Domain Adaptation for Memory-Efficient Dense Retrieval

Dense retrievers encode documents into fixed dimensional embeddings. However, storing all the document embeddings within an index produces bulky indexes which are expensive to serve. Recently, BPR (Yamada et al., 2021) and JPQ (Zhan et al., 2021a) have been proposed which train the model to produce binary document vectors, which reduce the index 32x and more. The authors showed these binary embedding models significantly outperform more traditional index compression techniques like Product Quantization (PQ). Previous work evaluated these approaches just in-domain, i.e. the methods were evaluated on tasks for which training data is available. In practice, retrieval models are often used in an out-of-domain setting, where they have been trained on a publicly available dataset, like MS MARCO, but are then used for some custom dataset for which no training data is available. In this work, we show that binary embedding models like BPR and JPQ can perform significantly worse than baselines once there is a domain-shift involved. We propose a modification to the training procedure of BPR and JPQ and combine it with a corpus specific generative procedure which allow the adaptation of BPR and JPQ to any corpus without requiring labeled training data. Our domain-adapted strategy known as GPL is model agnostic, achieves an improvement by up-to 19.3 and 11.6 points in nDCG@10 across the BEIR benchmark in comparison to BPR and JPQ while maintaining its 32x memory efficiency. JPQ+GPL even outperforms our upper baseline: uncompressed TAS-B model on average by 2.0 points.

Certified Error Control of Candidate Set Pruning for Two-Stage Relevance Ranking

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 off against computational efficiency in an empirical fashion, lacking 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 yields an average candidate set size of 27 out of 1,000 which increases the reranking speed by about 37 times, while the MRR@10 is greater than a pre-specified value of 0.38 with about 90% empirical coverage and the empirical baselines fail to provide such guarantee. Code and data are available at: https://github.com/alexlimh/CEC-Ranking.

To Interpolate or not to Interpolate: PRF, Dense and Sparse Retrievers

Current pre-trained language model approaches to information retrieval can be broadly divided into two categories: sparse retrievers (to which belong also non-neural approaches such as bag-of-words methods, e.g., BM25) and dense retrievers. Each of these categories appears to capture different characteristics of relevance. Previous work has investigated how relevance signals from sparse retrievers could be combined with those from dense retrievers via interpolation. Such interpolation would generally lead to higher retrieval effectiveness. In this paper we consider the problem of combining the relevance signals from sparse and dense retrievers in the context of Pseudo Relevance Feedback (PRF). This context poses two key challenges: (1) When should interpolation occur: before, after, or both before and after the PRF process? (2) Which sparse representation should be considered: a zero-shot bag-of-words model (BM25), or a learnt sparse representation? To answer these questions we perform a thorough empirical evaluation considering an effective and scalable neural PRF approach (Vector-PRF), three effective dense retrievers (ANCE, TCTv2, DistillBERT), and one state-of-the-art learnt sparse retriever (uniCOIL). The empirical findings from our experiments suggest that, regardless of sparse representation and dense retriever, interpolation both before and after PRF achieves the highest effectiveness across most datasets and metrics.

Towards Best Practices for Training Multilingual Dense Retrieval Models

Dense retrieval models using a transformer-based bi-encoder design have emerged as an active area of research. In this work, we focus on the task of monolingual retrieval in a variety of typologically diverse languages using one such design. Although recent work with multilingual transformers demonstrates that they exhibit strong cross-lingual generalization capabilities, there remain many open research questions, which we tackle here. Our study is organized as a "best practices" guide for training multilingual dense retrieval models, broken down into three main scenarios: where a multilingual transformer is available, but relevance judgments are not available in the language of interest; where both models and training data are available; and, where training data are available not but models. In considering these scenarios, we gain a better understanding of the role of multi-stage fine-tuning, the strength of cross-lingual transfer under various conditions, the usefulness of out-of-language data, and the advantages of multilingual vs. monolingual transformers. Our recommendations offer a guide for practitioners building search applications, particularly for low-resource languages, and while our work leaves open a number of research questions, we provide a solid foundation for future work.

Evaluating Token-Level and Passage-Level Dense Retrieval Models for Math Information Retrieval

With the recent success of dense retrieval methods based on bi-encoders, a number of studies have applied this approach to various interesting downstream retrieval tasks with good efficiency and in-domain effectiveness. Recently, we have also seen the presence of dense retrieval models in Math Information Retrieval (MIR) tasks, but the most effective systems remain "classic" retrieval methods that consider rich structure features. In this work, we try to combine the best of both worlds: a well-defined structure search method for effective formula search and bi-encoder dense retrieval models to capture contextual similarities in mathematical documents. Specifically, we have evaluated two representative bi-encoder models (ColBERT and DPR) for token-level and passage-level dense retrieval on recent MIR tasks. To our best knowledge, this is the first time a DPR model has been evaluated in the MIR domain. Our result shows that bi-encoder models are complementary to existing structure search methods, and we are able to advance the state of the art on a recent MIR dataset. We have made our model checkpoints and source code publicly available for the reproduction of our results.

Tevatron: An Efficient and Flexible Toolkit for Dense Retrieval

Recent rapid advancements in deep pre-trained language models and the introductions of large datasets have powered research in embedding-based dense retrieval. While several good research papers have emerged, many of them come with their own software stacks. These stacks are typically optimized for some particular research goals instead of efficiency or code structure. In this paper, we present Tevatron, a dense retrieval toolkit optimized for efficiency, flexibility, and code simplicity. Tevatron provides a standardized pipeline for dense retrieval including text processing, model training, corpus/query encoding, and search. This paper presents an overview of Tevatron and demonstrates its effectiveness and efficiency across several IR and QA data sets. We also show how Tevatron's flexible design enables easy generalization across datasets, model architectures, and accelerator platforms(GPU/TPU). We believe Tevatron can serve as an effective software foundation for dense retrieval system research including design, modeling, and optimization.

Can Old TREC Collections Reliably Evaluate Modern Neural Retrieval Models?

Neural retrieval models are generally regarded as fundamentally different from the retrieval techniques used in the late 1990's when the TREC ad hoc test collections were constructed. They thus provide the opportunity to empirically test the claim that pooling-built test collections can reliably evaluate retrieval systems that did not contribute to the construction of the collection (in other words, that such collections can be reusable). To test the reusability claim, we asked TREC assessors to judge new pools created from new search results for the TREC-8 ad hoc collection. These new search results consisted of five new runs (one each from three transformer-based models and two baseline runs that use BM25) plus the set of TREC-8 submissions that did not previously contribute to pools. The new runs did retrieve previously unseen documents, but the vast majority of those documents were not relevant. The ranking of all runs by mean evaluation score when evaluated using the official TREC-8 relevance judgment set and the newly expanded relevance set are almost identical, with Kendall's tau correlations greater than 0.99. Correlations for individual topics are also high. The TREC-8 ad hoc collection was originally constructed using deep pools over a diverse set of runs, including several effective manual runs. Its judgment budget, and hence construction cost, was relatively large. However, it does appear that the expense was well-spent: even with the advent of neural techniques, the collection has stood the test of time and remains a reliable evaluation instrument as retrieval techniques have advanced.

Sparsifying Sparse Representations for Passage Retrieval by Top-$k$ Masking

Sparse lexical representation learning has demonstrated much progress in improving passage retrieval effectiveness in recent models such as DeepImpact, uniCOIL, and SPLADE. This paper describes a straightforward yet effective approach for sparsifying lexical representations for passage retrieval, building on SPLADE by introducing a top-$k$ masking scheme to control sparsity and a self-learning method to coax masked representations to mimic unmasked representations. A basic implementation of our model is competitive with more sophisticated approaches and achieves a good balance between effectiveness and efficiency. The simplicity of our methods opens the door for future explorations in lexical representation learning for passage retrieval.

Improving Query Representations for Dense Retrieval with Pseudo Relevance Feedback: A Reproducibility Study

Pseudo-Relevance Feedback (PRF) utilises the relevance signals from the top-k passages from the first round of retrieval to perform a second round of retrieval aiming to improve search effectiveness. A recent research direction has been the study and development of PRF methods for deep language models based rankers, and in particular in the context of dense retrievers. Dense retrievers, compared to more complex neural rankers, provide a trade-off between effectiveness, which is often reduced compared to more complex neural rankers, and query latency, which also is reduced making the retrieval pipeline more efficient. The introduction of PRF methods for dense retrievers has been motivated as an attempt to further improve their effectiveness. In this paper, we reproduce and study a recent method for PRF with dense retrievers, called ANCE-PRF. This method concatenates the query text and that of the top-k feedback passages to form a new query input, which is then encoded into a dense representation using a newly trained query encoder based on the original dense retriever used for the first round of retrieval. While the method can potentially be applied to any of the existing dense retrievers, prior work has studied it only in the context of the ANCE dense retriever. We study the reproducibility of ANCE-PRF in terms of both its training (encoding of the PRF signal) and inference (ranking) steps. We further extend the empirical analysis provided in the original work to investigate the effect of the hyper-parameters that govern the training process and the robustness of the method across these different settings. Finally, we contribute a study of the generalisability of the ANCE-PRF method when dense retrievers other than ANCE are used for the first round of retrieval and for encoding the PRF signal.

Densifying Sparse Representations for Passage Retrieval by Representational Slicing

Learned sparse and dense representations capture different successful approaches to text retrieval and the fusion of their results has proven to be more effective and robust. Prior work combines dense and sparse retrievers by fusing their model scores. As an alternative, this paper presents a simple approach to densifying sparse representations for text retrieval that does not involve any training. Our densified sparse representations (DSRs) are interpretable and can be easily combined with dense representations for end-to-end retrieval. We demonstrate that our approach can jointly learn sparse and dense representations within a single model and then combine them for dense retrieval. Experimental results suggest that combining our DSRs and dense representations yields a balanced tradeoff between effectiveness and efficiency.