Recent advances in retrieval models based on learned sparse representations generated by transformers have led us to, once again, consider score-at-a-time query evaluation techniques for the top-k retrieval problem. Previous studies comparing document-at-a-time and score-at-a-time approaches have consistently found that the former approach yields lower mean query latency, although the latter approach has more predictable query latency. In our experiments with four different retrieval models that exploit representational learning with bags of words, we find that transformers generate "wacky weights" that appear to greatly reduce the opportunities for skipping and early exiting optimizations that lie at the core of standard document-at-a-time techniques. As a result, score-at-a-time approaches appear to be more competitive in terms of query evaluation latency than in previous studies. We find that, if an effectiveness loss of up to three percent can be tolerated, a score-at-a-time approach can yield substantial gains in mean query latency while at the same time dramatically reducing tail latency.
One key feature of dense passage retrievers (DPR) is the use of separate question and passage encoder in a bi-encoder design. Previous work on generalization of DPR mainly focus on testing both encoders in tandem on out-of-distribution (OOD) question-answering (QA) tasks, which is also known as domain adaptation. However, it is still unknown how DPR's individual question/passage encoder affects generalization. Specifically, in this paper, we want to know how an in-distribution (IND) question/passage encoder would generalize if paired with an OOD passage/question encoder from another domain. We refer to this challenge as \textit{encoder adaptation}. To answer this question, we inspect different combinations of DPR's question and passage encoder learned from five benchmark QA datasets on both in-domain and out-of-domain questions. We find that the passage encoder has more influence on the lower bound of generalization while the question encoder seems to affect the upper bound in general. For example, applying an OOD passage encoder usually hurts the retrieval accuracy while an OOD question encoder sometimes even improves the accuracy.
This paper outlines a conceptual framework for understanding recent developments in information retrieval and natural language processing that attempts to integrate dense and sparse retrieval methods. I propose a representational approach that breaks the core text retrieval problem into a logical scoring model and a physical retrieval model. The scoring model is defined in terms of encoders, which map queries and documents into a representational space, and a comparison function that computes query-document scores. The physical retrieval model defines how a system produces the top-k scoring documents from an arbitrarily large corpus with respect to a query. The scoring model can be further analyzed along two dimensions: dense vs. sparse representations and supervised (learned) vs. unsupervised approaches. I show that many recently proposed retrieval methods, including multi-stage ranking designs, can be seen as different parameterizations in this framework, and that a unified view suggests a number of open research questions, providing a roadmap for future work. As a bonus, this conceptual framework establishes connections to sentence similarity tasks in natural language processing and information access "technologies" prior to the dawn of computing.
Dense retrieval has shown great success in passage ranking in English. However, its effectiveness in document retrieval for non-English languages remains unexplored due to the limitation in training resources. In this work, we explore different transfer techniques for document ranking from English annotations to multiple non-English languages. Our experiments on the test collections in six languages (Chinese, Arabic, French, Hindi, Bengali, Spanish) from diverse language families reveal that zero-shot model-based transfer using mBERT improves the search quality in non-English mono-lingual retrieval. Also, we find that weakly-supervised target language transfer yields competitive performances against the generation-based target language transfer that requires external translators and query generators.
We present Mr. TyDi, a multi-lingual benchmark dataset for mono-lingual retrieval in eleven typologically diverse languages, designed to evaluate ranking with learned dense representations. The goal of this resource is to spur research in dense retrieval techniques in non-English languages, motivated by recent observations that existing techniques for representation learning perform poorly when applied to out-of-distribution data. As a starting point, we provide zero-shot baselines for this new dataset based on a multi-lingual adaptation of DPR that we call "mDPR". Experiments show that although the effectiveness of mDPR is much lower than BM25, dense representations nevertheless appear to provide valuable relevance signals, improving BM25 results in sparse-dense hybrids. In addition to analyses of our results, we also discuss future challenges and present a research agenda in multi-lingual dense retrieval. Mr. TyDi can be downloaded at https://github.com/castorini/mr.tydi.
Recent developments in representational learning for information retrieval can be organized in a conceptual framework that establishes two pairs of contrasts: sparse vs. dense representations and unsupervised vs. learned representations. Sparse learned representations can further be decomposed into expansion and term weighting components. This framework allows us to understand the relationship between recently proposed techniques such as DPR, ANCE, DeepCT, DeepImpact, and COIL, and furthermore, gaps revealed by our analysis point to "low hanging fruit" in terms of techniques that have yet to be explored. We present a novel technique dubbed "uniCOIL", a simple extension of COIL that achieves to our knowledge the current state-of-the-art in sparse retrieval on the popular MS MARCO passage ranking dataset. Our implementation using the Anserini IR toolkit is built on the Lucene search library and thus fully compatible with standard inverted indexes.
Evaluation efforts such as TREC, CLEF, NTCIR and FIRE, alongside public leaderboard such as MS MARCO, are intended to encourage research and track our progress, addressing big questions in our field. However, the goal is not simply to identify which run is "best", achieving the top score. The goal is to move the field forward by developing new robust techniques, that work in many different settings, and are adopted in research and practice. This paper uses the MS MARCO and TREC Deep Learning Track as our case study, comparing it to the case of TREC ad hoc ranking in the 1990s. We show how the design of the evaluation effort can encourage or discourage certain outcomes, and raising questions about internal and external validity of results. We provide some analysis of certain pitfalls, and a statement of best practices for avoiding such pitfalls. We summarize the progress of the effort so far, and describe our desired end state of "robust usefulness", along with steps that might be required to get us there.
Conversational search (CS) plays a vital role in information retrieval. The current state of the art approaches the task using a multi-stage pipeline comprising conversational query reformulation and information seeking modules. Despite of its effectiveness, such a pipeline often comprises multiple neural models and thus requires long inference times. In addition, independently optimizing the effectiveness of each module does not consider the relation between modules in the pipeline. Thus, in this paper, we propose a single-stage design, which supports end-to-end training and low-latency inference. To aid in this goal, we create a synthetic dataset for CS to overcome the lack of training data and explore different training strategies using this dataset. Experiments demonstrate that our model yields competitive retrieval effectiveness against state-of-the-art multi-stage approaches but with lower latency. Furthermore, we show that improved retrieval effectiveness benefits the downstream task of conversational question answering.
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.
Text retrieval using learned dense representations has recently emerged as a promising alternative to "traditional" text retrieval using sparse bag-of-words representations. One recent work that has garnered much attention is the dense passage retriever (DPR) technique proposed by Karpukhin et al. (2020) for end-to-end open-domain question answering. We present a replication study of this work, starting with model checkpoints provided by the authors, but otherwise from an independent implementation in our group's Pyserini IR toolkit and PyGaggle neural text ranking library. Although our experimental results largely verify the claims of the original paper, we arrived at two important additional findings that contribute to a better understanding of DPR: First, it appears that the original authors under-report the effectiveness of the BM25 baseline and hence also dense--sparse hybrid retrieval results. Second, by incorporating evidence from the retriever and an improved answer span scoring technique, we are able to improve end-to-end question answering effectiveness using exactly the same models as in the original work.