Retrieval-augmented generation models offer many benefits over standalone language models: besides a textual answer to a given query they provide provenance items retrieved from an updateable knowledge base. However, they are also more complex systems and need to handle long inputs. In this work, we introduce FiD-Light to strongly increase the efficiency of the state-of-the-art retrieval-augmented FiD model, while maintaining the same level of effectiveness. Our FiD-Light model constrains the information flow from the encoder (which encodes passages separately) to the decoder (using concatenated encoded representations). Furthermore, we adapt FiD-Light with re-ranking capabilities through textual source pointers, to improve the top-ranked provenance precision. Our experiments on a diverse set of seven knowledge intensive tasks (KILT) show FiD-Light consistently improves the Pareto frontier between query latency and effectiveness. FiD-Light with source pointing sets substantial new state-of-the-art results on six KILT tasks for combined text generation and provenance retrieval evaluation, while maintaining high efficiency.

Robust test collections are crucial for Information Retrieval research. Recently there is a growing interest in evaluating retrieval systems for domain-specific retrieval tasks, however these tasks often lack a reliable test collection with human-annotated relevance assessments following the Cranfield paradigm. In the medical domain, the TripClick collection was recently proposed, which contains click log data from the Trip search engine and includes two click-based test sets. However the clicks are biased to the retrieval model used, which remains unknown, and a previous study shows that the test sets have a low judgement coverage for the Top-10 results of lexical and neural retrieval models. In this paper we present the novel, relevance judgement test collection TripJudge for TripClick health retrieval. We collect relevance judgements in an annotation campaign and ensure the quality and reusability of TripJudge by a variety of ranking methods for pool creation, by multiple judgements per query-document pair and by an at least moderate inter-annotator agreement. We compare system evaluation with TripJudge and TripClick and find that that click and judgement-based evaluation can lead to substantially different system rankings.

Recent progress in neural information retrieval has demonstrated large gains in quality, while often sacrificing efficiency and interpretability compared to classical approaches. We propose ColBERTer, a neural retrieval model using contextualized late interaction (ColBERT) with enhanced reduction. Along the effectiveness Pareto frontier, ColBERTer dramatically lowers ColBERT's storage requirements while simultaneously improving the interpretability of its token-matching scores. To this end, ColBERTer fuses single-vector retrieval, multi-vector refinement, and optional lexical matching components into one model. For its multi-vector component, ColBERTer reduces the number of stored vectors by learning unique whole-word representations and learning to identify and remove word representations that are not essential to effective scoring. We employ an explicit multi-task, multi-stage training to facilitate using very small vector dimensions. Results on the MS MARCO and TREC-DL collection show that ColBERTer reduces the storage footprint by up to 2.5x, while maintaining effectiveness. With just one dimension per token in its smallest setting, ColBERTer achieves index storage parity with the plaintext size, with very strong effectiveness results. Finally, we demonstrate ColBERTer's robustness on seven high-quality out-of-domain collections, yielding statistically significant gains over traditional retrieval baselines.

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. 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. 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 [email protected], 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.

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.

The Transformer-Kernel (TK) model has demonstrated strong reranking performance on the TREC Deep Learning benchmark---and can be considered to be an efficient (but slightly less effective) alternative to other Transformer-based architectures that employ (i) large-scale pretraining (high training cost), (ii) joint encoding of query and document (high inference cost), and (iii) larger number of Transformer layers (both high training and high inference costs). Since, a variant of the TK model---called TKL---has been developed that incorporates local self-attention to efficiently process longer input sequences in the context of document ranking. In this work, we propose a novel Conformer layer as an alternative approach to scale TK to longer input sequences. Furthermore, we incorporate query term independence and explicit term matching to extend the model to the full retrieval setting. We benchmark our models under the strictly blind evaluation setting of the TREC 2020 Deep Learning track and find that our proposed architecture changes lead to improved retrieval quality over TKL. Our best model also outperforms all non-neural runs ("trad") and two-thirds of the pretrained Transformer-based runs ("nnlm") on [email protected]

The use of stopwords has been thoroughly studied in traditional Information Retrieval systems, but remains unexplored in the context of neural models. Neural re-ranking models take the full text of both the query and document into account. Naturally, removing tokens that do not carry relevance information provides us with an opportunity to improve the effectiveness by reducing noise and lower document representation caching-storage requirements. In this work we propose a novel contextualized stopword detection mechanism for neural re-ranking models. This mechanism consists of training a sparse vector in order to filter out document tokens from the ranking decision. This vector is learned end-to-end based on the contextualized document representations, allowing the model to filter terms on a per occurrence basis. This leads to a more explainable model, as it reduces noise. We integrate our component into the state-of-the-art interaction-based TK neural re-ranking model. Our experiments on the MS MARCO passage collection and queries from the TREC 2019 Deep Learning Track show that filtering out traditional stopwords prior to the neural model reduces its effectiveness, while learning to filter out contextualized representations improves it.

There are many existing retrieval and question answering datasets. However, most of them either focus on ranked list evaluation or single-candidate question answering. This divide makes it challenging to properly evaluate approaches concerned with ranking documents and providing snippets or answers for a given query. In this work, we present FiRA: a novel dataset of Fine-Grained Relevance Annotations. We extend the ranked retrieval annotations of the Deep Learning track of TREC 2019 with passage and word level graded relevance annotations for all relevant documents. We use our newly created data to study the distribution of relevance in long documents, as well as the attention of annotators to specific positions of the text. As an example, we evaluate the recently introduced TKL document ranking model. We find that although TKL exhibits state-of-the-art retrieval results for long documents, it misses many relevant passages.

The effective extraction of ranked disease-symptom relationships is a critical component in various medical tasks, including computer-assisted medical diagnosis or the discovery of unexpected associations between diseases. While existing disease-symptom relationship extraction methods are used as the foundation in the various medical tasks, no collection is available to systematically evaluate the performance of such methods. In this paper, we introduce the D isease- S ymptom R elation Collection ( dsr -collection), created by five physicians as expert annotators. We provide graded symptom judgments for diseases by differentiating between relevant symptoms and primary symptoms . Further, we provide several strong baselines, based on the methods used in previous studies. The first method is based on word embeddings, and the second on co-occurrences of MeSH-keywords of medical articles. For the co-occurrence method, we propose an adaption in which not only keywords are considered, but also the full text of medical articles. The evaluation on the dsr -collection shows the effectiveness of the proposed adaption in terms of nDCG, precision, and recall.

Relevance in ad-hoc retrieval is a fundamental problem of text understanding. Developing neural network methods for this foundational task of Information Retrieval (IR) has the potential to impact many search domains. Recently, a new generation of Transformerbased [6] neural re-ranking models initiated a new era, by providing substantial effectiveness increases in ad-hoc search tasks [1, 4, 5]. They operate on the full text of a query and a list of candidate documents from an initial retrieval. Using self-attention, they contextualize term occurrences conditioned on the sequence they are contained in. However, self-attention, because it is applied to a whole sequence and commonly uses many layers, is inefficient and re-ranking models still depended on the bottleneck of the initial retrieval of candidate documents. In our thesis we plan to address these shortcomings. First, we plan to make contextualization efficient enough to to be usable in resource constrained environments, and second we plan to use the efficient contextualization components to create a novel approach for learning to index and retrieve in a unified neural model. Hence, our main research questions are: RQ1 How can we balance the trade-off between efficiency and effectiveness in contextualized neural re-ranking? Solving efficiency problems with massively parallelized hardware is neither an economically nor environmentally friendly solution. We wish to tackle the problem of efficient contextualization in two common scenarios of ad-hoc ranking: passage retrieval and document retrieval. As part of this PhD we already proposed a novel efficient Transformer based passage re-ranking model: The TK (Transformer-Kernel) model [4] utilizes shallow Transformerlayers to contextualize query and passages separately, and the kernel-pooling technique [7] to score individual term interactions between a query and a passage. Additionally, we explored localattention as an effective approach for document-length ranking with an extension to the TK for long text (TKL) [3]. RQ2 What is needed to learn generalized contextualized representations for indexing and retrieval? Currently, IR systems are a patchwork of traditional and neural systems. Different parts of the pipeline are optimized in isolation and missing integration during training. To overcome bottlenecks and sub-optimal integration of pipeline components, we plan to develop a unified neural index and ranking model. A unified indexing and ranking model needs to learn how to efficiently store document representations, for example via trained sparsity [8], followed by the integration of all second-generation contextualized re-ranking components. We plan to train this unified model in a true end-toend approach, with gradients flowing through the complete system. The main challenge here is to generalize models beyond their training domain, to match the usability of traditional relevance models such as BM25, which can be used as a drop-in approach in almost any domain and language. As part of this PhD we already showed the importance of tuning the re-ranking depth as the interface between traditional and neural models [2]. Furthermore, we plan to create a truly novel indexing and ranking model, that removes all bottlenecks and patchwork systems. Contextualization is a key element to focus the model on the actual topic of a query and document, and reduce the search space for previously ambiguous words and topics. Parallel to our main research questions, we plan to keep explainability and thorough analysis of our approaches a priority during this PhD work.

Neural networks, particularly Transformer-based architectures, have achieved significant performance improvements on several retrieval benchmarks. When the items being retrieved are documents, the time and memory cost of employing Transformers over a full sequence of document terms can be prohibitive. A popular strategy involves considering only the first n terms of the document. This can, however, result in a biased system that under retrieves longer documents. In this work, we propose a local self-attention which considers a moving window over the document terms and for each term attends only to other terms in the same window. This local attention incurs a fraction of the compute and memory cost of attention over the whole document. The windowed approach also leads to more compact packing of padded documents in minibatches resulting in additional savings. We also employ a learned saturation function and a two-staged pooling strategy to identify relevant regions of the document. The Transformer-Kernel pooling model with these changes can efficiently elicit relevance information from documents with thousands of tokens. We benchmark our proposed modifications on the document ranking task from the TREC 2019 Deep Learning track and observe significant improvements in retrieval quality as well as increased retrieval of longer documents at moderate increase in compute and memory costs.

The success of crowdsourcing based annotation of text corpora depends on ensuring that crowdworkers are sufficiently well-trained to perform the annotation task accurately. To that end, a frequent approach to train annotators is to provide instructions and a few example cases that demonstrate how the task should be performed (referred to as the CONTROL approach). These globally defined "task-level examples", however, (i) often only cover the common cases that are encountered during an annotation task; and (ii) require effort from crowdworkers during the annotation process to find the most relevant example for the currently annotated sample. To overcome these limitations, we propose to support workers in addition to task-level examples, also with "task-instance level" examples that are semantically similar to the currently annotated data sample (referred to as Dynamic Examples for Annotation, DEXA). Such dynamic examples can be retrieved from collections previously labeled by experts, which are usually available as gold standard dataset. We evaluate DEXA on a complex task of annotating participants, interventions, and outcomes (known as PIO) in sentences of medical studies. The dynamic examples are retrieved using BioSent2Vec, an unsupervised semantic sentence similarity method specific to the biomedical domain. Results show that (i) workers of the DEXA approach reach on average much higher agreements (Cohen's Kappa) to experts than workers of the the CONTROL approach (avg. of 0.68 to experts in DEXA vs. 0.40 in CONTROL); (ii) already three per majority voting aggregated annotations of the DEXA approach reach substantial agreements to experts of 0.78/0.75/0.69 for P/I/O (in CONTROL 0.73/0.58/0.46). Finally, (iii) we acquire explicit feedback from workers and show that in the majority of cases (avg. 72%) workers find the dynamic examples useful.

Low-frequency terms are a recurring challenge for information retrieval models, especially neural IR frameworks struggle with adequately capturing infrequently observed words. While these terms are often removed from neural models - mainly as a concession to efficiency demands - they traditionally play an important role in the performance of IR models. In this paper, we analyze the effects of low-frequency terms on the performance and robustness of neural IR models. We conduct controlled experiments on three recent neural IR models, trained on a large-scale passage retrieval collection. We evaluate the neural IR models with various vocabulary sizes for their respective word embeddings, considering different levels of constraints on the available GPU memory. We observe that despite the significant benefits of using larger vocabularies, the performance gap between the vocabularies can be, to a great extent, mitigated by extensive tuning of a related parameter: the number of documents to re-rank. We further investigate the use of subword-token embedding models, and in particular FastText, for neural IR models. Our experiments show that using FastText brings slight improvements to the overall performance of the neural IR models in comparison to models trained on the full vocabulary, while the improvement becomes much more pronounced for queries containing low-frequency terms.