Knowledge-intensive language tasks require NLP systems to both provide the correct answer and retrieve supporting evidence for it in a given corpus. Autoregressive language models are emerging as the de-facto standard for generating answers, with newer and more powerful systems emerging at an astonishing pace. In this paper we argue that all this (and future) progress can be directly applied to the retrieval problem with minimal intervention to the models' architecture. Previous work has explored ways to partition the search space into hierarchical structures and retrieve documents by autoregressively generating their unique identifier. In this work we propose an alternative that doesn't force any structure in the search space: using all ngrams in a passage as its possible identifiers. This setup allows us to use an autoregressive model to generate and score distinctive ngrams, that are then mapped to full passages through an efficient data structure. Empirically, we show this not only outperforms prior autoregressive approaches but also leads to an average improvement of at least 10 points over more established retrieval solutions for passage-level retrieval on the KILT benchmark, establishing new state-of-the-art downstream performance on some datasets, while using a considerably lighter memory footprint than competing systems. Code available in the supplementary materials. Pre-trained models will be made available.

Search in social networks such as Facebook poses different challenges than in classical web search: besides the query text, it is important to take into account the searcher's context to provide relevant results. Their social graph is an integral part of this context and is a unique aspect of Facebook search. While embedding-based retrieval (EBR) has been applied in web search engines for years, Facebook search was still mainly based on a Boolean matching model. In this paper, we discuss the techniques for applying EBR to a Facebook Search system. We introduce the unified embedding framework developed to model semantic embeddings for personalized search, and the system to serve embedding-based retrieval in a typical search system based on an inverted index. We discuss various tricks and experiences on end-to-end optimization of the whole system, including ANN parameter tuning and full-stack optimization. Finally, we present our progress on two selected advanced topics about modeling. We evaluated EBR on verticals for Facebook Search with significant metrics gains observed in online A/B experiments. We believe this paper will provide useful insights and experiences to help people on developing embedding-based retrieval systems in search engines.

Query processing is one of the main bottlenecks in large-scale search engines. Retrieving the top k most relevant documents for a given query can be extremely expensive, as it involves scoring large amounts of documents. Several dynamic pruning techniques have been introduced in the literature to tackle this problem, such as BlockMaxWAND, which splits the inverted index into constant- sized blocks and stores the maximum document-term scores per block; this information can be used during query execution to safely skip low-score documents, producing many-fold speedups over exhaustive methods.

Graph reordering is a powerful technique to increase the locality of the representations of graphs, which can be helpful in several applications. We study how the technique can be used to improve compression of graphs and inverted indexes.

Entity linking deals with identifying entities from a knowledge base in a given piece of text and has become a fundamental building block for web search engines, enabling numerous downstream improvements from better document ranking to enhanced search results pages. A key problem in the context of web search queries is that this process needs to run under severe time constraints as it has to be performed before any actual retrieval takes place, typically within milliseconds.

Inverted indexes are usually represented by dividing posting lists into constant-sized blocks and representing them with an encoder for sequences of integers. Different encoders yield a different point in the space-time trade-off curve, with the fastest being several times larger than the most space-efficient. An important design decision for an index is thus the choice of the fastest encoding method such that the index fits in the available memory. However, a better usage of the space budget could be obtained by using faster encoders for frequently accessed blocks, and more space-efficient ones those that are rarely accessed. To perform this choice optimally, we introduce a linear time algorithm that, given a query distribution and a set of encoders, selects the best encoder for each index block to obtain the lowest expected query processing time respecting a given space constraint. To demonstrate the effectiveness of this approach we perform an extensive experimental analysis, which shows that our algorithm produces indexes which are significantly faster than single-encoder indexes under several query processing strategies, while respecting the same space constraints.

The computation of a peeling order in a randomly generated hypergraph is the most time-consuming step in a number of constructions, such as perfect hashing schemes, random r-SAT solvers, error-correcting codes, and approximate set encodings. While there exists a straightforward linear-time algorithm, its poor I/O performance makes it impractical for hypergraphs whose size exceeds the available internal memory. We show how to reduce the computation of a peeling order to a small number of sequential scans and sorts, and analyze its I/O complexity in the cache-oblivious model. The resulting algorithm requires O.sort.n// I/Os and O.n log n/ time to peel a random hypergraph with n edges. We experimentally evaluate the performance of our implementation of this algorithm in a real-world scenario by using the construction of minimal perfect hash functions (MPHF) as our test case: our algorithm builds a MPHF of 7:6 billion keys in less than 21 hours on a single machine. The resulting data structure is both more space-efficient and faster than that obtained with the current state-of-the-art MPHF construction for large-scale key sets.

The Elias-Fano representation of monotone sequences has been recently applied to the compression of inverted indexes, showing excellent query performance thanks to its efficient random access and search operations. While its space occupancy is competitive with some state-of-the-art methods such as gamma-delta-Golomb codes and PForDelta, it fails to exploit the local clustering that inverted lists usually exhibit, namely the presence of long subsequences of close identifiers. In this paper we describe a new representation based on partitioning the list into chunks and encoding both the chunks and their endpoints with Elias-Fano, hence forming a two-level data structure. This partitioning enables the encoding to better adapt to the local statistics of the chunk, thus exploiting clustering and improving compression. We present two partition strategies, respectively with fixed and variable-length chunks. For the latter case we introduce a linear-time optimization algorithm which identifies the minimum-space partition up to an arbitrarily small approximation factor.

Traditional video compression methods obtain a compact representation for image frames by computing coarse motion fields defined on patches of pixels called blocks, in order to compensate for the motion in the scene across frames. This piecewise constant approximation makes the motion field efficiently encodable, but it introduces block artifacts in the warped image frame. In this paper, we address the problem of estimating dense motion fields that, while accurately predicting one frame from a given reference frame by warping it with the field, are also compressible. We introduce a representation for motion fields based on wavelet bases, and approximate the compressibility of their coefficients with a piecewise smooth surrogate function that yields an objective function similar to classical optical flow formulations. We then show how to quantize and encode such coefficients with adaptive precision. We demonstrate the effectiveness of our approach by comparing its performance with a state-of-the-art wavelet video encoder. Experimental results on a number of standard flow and video datasets reveal that our method significantly outperforms both block-based and optical-flow-based motion compensation algorithms.

Virtually every modern search application, either desktop, web, or mobile, features some kind of query auto-completion. In its basic form, the problem consists in retrieving from a string set a small number of completions, i.e. strings beginning with a given prefix, that have the highest scores according to some static ranking. In this paper, we focus on the case where the string set is so large that compression is needed to fit the data structure in memory. This is a compelling case for web search engines and social networks, where it is necessary to index hundreds of millions of distinct queries to guarantee a reasonable coverage; and for mobile devices, where the amount of memory is limited.

Tries are popular data structures for storing a set of strings, where common prefixes are represented by common root-to-node paths. Over fifty years of usage have produced many variants and implementations to overcome some of their limitations. We explore new succinct representations of path-decomposed tries and experimentally evaluate the corresponding reduction in space usage and memory latency, comparing with the state of the art. We study two cases of applications: (1) a compressed dictionary for (compressed) strings, and (2) a monotone minimal perfect hash for strings that preserves their lexicographic order. For (1), we obtain data structures that outperform other state-of-the-art compressed dictionaries in space efficiency, while obtaining predictable query times that are competitive with data structures preferred by the practitioners. In (2), our tries perform several times faster than other trie-based monotone perfect hash functions, while occupying nearly the same space.

An indexed sequence of strings is a data structure for storing a string sequence that supports random access, searching, range counting and analytics operations, both for exact matches and prefix search. String sequences lie at the core of column-oriented databases, log processing, and other storage and query tasks. In these applications each string can appear several times and the order of the strings in the sequence is relevant. The prefix structure of the strings is relevant as well: common prefixes are sought in strings to extract interesting features from the sequence. Moreover, space-efficiency is highly desirable as it translates directly into higher performance, since more data can fit in fast memory.

Semi-structured textual formats are gaining increasing popularity for the storage of document collections and rich logs. Their flexibility comes at the cost of having to load and parse a document entirely even if just a small part of it needs to be accessed. For instance, in data analytics massive collections are usually scanned sequentially, selecting a small number of attributes from each document. We propose a technique to attach to a raw, unparsed document (even in compressed form) a "semi-index": a succinct data structure that supports operations on the document tree at speed comparable with an in-memory deserialized object, thus bridging textual formats with binary formats. After describing the general technique, we focus on the JSON format: our experiments show that avoiding the full loading and parsing step can give speedups of up to 12 times for on-disk documents using a small space overhead.