In text ranking, it is generally believed that the cross-encoders already gather sufficient token interaction information via the attention mechanism in the hidden layers. However, our results show that the cross-encoders can consistently benefit from additional token interaction in the similarity computation at the last layer. We introduce CELI (Cross-Encoder with Late Interaction), which incorporates a late interaction layer into the current cross-encoder models. This simple method brings 5% improvement on BEIR without compromising in-domain effectiveness or search latency. Extensive experiments show that this finding is consistent across different sizes of the cross-encoder models and the first-stage retrievers. Our findings suggest that boiling all information into the [CLS] token is a suboptimal use for cross-encoders, and advocate further studies to investigate its relevance score mechanism.
Recent studies have demonstrated that the ability of dense retrieval models to generalize to target domains with different distributions is limited, which contrasts with the results obtained with interaction-based models. Prior attempts to mitigate this challenge involved leveraging adversarial learning and query generation approaches, but both approaches nevertheless resulted in limited improvements. In this paper, we propose to combine the query-generation approach with a self-supervision approach in which pseudo-relevance labels are automatically generated on the target domain. To accomplish this, a T5-3B model is utilized for pseudo-positive labeling, and meticulous hard negatives are chosen. We also apply this strategy on conversational dense retrieval model for conversational search. A similar pseudo-labeling approach is used, but with the addition of a query-rewriting module to rewrite conversational queries for subsequent labeling. This proposed approach enables a model’s domain adaptation with real queries and documents from the target dataset. Experiments on standard dense retrieval and conversational dense retrieval models both demonstrate improvements on baseline models when they are fine-tuned on the pseudo-relevance labeled data.
Multi-vector retrieval methods combine the merits of sparse (e.g. BM25) and dense (e.g. DPR) retrievers and have achieved state-of-the-art performance on various retrieval tasks. These methods, however, are orders of magnitude slower and need much more space to store their indices compared to their single-vector counterparts. In this paper, we unify different multi-vector retrieval models from a token routing viewpoint and propose conditional token interaction via dynamic lexical routing, namely CITADEL, for efficient and effective multi-vector retrieval.CITADEL learns to route different token vectors to the predicted lexical keys such that a query token vector only interacts with document token vectors routed to the same key. This design significantly reduces the computation cost while maintaining high accuracy. Notably, CITADEL achieves the same or slightly better performance than the previous state of the art, ColBERT-v2, on both in-domain (MS MARCO) and out-of-domain (BEIR) evaluations, while being nearly 40 times faster. Source code and data are available at https://github.com/facebookresearch/dpr-scale/tree/citadel.
Various techniques have been developed in recent years to improve dense retrieval (DR), such as unsupervised contrastive learning and pseudo-query generation. Existing DRs, however, often suffer from effectiveness tradeoffs between supervised and zero-shot retrieval, which some argue was due to the limited model capacity. We contradict this hypothesis and show that a generalizable DR can be trained to achieve high accuracy in both supervised and zero-shot retrieval without increasing model size. In particular, we systematically examine the contrastive learning of DRs, under the framework of Data Augmentation (DA). Our study shows that common DA practices such as query augmentation with generative models and pseudo-relevance label creation using a cross-encoder, are often inefficient and sub-optimal. We hence propose a new DA approach with diverse queries and sources of supervision to progressively train a generalizable DR. As a result, DRAGON, our Dense Retriever trained with diverse AuGmentatiON, is the first BERT-base-sized DR to achieve state-of-the-art effectiveness in both supervised and zero-shot evaluations and even competes with models using more complex late interaction.
Bien que la recherche d’information neuronale ait connu des améliorations, les modèles de recherche dense ont une capacité de généralisation à de nouveaux domaines limitée, contrairement aux modèles basés sur l’interaction. Les approches d’apprentissage adversarial et de génération de requêtes n’ont pas résolu ce problème. Cet article propose une approche d’auto-supervision utilisant des étiquettes de pseudo-pertinence automatiquement générées pour le domaine cible. Le modèle T53B est utilisé pour réordonner une liste de documents fournie par BM25 afin d’obtenir une annotation des exemples positifs. L’extraction des exemples négatifs est effectuée en explorant différentes stratégies. Les expériences montrent que cette approche aide le modèle dense sur le domaine cible et améliore l’approche de génération de requêtes GPL.
Les réseaux neuronaux profonds et les modèles fondés sur les transformeurs comme BERT ont envahi le domaine de la recherche d’informations (RI) ces dernières années. Leur succès est lié au mécanisme d’auto-attention qui permet de capturer les dépendances entre les mots indépendamment de leur distance. Cependant, en raison de sa complexité quadratique dans le nombre de mots, ce mécanisme ne peut être directement utilisé sur de longues séquences, ce qui ne permet pas de déployer entièrement les modèles neuronaux sur des documents longs pouvant contenir des milliers de mots. Trois stratégies standard ont été adoptées pour contourner ce problème. La première consiste à tronquer les documents longs, la deuxième à segmenter les documents longs en passages plus courts et la dernière à remplacer le module d’auto-attention par des modules d’attention parcimonieux. Dans le premier cas, des informations importantes peuvent être perdues et le jugement de pertinence n’est fondé que sur une partie de l’information contenue dans le document. Dans le deuxième cas, une architecture hiérarchique peut être adoptée pour construire une représentation du document sur la base des représentations de chaque passage. Cela dit, malgré ses résultats prometteurs, cette stratégie reste coûteuse en temps, en mémoire et en énergie. Dans le troisième cas, les contraintes de parcimonie peuvent conduire à manquer des dépendances importantes et, in fine, à des résultats sous-optimaux. L’approche que nous proposons est légèrement différente de ces stratégies et vise à capturer, dans les documents longs, les blocs les plus importants permettant de décider du statut, pertinent ou non, de l’ensemble du document. Elle repose sur trois étapes principales : (a) la sélection de blocs clés (c’est-à-dire susceptibles d’être pertinents) avec un pré-classement local en utilisant soit des modèles de RI classiques, soit un module d’apprentissage, (b) l’apprentissage d’une représentation conjointe des requêtes et des blocs clés à l’aide d’un modèle BERT standard, et (c) le calcul d’un score de pertinence final qui peut être considéré comme une agrégation d’informations de pertinence locale. Dans cet article, nous menons tout d’abord une analyse qui révèle que les signaux de pertinence peuvent apparaître à différents endroits dans les documents et que de tels signaux sont mieux capturés par des relations sémantiques que par des correspondances exactes. Nous examinons ensuite plusieurs méthodes pour sélectionner les blocs pertinents et montrons comment intégrer ces méthodes dans les modèles récents de RI.
Pre-trained language models have been successful in many knowledge-intensive NLP tasks. However, recent work has shown that models such as BERT are not “structurally ready” to aggregate textual information into a [CLS] vector for dense passage retrieval (DPR). This “lack of readiness” results from the gap between language model pre-training and DPR fine-tuning. Previous solutions call for computationally expensive techniques such as hard negative mining, cross-encoder distillation, and further pre-training to learn a robust DPR model. In this work, we instead propose to fully exploit knowledge in a pre-trained language model for DPR by aggregating the contextualized token embeddings into a dense vector, which we call agg★. By concatenating vectors from the [CLS] token and agg★, our Aggretriever model substantially improves the effectiveness of dense retrieval models on both in-domain and zero-shot evaluations without introducing substantial training overhead. Code is available at https://github.com/castorini/dhr.
In information retrieval (IR), candidate set pruning has been commonly used to speed up two-stage relevance ranking. However, such an approach lacks accurate error control and often trades accuracy against computational efficiency in an empirical fashion, missing theoretical guarantees. In this paper, we propose the concept of certified error control of candidate set pruning for relevance ranking, which means that the test error after pruning is guaranteed to be controlled under a user-specified threshold with high probability. Both in-domain and out-of-domain experiments show that our method successfully prunes the first-stage retrieved candidate sets to improve the second-stage reranking speed while satisfying the pre-specified accuracy constraints in both settings. For example, on MS MARCO Passage v1, our method reduces the average candidate set size from 1000 to 27, increasing reranking speed by about 37 times, while keeping MRR@10 greater than a pre-specified value of 0.38 with about 90% empirical coverage. In contrast, empirical baselines fail to meet such requirements. Code and data are available at: https://github.com/alexlimh/CEC-Ranking.
The bi-encoder design of dense passage retriever (DPR) is a key factor to its success in open-domain question answering (QA), yet it is unclear how DPR’s question encoder and passage encoder individually contributes to overall performance, which we refer to as the encoder attribution problem. The problem is important as it helps us identify the factors that affect individual encoders to further improve overall performance. In this paper, we formulate our analysis under a probabilistic framework called encoder marginalization, where we quantify the contribution of a single encoder by marginalizing other variables. First, we find that the passage encoder contributes more than the question encoder to in-domain retrieval accuracy. Second, we demonstrate how to find the affecting factors for each encoder, where we train DPR with different amounts of data and use encoder marginalization to analyze the results. We find that positive passage overlap and corpus coverage of training data have big impacts on the passage encoder, while the question encoder is mainly affected by training sample complexity under this setting. Based on this framework, we can devise data-efficient training regimes: for example, we manage to train a passage encoder on SQuAD using 60% less training data without loss of accuracy.
Multi-task dense retrieval models can be used to retrieve documents from a common corpus (e.g., Wikipedia) for different open-domain question-answering (QA) tasks. However, Karpukhin et al. (2020) shows that jointly learning different QA tasks with one dense model is not always beneficial due to corpus inconsistency. For example, SQuAD only focuses on a small set of Wikipedia articles while datasets like NQ and Trivia cover more entries, and joint training on their union can cause performance degradation. To solve this problem, we propose to train individual dense passage retrievers (DPR) for different tasks and aggregate their predictions during test time, where we use uncertainty estimation as weights to indicate how probable a specific query belongs to each expert’s expertise. Our method reaches state-of-the-art performance on 5 benchmark QA datasets, with up to 10% improvement in top-100 accuracy compared to a joint-training multi-task DPR on SQuAD. We also show that our method handles corpus inconsistency better than the joint-training DPR on a mixed subset of different QA datasets. Code and data are available at https://github.com/alexlimh/DPR_MUF.
Recent work has shown that dense passage retrieval techniques achieve better ranking accuracy in open-domain question answering compared to sparse retrieval techniques such as BM25, but at the cost of large space and memory requirements. In this paper, we analyze the redundancy present in encoded dense vectors and show that the default dimension of 768 is unnecessarily large. To improve space efficiency, we propose a simple unsupervised compression pipeline that consists of principal component analysis (PCA), product quantization, and hybrid search. We further investigate other supervised baselines and find surprisingly that unsupervised PCA outperforms them in some settings. We perform extensive experiments on five question answering datasets and demonstrate that our best pipeline achieves good accuracy–space trade-offs, for example, 48× compression with less than 3% drop in top-100 retrieval accuracy on average or 96× compression with less than 4% drop. Code and data are available at http://pyserini.io/.