An emerging recipe for achieving state-of-the-art effectiveness in neural document re-ranking involves utilizing large pre-trained language models - e.g., BERT - to evaluate all individual passages in the document and then aggregating the outputs by pooling or additional Transformer layers. A major drawback of this approach is high query latency due to the cost of evaluating every passage in the document with BERT. To make matters worse, this high inference cost and latency varies based on the length of the document, with longer documents requiring more time and computation. To address this challenge, we adopt an intra-document cascading strategy, which prunes passages of a candidate document using a less expensive model, called ESM, before running a scoring model that is more expensive and effective, called ETM. We found it best to train ESM (short for Efficient Student Model) via knowledge distillation from the ETM (short for Effective Teacher Model) e.g., BERT. This pruning allows us to only run the ETM model on a smaller set of passages whose size does not vary by document length. Our experiments on the MS MARCO and TREC Deep Learning Track benchmarks suggest that the proposed Intra-Document Cascaded Ranking Model (IDCM) leads to over 400% lower query latency by providing essentially the same effectiveness as the state-of-the-art BERT-based document ranking models.
This paper introduces our solution for the Track2 in AI City Challenge 2021 (AICITY21). The Track2 is a vehicle re-identification (ReID) task with both the real-world data and synthetic data. We mainly focus on four points, i.e. training data, unsupervised domain-adaptive (UDA) training, post-processing, model ensembling in this challenge. (1) Both cropping training data and using synthetic data can help the model learn more discriminative features. (2) Since there is a new scenario in the test set that dose not appear in the training set, UDA methods perform well in the challenge. (3) Post-processing techniques including re-ranking, image-to-track retrieval, inter-camera fusion, etc, significantly improve final performance. (4) We ensemble CNN-based models and transformer-based models which provide different representation diversity. With aforementioned techniques, our method finally achieves 0.7445 mAP score, yielding the first place in the competition. Codes are available at https://github.com/michuanhaohao/AICITY2021_Track2_DMT.
Societal biases resonate in the retrieved contents of information retrieval (IR) systems, resulting in reinforcing existing stereotypes. Approaching this issue requires established measures of fairness in respect to the representation of various social groups in retrieval results, as well as methods to mitigate such biases, particularly in the light of the advances in deep ranking models. In this work, we first provide a novel framework to measure the fairness in the retrieved text contents of ranking models. Introducing a ranker-agnostic measurement, the framework also enables the disentanglement of the effect on fairness of collection from that of rankers. To mitigate these biases, we propose AdvBert, a ranking model achieved by adapting adversarial bias mitigation for IR, which jointly learns to predict relevance and remove protected attributes. We conduct experiments on two passage retrieval collections (MSMARCO Passage Re-ranking and TREC Deep Learning 2019 Passage Re-ranking), which we extend by fairness annotations of a selected subset of queries regarding gender attributes. Our results on the MSMARCO benchmark show that, (1) all ranking models are less fair in comparison with ranker-agnostic baselines, and (2) the fairness of Bert rankers significantly improves when using the proposed AdvBert models. Lastly, we investigate the trade-off between fairness and utility, showing that we can maintain the significant improvements in fairness without any significant loss in utility.
Neural information retrieval systems typically use a cascading pipeline, in which a first-stage model retrieves a candidate set of documents and one or more subsequent stages re-rank this set using contextualized language models such as BERT. In this paper, we propose DeepImpact, a new document term-weighting scheme suitable for efficient retrieval using a standard inverted index. Compared to existing methods, DeepImpact improves impact-score modeling and tackles the vocabulary-mismatch problem. In particular, DeepImpact leverages DocT5Query to enrich the document collection and, using a contextualized language model, directly estimates the semantic importance of tokens in a document, producing a single-value representation for each token in each document. Our experiments show that DeepImpact significantly outperforms prior first-stage retrieval approaches by up to 17% on effectiveness metrics w.r.t. DocT5Query, and, when deployed in a re-ranking scenario, can reach the same effectiveness of state-of-the-art approaches with up to 5.1x speedup in efficiency.
As a highly data-driven application, recommender systems could be affected by data bias, resulting in unfair results for different data groups, which could be a reason that affects the system performance. Therefore, it is important to identify and solve the unfairness issues in recommendation scenarios. In this paper, we address the unfairness problem in recommender systems from the user perspective. We group users into advantaged and disadvantaged groups according to their level of activity, and conduct experiments to show that current recommender systems will behave unfairly between two groups of users. Specifically, the advantaged users (active) who only account for a small proportion in data enjoy much higher recommendation quality than those disadvantaged users (inactive). Such bias can also affect the overall performance since the disadvantaged users are the majority. To solve this problem, we provide a re-ranking approach to mitigate this unfairness problem by adding constraints over evaluation metrics. The experiments we conducted on several real-world datasets with various recommendation algorithms show that our approach can not only improve group fairness of users in recommender systems, but also achieve better overall recommendation performance.
Existing neural information retrieval (IR) models have often been studied in homogeneous and narrow settings, which has considerably limited insights into their out-of-distribution (OOD) generalization capabilities. To address this, and to facilitate researchers to broadly evaluate the effectiveness of their models, we introduce Benchmarking-IR (BEIR), a robust and heterogeneous evaluation benchmark for information retrieval. We leverage a careful selection of 18 publicly available datasets from diverse text retrieval tasks and domains and evaluate 10 state-of-the-art retrieval systems including lexical, sparse, dense, late-interaction and re-ranking architectures on the BEIR benchmark. Our results show BM25 is a robust baseline and re-ranking and late-interaction-based models on average achieve the best zero-shot performances, however, at high computational costs. In contrast, dense and sparse-retrieval models are computationally more efficient but often underperform other approaches, highlighting the considerable room for improvement in their generalization capabilities. We hope this framework allows us to better evaluate and understand existing retrieval systems, and contributes to accelerating progress towards better robust and generalizable systems in the future. BEIR is publicly available at https://github.com/UKPLab/beir.
Learning sentence embeddings often requires a large amount of labeled data. However, for most tasks and domains, labeled data is seldom available and creating it is expensive. In this work, we present a new state-of-the-art unsupervised method based on pre-trained Transformers and Sequential Denoising Auto-Encoder (TSDAE) which outperforms previous approaches by up to 6.4 points. It can achieve up to 93.1% of the performance of in-domain supervised approaches. Further, we show that TSDAE is a strong domain adaptation and pre-training method for sentence embeddings, significantly outperforming other approaches like Masked Language Model. A crucial shortcoming of previous studies is the narrow evaluation: Most work mainly evaluates on the single task of Semantic Textual Similarity (STS), which does not require any domain knowledge. It is unclear if these proposed methods generalize to other domains and tasks. We fill this gap and evaluate TSDAE and other recent approaches on four different datasets from heterogeneous domains.
A vital step towards the widespread adoption of neural retrieval models is their resource efficiency throughout the training, indexing and query workflows. The neural IR community made great advancements in training effective dual-encoder dense retrieval (DR) models recently. A dense text retrieval model uses a single vector representation per query and passage to score a match, which enables low-latency first stage retrieval with a nearest neighbor search. Increasingly common, training approaches require enormous compute power, as they either conduct negative passage sampling out of a continuously updating refreshing index or require very large batch sizes for in-batch negative sampling. Instead of relying on more compute capability, we introduce an efficient topic-aware query and balanced margin sampling technique, called TAS-Balanced. We cluster queries once before training and sample queries out of a cluster per batch. We train our lightweight 6-layer DR model with a novel dual-teacher supervision that combines pairwise and in-batch negative teachers. Our method is trainable on a single consumer-grade GPU in under 48 hours (as opposed to a common configuration of 8x V100s). We show that our TAS-Balanced training method achieves state-of-the-art low-latency (64ms per query) results on two TREC Deep Learning Track query sets. Evaluated on NDCG@10, we outperform BM25 by 44%, a plainly trained DR by 19%, docT5query by 11%, and the previous best DR model by 5%. Additionally, TAS-Balanced produces the first dense retriever that outperforms every other method on recall at any cutoff on TREC-DL and allows more resource intensive re-ranking models to operate on fewer passages to improve results further.
Person Re-Identification (ReID) across non-overlapping cameras is a challenging task and, for this reason, most works in the prior art rely on supervised feature learning from a labeled dataset to match the same person in different views. However, it demands the time-consuming task of labeling the acquired data, prohibiting its fast deployment, specially in forensic scenarios. Unsupervised Domain Adaptation (UDA) emerges as a promising alternative, as it performs feature-learning adaptation from a model trained on a source to a target domain without identity-label annotation. However, most UDA-based algorithms rely upon a complex loss function with several hyper-parameters, which hinders the generalization to different scenarios. Moreover, as UDA depends on the translation between domains, it is important to select the most reliable data from the unseen domain, thus avoiding error propagation caused by noisy examples on the target data -- an often overlooked problem. In this sense, we propose a novel UDA-based ReID method that optimizes a simple loss function with only one hyper-parameter and that takes advantage of triplets of samples created by a new offline strategy based on the diversity of cameras within a cluster. This new strategy adapts the model and also regularizes it, avoiding overfitting on the target domain. We also introduce a new self-ensembling strategy, in which weights from different iterations are aggregated to create a final model combining knowledge from distinct moments of the adaptation. For evaluation, we consider three well-known deep learning architectures and combine them for final decision-making. The proposed method does not use person re-ranking nor any label on the target domain, and outperforms the state of the art, with a much simpler setup, on the Market to Duke, the challenging Market1501 to MSMT17, and Duke to MSMT17 adaptation scenarios.
Multimodal pre-training has propelled great advancement in vision-and-language research. These large-scale pre-trained models, although successful, fatefully suffer from slow inference speed due to enormous computation cost mainly from cross-modal attention in Transformer architecture. When applied to real-life applications, such latency and computation demand severely deter the practical use of pre-trained models. In this paper, we study Image-text retrieval (ITR), the most mature scenario of V+L application, which has been widely studied even prior to the emergence of recent pre-trained models. We propose a simple yet highly effective approach, LightningDOT that accelerates the inference time of ITR by thousands of times, without sacrificing accuracy. LightningDOT removes the time-consuming cross-modal attention by pre-training on three novel learning objectives, extracting feature indexes offline, and employing instant dot-product matching with further re-ranking, which significantly speeds up retrieval process. In fact, LightningDOT achieves new state of the art across multiple ITR benchmarks such as Flickr30k, COCO and Multi30K, outperforming existing pre-trained models that consume 1000x magnitude of computational hours. Code and pre-training checkpoints are available at https://github.com/intersun/LightningDOT.
Multi-stage ranking pipelines have been a practical solution in modern search systems, where the first-stage retrieval is to return a subset of candidate documents, and latter stages attempt to re-rank those candidates. Unlike re-ranking stages going through quick technique shifts during past decades, the first-stage retrieval has long been dominated by classical term-based models. Unfortunately, these models suffer from the vocabulary mismatch problem, which may block re-ranking stages from relevant documents at the very beginning. Therefore, it has been a long-term desire to build semantic models for the first-stage retrieval that can achieve high recall efficiently. Recently, we have witnessed an explosive growth of research interests on the first-stage semantic retrieval models. We believe it is the right time to survey current status, learn from existing methods, and gain some insights for future development. In this paper, we describe the current landscape of the first-stage retrieval models under a unified framework to clarify the connection between classical term-based retrieval methods, early semantic retrieval methods and neural semantic retrieval methods. Moreover, we identify some open challenges and envision some future directions, with the hope of inspiring more researches on these important yet less investigated topics.
Supervised machine learning models and their evaluation strongly depends on the quality of the underlying dataset. When we search for a relevant piece of information it may appear anywhere in a given passage. However, we observe a bias in the position of the correct answer in the text in two popular Question Answering datasets used for passage re-ranking. The excessive favoring of earlier positions inside passages is an unwanted artefact. This leads to three common Transformer-based re-ranking models to ignore relevant parts in unseen passages. More concerningly, as the evaluation set is taken from the same biased distribution, the models overfitting to that bias overestimate their true effectiveness. In this work we analyze position bias on datasets, the contextualized representations, and their effect on retrieval results. We propose a debiasing method for retrieval datasets. Our results show that a model trained on a position-biased dataset exhibits a significant decrease in re-ranking effectiveness when evaluated on a debiased dataset. We demonstrate that by mitigating the position bias, Transformer-based re-ranking models are equally effective on a biased and debiased dataset, as well as more effective in a transfer-learning setting between two differently biased datasets.
Machine translation systems are vulnerable to domain mismatch, especially in a low-resource scenario. Out-of-domain translations are often of poor quality and prone to hallucinations, due to exposure bias and the decoder acting as a language model. We adopt two approaches to alleviate this problem: lexical shortlisting restricted by IBM statistical alignments, and hypothesis re-ranking based on similarity. The methods are computationally cheap, widely known, but not extensively experimented on domain adaptation. We demonstrate success on low-resource out-of-domain test sets, however, the methods are ineffective when there is sufficient data or too great domain mismatch. This is due to both the IBM model losing its advantage over the implicitly learned neural alignment, and issues with subword segmentation of out-of-domain words.
Explainable question answering for science questions is a challenging task that requires multi-hop inference over a large set of fact sentences. To counter the limitations of methods that view each query-document pair in isolation, we propose the LSTM-Interleaved Transformer which incorporates cross-document interactions for improved multi-hop ranking. The LIT architecture can leverage prior ranking positions in the re-ranking setting. Our model is competitive on the current leaderboard for the TextGraphs 2020 shared task, achieving a test-set MAP of 0.5607, and would have gained third place had we submitted before the competition deadline. Our code implementation is made available at https://github.com/mdda/worldtree_corpus/tree/textgraphs_2020
This short document describes a traditional IR system that achieved MRR@100 equal to 0.298 on the MS MARCO Document Ranking leaderboard (on 2020-12-06). Although inferior to most BERT-based models, it outperformed several neural runs (as well as all non-neural ones), including two submissions that used a large pretrained Transformer model for re-ranking. We provide software and data to reproduce our results.
The re-ranking approach leverages high-confidence retrieved samples to refine retrieval results, which have been widely adopted as a post-processing tool for image retrieval tasks. However, we notice one main flaw of re-ranking, i.e., high computational complexity, which leads to an unaffordable time cost for real-world applications. In this paper, we revisit re-ranking and demonstrate that re-ranking can be reformulated as a high-parallelism Graph Neural Network (GNN) function. In particular, we divide the conventional re-ranking process into two phases, i.e., retrieving high-quality gallery samples and updating features. We argue that the first phase equals building the k-nearest neighbor graph, while the second phase can be viewed as spreading the message within the graph. In practice, GNN only needs to concern vertices with the connected edges. Since the graph is sparse, we can efficiently update the vertex features. On the Market-1501 dataset, we accelerate the re-ranking processing from 89.2s to 9.4ms with one K40m GPU, facilitating the real-time post-processing. Similarly, we observe that our method achieves comparable or even better retrieval results on the other four image retrieval benchmarks, i.e., VeRi-776, Oxford-5k, Paris-6k and University-1652, with limited time cost. Our code is publicly available.
Rank fusion is a powerful technique that allows multiple sources of information to be combined into a single result set. However, to date fusion has not been regarded as being cost-effective in cases where strict per-query efficiency guarantees are required, such as in web search. In this work we propose a novel solution to rank fusion by splitting the computation into two parts -- one phase that is carried out offline to generate pre-computed centroid answers for queries with broadly similar information needs, and then a second online phase that uses the corresponding topic centroid to compute a result page for each query. We explore efficiency improvements to classic fusion algorithms whose costs can be amortized as a pre-processing step, and can then be combined with re-ranking approaches to dramatically improve effectiveness in multi-stage retrieval systems with little efficiency overhead at query time. Experimental results using the ClueWeb12B collection and the UQV100 query variations demonstrate that centroid-based approaches allow improved retrieval effectiveness at little or no loss in query throughput or latency, and with reasonable pre-processing requirements. We additionally show that queries that do not match any of the pre-computed clusters can be accurately identified and efficiently processed in our proposed ranking pipeline.
Our objective is to introduce to the NLP community an existing k-NN search library NMSLIB, a new retrieval toolkit FlexNeuART, as well as their integration capabilities. NMSLIB, while being one the fastest k-NN search libraries, is quite generic and supports a variety of distance/similarity functions. Because the library relies on the distance-based structure-agnostic algorithms, it can be further extended by adding new distances. FlexNeuART is a modular, extendible and flexible toolkit for candidate generation in IR and QA applications, which supports mixing of classic and neural ranking signals. FlexNeuART can efficiently retrieve mixed dense and sparse representations (with weights learned from training data), which is achieved by extending NMSLIB. In that, other retrieval systems work with purely sparse representations (e.g., Lucene), purely dense representations (e.g., FAISS and Annoy), or only perform mixing at the re-ranking stage.
Many recent approaches towards neural information retrieval mitigate their computational costs by using a multi-stage ranking pipeline. In the first stage, a number of potentially relevant candidates are retrieved using an efficient retrieval model such as BM25. Although BM25 has proven decent performance as a first-stage ranker, it tends to miss relevant passages. In this context we propose CoRT, a simple neural first-stage ranking model that leverages contextual representations from pretrained language models such as BERT to complement term-based ranking functions while causing no significant delay at query time. Using the MS MARCO dataset, we show that CoRT significantly increases the candidate recall by complementing BM25 with missing candidates. Consequently, we find subsequent re-rankers achieve superior results with less candidates. We further demonstrate that passage retrieval using CoRT can be realized with surprisingly low latencies.
Retrieval and ranking models are the backbone of many applications such as web search, open domain QA, or text-based recommender systems. The latency of neural ranking models at query time is largely dependent on the architecture and deliberate choices by their designers to trade-off effectiveness for higher efficiency. This focus on low query latency of a rising number of efficient ranking architectures make them feasible for production deployment. In machine learning an increasingly common approach to close the effectiveness gap of more efficient models is to apply knowledge distillation from a large teacher model to a smaller student model. We find that different ranking architectures tend to produce output scores in different magnitudes. Based on this finding, we propose a cross-architecture training procedure with a margin focused loss (Margin-MSE), that adapts knowledge distillation to the varying score output distributions of different BERT and non-BERT passage ranking architectures. We apply the teachable information as additional fine-grained labels to existing training triples of the MSMARCO-Passage collection. We evaluate our procedure of distilling knowledge from state-of-the-art concatenated BERT models to four different efficient architectures (TK, ColBERT, PreTT, and a BERT CLS dot product model). We show that across our evaluated architectures our Margin-MSE knowledge distillation significantly improves re-ranking effectiveness without compromising their efficiency. Additionally, we show our general distillation method to improve nearest neighbor based index retrieval with the BERT dot product model, offering competitive results with specialized and much more costly training methods. To benefit the community, we publish the teacher-score training files in a ready-to-use package.