. The ability to mix various data sources based on how relevant they are to a mismatched held-out set is probably the most attractive feature of the new estimation method for SNM LM.

Noam Shazeer, Joris Pelemans, Ciprian Chelba

We present a novel family of language model (LM) estimation techniques named Sparse Non-negative Matrix (SNM) estimation. A first set of experiments empirically evaluating it on the One Billion Word Benchmark shows that SNM $n$-gram LMs perform almost as well as the well-established Kneser-Ney (KN) models. When using skip-gram features the models are able to match the state-of-the-art recurrent neural network (RNN) LMs; combining the two modeling techniques yields the best known result on the benchmark. The computational advantages of SNM over both maximum entropy and RNN LM estimation are probably its main strength, promising an approach that has the same flexibility in combining arbitrary features effectively and yet should scale to very large amounts of data as gracefully as $n$-gram LMs do.

Ciprian Chelba, Tomas Mikolov, Mike Schuster et al.

We propose a new benchmark corpus to be used for measuring progress in statistical language modeling. With almost one billion words of training data, we hope this benchmark will be useful to quickly evaluate novel language modeling techniques, and to compare their contribution when combined with other advanced techniques. We show performance of several well-known types of language models, with the best results achieved with a recurrent neural network based language model. The baseline unpruned Kneser-Ney 5-gram model achieves perplexity 67.6; a combination of techniques leads to 35% reduction in perplexity, or 10% reduction in cross-entropy (bits), over that baseline. The benchmark is available as a code.google.com project; besides the scripts needed to rebuild the training/held-out data, it also makes available log-probability values for each word in each of ten held-out data sets, for each of the baseline n-gram models.

Ciprian Chelba, Peng Xu, Fernando Pereira et al.

The paper revives an older approach to acoustic modeling that borrows from n-gram language modeling in an attempt to scale up both the amount of training data and model size (as measured by the number of parameters in the model), to approximately 100 times larger than current sizes used in automatic speech recognition. In such a data-rich setting, we can expand the phonetic context significantly beyond triphones, as well as increase the number of Gaussian mixture components for the context-dependent states that allow it. We have experimented with contexts that span seven or more context-independent phones, and up to 620 mixture components per state. Dealing with unseen phonetic contexts is accomplished using the familiar back-off technique used in language modeling due to implementation simplicity. The back-off acoustic model is estimated, stored and served using MapReduce distributed computing infrastructure. Speech recognition experiments are carried out in an N-best list rescoring framework for Google Voice Search. Training big models on large amounts of data proves to be an effective way to increase the accuracy of a state-of-the-art automatic speech recognition system. We use 87,000 hours of training data (speech along with transcription) obtained by filtering utterances in Voice Search logs on automatic speech recognition confidence. Models ranging in size between 20--40 million Gaussians are estimated using maximum likelihood training. They achieve relative reductions in word-error-rate of 11% and 6% when combined with first-pass models trained using maximum likelihood, and boosted maximum mutual information, respectively. Increasing the context size beyond five phones (quinphones) does not help.

Ciprian Chelba, Dan Bikel, Maria Shugrina et al.

Large language models have been proven quite beneficial for a variety of automatic speech recognition tasks in Google. We summarize results on Voice Search and a few YouTube speech transcription tasks to highlight the impact that one can expect from increasing both the amount of training data, and the size of the language model estimated from such data. Depending on the task, availability and amount of training data used, language model size and amount of work and care put into integrating them in the lattice rescoring step we observe reductions in word error rate between 6% and 10% relative, for systems on a wide range of operating points between 17% and 52% word error rate.

Maryam Kamvar, Ciprian Chelba

In this paper, we investigate how to optimize the vocabulary for a voice search language model. The metric we optimize over is the out-of-vocabulary (OoV) rate since it is a strong indicator of user experience. In a departure from the usual way of measuring OoV rates, web search logs allow us to compute the per-session OoV rate and thus estimate the percentage of users that experience a given OoV rate. Under very conservative text normalization, we find that a voice search vocabulary consisting of 2 to 2.5 million words extracted from 1 week of search query data will result in an aggregate OoV rate of 1%; at that size, the same OoV rate will also be experienced by 90% of users. The number of words included in the vocabulary is a stable indicator of the OoV rate. Altering the freshness of the vocabulary or the duration of the time window over which the training data is gathered does not significantly change the OoV rate. Surprisingly, a significantly larger vocabulary (approximately 10 million words) is required to guarantee OoV rates below 1% for 95% of the users.