Search

Retrieval-augmented generation improves various aspects of large language models (LLMs) generation, but suffers from computational overhead caused by long contexts as well as the propagation of irrelevant retrieved information into generated responses. Context pruning deals with both aspects, by removing irrelevant parts of retrieved contexts before LLM generation. Existing context pruning approaches are however limited, and do not provide a universal model that would be both efficient and robust in a wide range of scenarios, e.g., when contexts contain a variable amount of relevant information or vary in length, or when evaluated on various domains. In this work, we close this gap and introduce Provence (Pruning and Reranking Of retrieVEd relevaNt ContExts), an efficient and robust context pruner for Question Answering, which dynamically detects the needed amount of pruning for a given context and can be used out-of-the-box for various domains. The three key ingredients of Provence are formulating the context pruning task as sequence labeling, unifying context pruning capabilities with context reranking, and training on diverse data. Our experimental results show that Provence enables context pruning with negligible to no drop in performance, in various domains and settings, at almost no cost in a standard RAG pipeline. We also conduct a deeper analysis alongside various ablations to provide insights into training context pruners for future work., Comment: Accepted to ICLR 2025

Retrieval-Augmented Generation (RAG) pipelines enhance Large Language Models (LLMs) by retrieving relevant documents, but they face scalability issues due to high inference costs and limited context size. Document compression is a practical solution, but current soft compression methods suffer from accuracy losses and require extensive pretraining. In this paper, we introduce PISCO, a novel method that achieves a 16x compression rate with minimal accuracy loss (0-3%) across diverse RAG-based question-answering (QA) tasks. Unlike existing approaches, PISCO requires no pretraining or annotated data, relying solely on sequence-level knowledge distillation from document-based questions. With the ability to fine-tune a 7-10B LLM in 48 hours on a single A100 GPU, PISCO offers a highly efficient and scalable solution. We present comprehensive experiments showing that PISCO outperforms existing compression models by 8% in accuracy.

Retrieval-Augmented Generation (RAG) allows overcoming the limited knowledge of LLMs by extending the input with external information. As a consequence, the contextual inputs to the model become much longer which slows down decoding time directly translating to the time a user has to wait for an answer. We address this challenge by presenting COCOM, an effective context compression method, reducing long contexts to only a handful of Context Embeddings speeding up the generation time by a large margin. Our method allows for different compression rates trading off decoding time for answer quality. Compared to earlier methods, COCOM allows for handling multiple contexts more effectively, significantly reducing decoding time for long inputs. Our method demonstrates a speed-up of up to 5.69 $\times$ while achieving higher performance compared to existing efficient context compression methods., Comment: 10 pages

Retrieval-Augmented Generation allows to enhance Large Language Models with external knowledge. In response to the recent popularity of generative LLMs, many RAG approaches have been proposed, which involve an intricate number of different configurations such as evaluation datasets, collections, metrics, retrievers, and LLMs. Inconsistent benchmarking poses a major challenge in comparing approaches and understanding the impact of each component in the pipeline. In this work, we study best practices that lay the groundwork for a systematic evaluation of RAG and present BERGEN, an end-to-end library for reproducible research standardizing RAG experiments. In an extensive study focusing on QA, we benchmark different state-of-the-art retrievers, rerankers, and LLMs. Additionally, we analyze existing RAG metrics and datasets. Our open-source library BERGEN is available under \url{https://github.com/naver/bergen}., Comment: 29 pages

Retrieval-augmented generation (RAG) has recently emerged as a promising solution for incorporating up-to-date or domain-specific knowledge into large language models (LLMs) and improving LLM factuality, but is predominantly studied in English-only settings. In this work, we consider RAG in the multilingual setting (mRAG), i.e. with user queries and the datastore in 13 languages, and investigate which components and with which adjustments are needed to build a well-performing mRAG pipeline, that can be used as a strong baseline in future works. Our findings highlight that despite the availability of high-quality off-the-shelf multilingual retrievers and generators, task-specific prompt engineering is needed to enable generation in user languages. Moreover, current evaluation metrics need adjustments for multilingual setting, to account for variations in spelling named entities. The main limitations to be addressed in future works include frequent code-switching in non-Latin alphabet languages, occasional fluency errors, wrong reading of the provided documents, or irrelevant retrieval. We release the code for the resulting mRAG baseline pipeline at https://github.com/naver/bergen.

The late interaction paradigm introduced with ColBERT stands out in the neural Information Retrieval space, offering a compelling effectiveness-efficiency trade-off across many benchmarks. Efficient late interaction retrieval is based on an optimized multi-step strategy, where an approximate search first identifies a set of candidate documents to re-rank exactly. In this work, we introduce SPLATE, a simple and lightweight adaptation of the ColBERTv2 model which learns an ``MLM adapter'', mapping its frozen token embeddings to a sparse vocabulary space with a partially learned SPLADE module. This allows us to perform the candidate generation step in late interaction pipelines with traditional sparse retrieval techniques, making it particularly appealing for running ColBERT in CPU environments. Our SPLATE ColBERTv2 pipeline achieves the same effectiveness as the PLAID ColBERTv2 engine by re-ranking 50 documents that can be retrieved under 10ms., Comment: To appear at SIGIR'24 (short paper track)

Learned sparse models such as SPLADE have successfully shown how to incorporate the benefits of state-of-the-art neural information retrieval models into the classical inverted index data structure. Despite their improvements in effectiveness, learned sparse models are not as efficient as classical sparse model such as BM25. The problem has been investigated and addressed by recently developed strategies, such as guided traversal query processing and static pruning, with different degrees of success on in-domain and out-of-domain datasets. In this work, we propose a new query processing strategy for SPLADE based on a two-step cascade. The first step uses a pruned and reweighted version of the SPLADE sparse vectors, and the second step uses the original SPLADE vectors to re-score a sample of documents retrieved in the first stage. Our extensive experiments, performed on 30 different in-domain and out-of-domain datasets, show that our proposed strategy is able to improve mean and tail response times over the original single-stage SPLADE processing by up to $30\times$ and $40\times$, respectively, for in-domain datasets, and by 12x to 25x, for mean response on out-of-domain datasets, while not incurring in statistical significant difference in 60\% of datasets., Comment: published in Findings at ECIR'24

We present a comparative study between cross-encoder and LLMs rerankers in the context of re-ranking effective SPLADE retrievers. We conduct a large evaluation on TREC Deep Learning datasets and out-of-domain datasets such as BEIR and LoTTE. In the first set of experiments, we show how cross-encoder rerankers are hard to distinguish when it comes to re-rerank SPLADE on MS MARCO. Observations shift in the out-of-domain scenario, where both the type of model and the number of documents to re-rank have an impact on effectiveness. Then, we focus on listwise rerankers based on Large Language Models -- especially GPT-4. While GPT-4 demonstrates impressive (zero-shot) performance, we show that traditional cross-encoders remain very competitive. Overall, our findings aim to to provide a more nuanced perspective on the recent excitement surrounding LLM-based re-rankers -- by positioning them as another factor to consider in balancing effectiveness and efficiency in search systems.

A companion to the release of the latest version of the SPLADE library. We describe changes to the training structure and present our latest series of models -- SPLADE-v3. We compare this new version to BM25, SPLADE++, as well as re-rankers, and showcase its effectiveness via a meta-analysis over more than 40 query sets. SPLADE-v3 further pushes the limit of SPLADE models: it is statistically significantly more effective than both BM25 and SPLADE++, while comparing well to cross-encoder re-rankers. Specifically, it gets more than 40 MRR@10 on the MS MARCO dev set, and improves by 2% the out-of-domain results on the BEIR benchmark., Comment: Technical report

Middle training methods aim to bridge the gap between the Masked Language Model (MLM) pre-training and the final finetuning for retrieval. Recent models such as CoCondenser, RetroMAE, and LexMAE argue that the MLM task is not sufficient enough to pre-train a transformer network for retrieval and hence propose various tasks to do so. Intrigued by those novel methods, we noticed that all these models used different finetuning protocols, making it hard to assess the benefits of middle training. We propose in this paper a benchmark of CoCondenser, RetroMAE, and LexMAE, under the same finetuning conditions. We compare both dense and sparse approaches under various finetuning protocols and middle training on different collections (MS MARCO, Wikipedia or Tripclick). We use additional middle training baselines, such as a standard MLM finetuning on the retrieval collection, optionally augmented by a CLS predicting the passage term frequency. For the sparse approach, our study reveals that there is almost no statistical difference between those methods: the more effective the finetuning procedure is, the less difference there is between those models. For the dense approach, RetroMAE using MS MARCO as middle-training collection shows excellent results in almost all the settings. Finally, we show that middle training on the retrieval collection, thus adapting the language model to it, is a critical factor. Overall, a better experimental setup should be adopted to evaluate middle training methods. Code available at https://github.com/naver/splade/tree/benchmarch-SIGIR23

The MS MARCO-passage dataset has been the main large-scale dataset open to the IR community and it has fostered successfully the development of novel neural retrieval models over the years. But, it turns out that two different corpora of MS MARCO are used in the literature, the official one and a second one where passages were augmented with titles, mostly due to the introduction of the Tevatron code base. However, the addition of titles actually leaks relevance information, while breaking the original guidelines of the MS MARCO-passage dataset. In this work, we investigate the differences between the two corpora and demonstrate empirically that they make a significant difference when evaluating a new method. In other words, we show that if a paper does not properly report which version is used, reproducing fairly its results is basically impossible. Furthermore, given the current status of reviewing, where monitoring state-of-the-art results is of great importance, having two different versions of a dataset is a large problem. This is why this paper aims to report the importance of this issue so that researchers can be made aware of this problem and appropriately report their results., Comment: Short paper accepted at SIGIR 2023

Sparse neural retrievers, such as DeepImpact, uniCOIL and SPLADE, have been introduced recently as an efficient and effective way to perform retrieval with inverted indexes. They aim to learn term importance and, in some cases, document expansions, to provide a more effective document ranking compared to traditional bag-of-words retrieval models such as BM25. However, these sparse neural retrievers have been shown to increase the computational costs and latency of query processing compared to their classical counterparts. To mitigate this, we apply a well-known family of techniques for boosting the efficiency of query processing over inverted indexes: static pruning. We experiment with three static pruning strategies, namely document-centric, term-centric and agnostic pruning, and we assess, over diverse datasets, that these techniques still work with sparse neural retrievers. In particular, static pruning achieves $2\times$ speedup with negligible effectiveness loss ($\leq 2\%$ drop) and, depending on the use case, even $4\times$ speedup with minimal impact on the effectiveness ($\leq 8\%$ drop). Moreover, we show that neural rerankers are robust to candidates from statically pruned indexes., Comment: Short accepted at SIGIR 2023. Has extra content in the Appendix that is not present in the ACM version

This paper presents the AToMiC (Authoring Tools for Multimedia Content) dataset, designed to advance research in image/text cross-modal retrieval. While vision-language pretrained transformers have led to significant improvements in retrieval effectiveness, existing research has relied on image-caption datasets that feature only simplistic image-text relationships and underspecified user models of retrieval tasks. To address the gap between these oversimplified settings and real-world applications for multimedia content creation, we introduce a new approach for building retrieval test collections. We leverage hierarchical structures and diverse domains of texts, styles, and types of images, as well as large-scale image-document associations embedded in Wikipedia. We formulate two tasks based on a realistic user model and validate our dataset through retrieval experiments using baseline models. AToMiC offers a testbed for scalable, diverse, and reproducible multimedia retrieval research. Finally, the dataset provides the basis for a dedicated track at the 2023 Text Retrieval Conference (TREC), and is publicly available at https://github.com/TREC-AToMiC/AToMiC.

Parameter-Efficient transfer learning with Adapters have been studied in Natural Language Processing (NLP) as an alternative to full fine-tuning. Adapters are memory-efficient and scale well with downstream tasks by training small bottle-neck layers added between transformer layers while keeping the large pretrained language model (PLMs) frozen. In spite of showing promising results in NLP, these methods are under-explored in Information Retrieval. While previous studies have only experimented with dense retriever or in a cross lingual retrieval scenario, in this paper we aim to complete the picture on the use of adapters in IR. First, we study adapters for SPLADE, a sparse retriever, for which adapters not only retain the efficiency and effectiveness otherwise achieved by finetuning, but are memory-efficient and orders of magnitude lighter to train. We observe that Adapters-SPLADE not only optimizes just 2\% of training parameters, but outperforms fully fine-tuned counterpart and existing parameter-efficient dense IR models on IR benchmark datasets. Secondly, we address domain adaptation of neural retrieval thanks to adapters on cross-domain BEIR datasets and TripClick. Finally, we also consider knowledge sharing between rerankers and first stage rankers. Overall, our study complete the examination of adapters for neural IR, Comment: accepted at ECIR'23

This paper describes our participation in the 2022 TREC NeuCLIR challenge. We submitted runs to two out of the three languages (Farsi and Russian), with a focus on first-stage rankers and comparing mono-lingual strategies to Adhoc ones. For monolingual runs, we start from pretraining models on the target language using MLM+FLOPS and then finetuning using the MSMARCO translated to the language either with ColBERT or SPLADE as the retrieval model. While for the Adhoc task, we test both query translation (to the target language) and back-translation of the documents (to English). Initial result analysis shows that the monolingual strategy is strong, but that for the moment Adhoc achieved the best results, with back-translating documents being better than translating queries., Comment: Notebook detailing our participation and analysis on the TREC NeuCLIR 2022

This paper describes our participation to the 2022 TREC Deep Learning challenge. We submitted runs to all four tasks, with a focus on the full retrieval passage task. The strategy is almost the same as 2021, with first stage retrieval being based around SPLADE, with some added ensembling with ColBERTv2 and DocT5. We also use the same strategy of last year for the second stage, with an ensemble of re-rankers trained using hard negatives selected by SPLADE. Initial result analysis show that the strategy is still strong, but is still unclear to us what next steps should we take.

Evaluation in Information Retrieval relies on post-hoc empirical procedures, which are time-consuming and expensive operations. To alleviate this, Query Performance Prediction (QPP) models have been developed to estimate the performance of a system without the need for human-made relevance judgements. Such models, usually relying on lexical features from queries and corpora, have been applied to traditional sparse IR methods - with various degrees of success. With the advent of neural IR and large Pre-trained Language Models, the retrieval paradigm has significantly shifted towards more semantic signals. In this work, we study and analyze to what extent current QPP models can predict the performance of such systems. Our experiments consider seven traditional bag-of-words and seven BERT-based IR approaches, as well as nineteen state-of-the-art QPPs evaluated on two collections, Deep Learning '19 and Robust '04. Our findings show that QPPs perform statistically significantly worse on neural IR systems. In settings where semantic signals are prominent (e.g., passage retrieval), their performance on neural models drops by as much as 10% compared to bag-of-words approaches. On top of that, in lexical-oriented scenarios, QPPs fail to predict performance for neural IR systems on those queries where they differ from traditional approaches the most.

Neural retrieval models have acquired significant effectiveness gains over the last few years compared to term-based methods. Nevertheless, those models may be brittle when faced to typos, distribution shifts or vulnerable to malicious attacks. For instance, several recent papers demonstrated that such variations severely impacted models performances, and then tried to train more resilient models. Usual approaches include synonyms replacements or typos injections -- as data-augmentation -- and the use of more robust tokenizers (characterBERT, BPE-dropout). To further complement the literature, we investigate in this paper adversarial training as another possible solution to this robustness issue. Our comparison includes the two main families of BERT-based neural retrievers, i.e. dense and sparse, with and without distillation techniques. We then demonstrate that one of the most simple adversarial training techniques -- the Fast Gradient Sign Method (FGSM) -- can improve first stage rankers robustness and effectiveness. In particular, FGSM increases models performances on both in-domain and out-of-domain distributions, and also on queries with typos, for multiple neural retrievers., Comment: Accepted at ECIR 2023

Finetuning Pretrained Language Models (PLM) for IR has been de facto the standard practice since their breakthrough effectiveness few years ago. But, is this approach well understood? In this paper, we study the impact of the pretraining collection on the final IR effectiveness. In particular, we challenge the current hypothesis that PLM shall be trained on a large enough generic collection and we show that pretraining from scratch on the collection of interest is surprisingly competitive with the current approach. We benchmark first-stage ranking rankers and cross-encoders for reranking on the task of general passage retrieval on MSMARCO, Mr-Tydi for Arabic, Japanese and Russian, and TripClick for specific domain. Contrary to popular belief, we show that, for finetuning first-stage rankers, models pretrained solely on their collection have equivalent or better effectiveness compared to more general models. However, there is a slight effectiveness drop for rerankers pretrained only on the target collection. Overall, our study sheds a new light on the role of the pretraining collection and should make our community ponder on building specialized models by pretraining from scratch. Last but not least, doing so could enable better control of efficiency, data bias and replicability, which are key research questions for the IR community., Comment: Accepted at ECIR 2023

Latency and efficiency issues are often overlooked when evaluating IR models based on Pretrained Language Models (PLMs) in reason of multiple hardware and software testing scenarios. Nevertheless, efficiency is an important part of such systems and should not be overlooked. In this paper, we focus on improving the efficiency of the SPLADE model since it has achieved state-of-the-art zero-shot performance and competitive results on TREC collections. SPLADE efficiency can be controlled via a regularization factor, but solely controlling this regularization has been shown to not be efficient enough. In order to reduce the latency gap between SPLADE and traditional retrieval systems, we propose several techniques including L1 regularization for queries, a separation of document/query encoders, a FLOPS-regularized middle-training, and the use of faster query encoders. Our benchmark demonstrates that we can drastically improve the efficiency of these models while increasing the performance metrics on in-domain data. To our knowledge, {we propose the first neural models that, under the same computing constraints, \textit{achieve similar latency (less than 4ms difference) as traditional BM25}, while having \textit{similar performance (less than 10\% MRR@10 reduction)} as the state-of-the-art single-stage neural rankers on in-domain data}., Comment: Accepted at SIGIR 2022 and ReNeuIR. Huggingface models available at https://huggingface.co/naver and code at https://github.com/naver/splade

Neural retrievers based on dense representations combined with Approximate Nearest Neighbors search have recently received a lot of attention, owing their success to distillation and/or better sampling of examples for training -- while still relying on the same backbone architecture. In the meantime, sparse representation learning fueled by traditional inverted indexing techniques has seen a growing interest, inheriting from desirable IR priors such as explicit lexical matching. While some architectural variants have been proposed, a lesser effort has been put in the training of such models. In this work, we build on SPLADE -- a sparse expansion-based retriever -- and show to which extent it is able to benefit from the same training improvements as dense models, by studying the effect of distillation, hard-negative mining as well as the Pre-trained Language Model initialization. We furthermore study the link between effectiveness and efficiency, on in-domain and zero-shot settings, leading to state-of-the-art results in both scenarios for sufficiently expressive models., Comment: Accepted at SIGIR22 as a short paper (this work is the extension of SPLADE v2)

This paper investigates the Relation Extraction task in documents by benchmarking two different neural network models: a multi-modal language model (LayoutXLM) and a Graph Neural Network: Edge Convolution Network (ECN). For this benchmark, we use the XFUND dataset, released along with LayoutXLM. While both models reach similar results, they both exhibit very different characteristics. This raises the question on how to integrate various modalities in a neural network: by merging all modalities thanks to additional pretraining (LayoutXLM), or in a cascaded way (ECN). We conclude by discussing some methodological issues that must be considered for new datasets and task definition in the domain of Information Extraction with complex documents.

Pre-trained Language Models have recently emerged in Information Retrieval as providing the backbone of a new generation of neural systems that outperform traditional methods on a variety of tasks. However, it is still unclear to what extent such approaches generalize in zero-shot conditions. The recent BEIR benchmark provides partial answers to this question by comparing models on datasets and tasks that differ from the training conditions. We aim to address the same question by comparing models under more explicit distribution shifts. To this end, we build three query-based distribution shifts within MS MARCO (query-semantic, query-intent, query-length), which are used to evaluate the three main families of neural retrievers based on BERT: sparse, dense, and late-interaction -- as well as a monoBERT re-ranker. We further analyse the performance drops between the train and test query distributions. In particular, we experiment with two generalization indicators: the first one based on train/test query vocabulary overlap, and the second based on representations of a trained bi-encoder. Intuitively, those indicators verify that the further away the test set is from the train one, the worse the drop in performance. We also show that models respond differently to the shifts -- dense approaches being the most impacted. Overall, our study demonstrates that it is possible to design more controllable distribution shifts as a tool to better understand generalization of IR models. Finally, we release the MS MARCO query subsets, which provide an additional resource to benchmark zero-shot transfer in Information Retrieval., Comment: Accepted at ECIR 2023

We propose a Composite Code Sparse Autoencoder (CCSA) approach for Approximate Nearest Neighbor (ANN) search of document representations based on Siamese-BERT models. In Information Retrieval (IR), the ranking pipeline is generally decomposed in two stages: the first stage focus on retrieving a candidate set from the whole collection. The second stage re-ranks the candidate set by relying on more complex models. Recently, Siamese-BERT models have been used as first stage ranker to replace or complement the traditional bag-of-word models. However, indexing and searching a large document collection require efficient similarity search on dense vectors and this is why ANN techniques come into play. Since composite codes are naturally sparse, we first show how CCSA can learn efficient parallel inverted index thanks to an uniformity regularizer. Second, CCSA can be used as a binary quantization method and we propose to combine it with the recent graph based ANN techniques. Our experiments on MSMARCO dataset reveal that CCSA outperforms IVF with product quantization. Furthermore, CCSA binary quantization is beneficial for the index size, and memory usage for the graph-based HNSW method, while maintaining a good level of recall and MRR. Third, we compare with recent supervised quantization methods for image retrieval and find that CCSA is able to outperform them., Comment: 24 pages, longer version of short paper accepted at SIGIR 2021

Learning with noisy labels is an active research area for image classification. However, the effect of noisy labels on image retrieval has been less studied. In this work, we propose a noise-resistant method for image retrieval named Teacher-based Selection of Interactions, T-SINT, which identifies noisy interactions, ie. elements in the distance matrix, and selects correct positive and negative interactions to be considered in the retrieval loss by using a teacher-based training setup which contributes to the stability. As a result, it consistently outperforms state-of-the-art methods on high noise rates across benchmark datasets with synthetic noise and more realistic noise., Comment: Accepted at WACV 2022. 13 pages, 5 figures

The ColBERT model has recently been proposed as an effective BERT based ranker. By adopting a late interaction mechanism, a major advantage of ColBERT is that document representations can be precomputed in advance. However, the big downside of the model is the index size, which scales linearly with the number of tokens in the collection. In this paper, we study various designs for ColBERT models in order to attack this problem. While compression techniques have been explored to reduce the index size, in this paper we study token pruning techniques for ColBERT. We compare simple heuristics, as well as a single layer of attention mechanism to select the tokens to keep at indexing time. Our experiments show that ColBERT indexes can be pruned up to 30\% on the MS MARCO passage collection without a significant drop in performance. Finally, we experiment on MS MARCO documents, which reveal several challenges for such mechanism., Comment: 5 pages

Neural Information Retrieval models hold the promise to replace lexical matching models, e.g. BM25, in modern search engines. While their capabilities have fully shone on in-domain datasets like MS MARCO, they have recently been challenged on out-of-domain zero-shot settings (BEIR benchmark), questioning their actual generalization capabilities compared to bag-of-words approaches. Particularly, we wonder if these shortcomings could (partly) be the consequence of the inability of neural IR models to perform lexical matching off-the-shelf. In this work, we propose a measure of discrepancy between the lexical matching performed by any (neural) model and an 'ideal' one. Based on this, we study the behavior of different state-of-the-art neural IR models, focusing on whether they are able to perform lexical matching when it's actually useful, i.e. for important terms. Overall, we show that neural IR models fail to properly generalize term importance on out-of-domain collections or terms almost unseen during training, Comment: Accepted in the 44nd European Conference on Information Retrieval (ECIR'22), short paper track

In neural Information Retrieval (IR), ongoing research is directed towards improving the first retriever in ranking pipelines. Learning dense embeddings to conduct retrieval using efficient approximate nearest neighbors methods has proven to work well. Meanwhile, there has been a growing interest in learning \emph{sparse} representations for documents and queries, that could inherit from the desirable properties of bag-of-words models such as the exact matching of terms and the efficiency of inverted indexes. Introduced recently, the SPLADE model provides highly sparse representations and competitive results with respect to state-of-the-art dense and sparse approaches. In this paper, we build on SPLADE and propose several significant improvements in terms of effectiveness and/or efficiency. More specifically, we modify the pooling mechanism, benchmark a model solely based on document expansion, and introduce models trained with distillation. We also report results on the BEIR benchmark. Overall, SPLADE is considerably improved with more than $9$\% gains on NDCG@10 on TREC DL 2019, leading to state-of-the-art results on the BEIR benchmark., Comment: 5 pages. arXiv admin note: substantial text overlap with arXiv:2107.05720

Multilingual NMT has become an attractive solution for MT deployment in production. But to match bilingual quality, it comes at the cost of larger and slower models. In this work, we consider several ways to make multilingual NMT faster at inference without degrading its quality. We experiment with several "light decoder" architectures in two 20-language multi-parallel settings: small-scale on TED Talks and large-scale on ParaCrawl. Our experiments demonstrate that combining a shallow decoder with vocabulary filtering leads to more than twice faster inference with no loss in translation quality. We validate our findings with BLEU and chrF (on 380 language pairs), robustness evaluation and human evaluation., Comment: EMNLP 2021 long paper

Attacking Neural Machine Translation models is an inherently combinatorial task on discrete sequences, solved with approximate heuristics. Most methods use the gradient to attack the model on each sample independently. Instead of mechanically applying the gradient, could we learn to produce meaningful adversarial attacks ? In contrast to existing approaches, we learn to attack a model by training an adversarial generator based on a language model. We propose the Masked Adversarial Generation (MAG) model, that learns to perturb the translation model throughout the training process. The experiments show that it improves the robustness of machine translation models, while being faster than competing methods., Comment: 5 pages

In neural Information Retrieval, ongoing research is directed towards improving the first retriever in ranking pipelines. Learning dense embeddings to conduct retrieval using efficient approximate nearest neighbors methods has proven to work well. Meanwhile, there has been a growing interest in learning sparse representations for documents and queries, that could inherit from the desirable properties of bag-of-words models such as the exact matching of terms and the efficiency of inverted indexes. In this work, we present a new first-stage ranker based on explicit sparsity regularization and a log-saturation effect on term weights, leading to highly sparse representations and competitive results with respect to state-of-the-art dense and sparse methods. Our approach is simple, trained end-to-end in a single stage. We also explore the trade-off between effectiveness and efficiency, by controlling the contribution of the sparsity regularization., Comment: 5 pages, SIGIR'21 short paper

Transformer-based models are nowadays state-of-the-art in ad-hoc Information Retrieval, but their behavior is far from being understood. Recent work has claimed that BERT does not satisfy the classical IR axioms. However, we propose to dissect the matching process of ColBERT, through the analysis of term importance and exact/soft matching patterns. Even if the traditional axioms are not formally verified, our analysis reveals that ColBERT: (i) is able to capture a notion of term importance; (ii) relies on exact matches for important terms., Comment: to appear in ECIR'21