Search

As more and more online search queries come from voice, automatic speech recognition becomes a key component to deliver relevant search results. Errors introduced by automatic speech recognition (ASR) lead to irrelevant search results returned to the user, thus causing user dissatisfaction. In this paper, we introduce an approach, Mondegreen, to correct voice queries in text space without depending on audio signals, which may not always be available due to system constraints or privacy or bandwidth (for example, some ASR systems run on-device) considerations. We focus on voice queries transcribed via several proprietary commercial ASR systems. These queries come from users making internet, or online service search queries. We first present an analysis showing how different the language distribution coming from user voice queries is from that in traditional text corpora used to train off-the-shelf ASR systems. We then demonstrate that Mondegreen can achieve significant improvements in increased user interaction by correcting user voice queries in one of the largest search systems in Google. Finally, we see Mondegreen as complementing existing highly-optimized production ASR systems, which may not be frequently retrained and thus lag behind due to vocabulary drifts., Comment: Accepted in KDD 2021

Many recent advances in neural information retrieval models, which predict top-K items given a query, learn directly from a large training set of (query, item) pairs. However, they are often insufficient when there are many previously unseen (query, item) combinations, often referred to as the cold start problem. Furthermore, the search system can be biased towards items that are frequently shown to a query previously, also known as the 'rich get richer' (a.k.a. feedback loop) problem. In light of these problems, we observed that most online content platforms have both a search and a recommender system that, while having heterogeneous input spaces, can be connected through their common output item space and a shared semantic representation. In this paper, we propose a new Zero-Shot Heterogeneous Transfer Learning framework that transfers learned knowledge from the recommender system component to improve the search component of a content platform. First, it learns representations of items and their natural-language features by predicting (item, item) correlation graphs derived from the recommender system as an auxiliary task. Then, the learned representations are transferred to solve the target search retrieval task, performing query-to-item prediction without having seen any (query, item) pairs in training. We conduct online and offline experiments on one of the world's largest search and recommender systems from Google, and present the results and lessons learned. We demonstrate that the proposed approach can achieve high performance on offline search retrieval tasks, and more importantly, achieved significant improvements on relevance and user interactions over the highly-optimized production system in online experiments., Comment: Accepted at CIKM 2020

Reinforcement learning methods for recommender systems optimize recommendations for long-term user engagement. However, since users are often presented with slates of multiple items---which may have interacting effects on user choice---methods are required to deal with the combinatorics of the RL action space. We develop SlateQ, a decomposition of value-based temporal-difference and Q-learning that renders RL tractable with slates. Under mild assumptions on user choice behavior, we show that the long-term value (LTV) of a slate can be decomposed into a tractable function of its component item-wise LTVs. We demonstrate our methods in simulation, and validate the scalability and effectiveness of decomposed TD-learning on YouTube.

Generalized linear models with nonlinear feature transformations are widely used for large-scale regression and classification problems with sparse inputs. Memorization of feature interactions through a wide set of cross-product feature transformations are effective and interpretable, while generalization requires more feature engineering effort. With less feature engineering, deep neural networks can generalize better to unseen feature combinations through low-dimensional dense embeddings learned for the sparse features. However, deep neural networks with embeddings can over-generalize and recommend less relevant items when the user-item interactions are sparse and high-rank. In this paper, we present Wide & Deep learning---jointly trained wide linear models and deep neural networks---to combine the benefits of memorization and generalization for recommender systems. We productionized and evaluated the system on Google Play, a commercial mobile app store with over one billion active users and over one million apps. Online experiment results show that Wide & Deep significantly increased app acquisitions compared with wide-only and deep-only models. We have also open-sourced our implementation in TensorFlow.

The problem. We consider the problem of implementing a consistent replicated object in a partially synchronous message passing distributed system susceptible to process and communication failures. The object is a generic shared resource, such as a data structure, a file, or a lock. The processes implementing the replicated object access it by applying operations to it at unpredictable times and potentially concurrently.1 The object should be linearizable: it should behave as if each operation applied to it takes effect at a distinct instant in time during the interval between its invocation and its response.The main reason for replicating an object is fault tolerance; if a copy of the object becomes inaccessible, the object can still be accessed through the other copies. Another reason for replicating an object is performance; a process that requires the object can access its local copy. The benefit of accessing a local copy, however, must be weighed against the cost of keeping the copies consistent.

Bigtable is a distributed storage system for managing structured data that is designed to scale to a very large size: petabytes of data across thousands of commodity servers. Many projects at Google store data in Bigtable, including web indexing, Google Earth, and Google Finance. These applications place very different demands on Bigtable, both in terms of data size (from URLs to web pages to satellite imagery) and latency requirements (from backend bulk processing to real-time data serving). Despite these varied demands, Bigtable has successfully provided a flexible, high-performance solution for all of these Google products. In this article, we describe the simple data model provided by Bigtable, which gives clients dynamic control over data layout and format, and we describe the design and implementation of Bigtable.

An implementation of a shared object O is t-tolerant if the object remains correct and wait-free even when up to t base objects (objects used in the implementation of O) fail. The implementation is gracefully degrading if, no matter how many base objects fail, O does not fail more severely than its base objects. For the omission failure mode, we derive a lower bound on the space complexity of a gracefully degrading t-tolerant implementation. This result lets us conclude that, for omission failures, graceful degradation can be achieved only at the cost of increased space complexity.

Wait-free implementations of shared objects tolerate the failure of processes, but not the failure of base objects from which they are implemented. We consider the problem of implementing shared objects that tolerate the failure of both processes and base objects.We identify two classes of object failures:responsiveandnonresponsive. With responsive failures, a faulty object responds to every operation, but its responses may be incorrect. With nonresponsive failures, a faulty object may also “hang” without responding. In each class, we definecrash, omission,andarbitrarymodes of failure.We show that all responsive failure modes can be tolerated. More precisely, for all responsive failure modes ℱ, object typesT, andt≥ 0, we show how to implement a shared object of typeTwhich ist-tolerant for ℱ. Such an object remains correct and wait-free even if up totbase objects fail according to ℱ. In contrast to responsive failures, we show that even the most benign non-responsive failure mode cannot be tolerated. We also show that randomization can be used to circumvent this impossibility result.Graceful degradationis a desirable property of fault-tolerant implementations: the implemented object never fails more severely than the base objects it is derived from, even if all the base objects fail. For several failure modes, we show wheter this property can be achieved, and, if so, how.

We determine what information about failures is necessary and sufficient to solve Consensus in asynchronous distributed systems subject to crash failures. In Chandra and Toueg [1996], it is shown thatW, a failure detector that provides surprisingly little information about which processes have crashed, is sufficient to solve Consensus in asynchronous systems with a majority of correct processes. In this paper, we prove that to solve Consensus, any failure detector has to provide at least as much information as W. Thus, W is indeed the weakest failure detector for solving Consensus in asynchronous systems with a majority of correct processes.

We introduce the concept of unreliable failure detectors and study how they can be used to solve Consensus in asynchronous systems with crash failures. We characterise unreliable failure detectors in terms of two properties—completeness and accuracy. We show that Consensus can be solved even with unreliable failure detectors that make an infinite number of mistakes, and determine which ones can be used to solve Consensus despite any number of crashes, and which ones require a majority of correct processes. We prove that Consensus and Atomic Broadcast are reducible to each other in asynchronous systems with crash failures; thus, the above results also apply to Atomic Broadcast. A companion paper shows that one of the failure detectors introduced here is the weakest failure detector for solving Consensus [Chandra et al. 1992].

We describe our experience in building a fault-tolerant data-base using the Paxos consensus algorithm. Despite the existing literature in the field, building such a database proved to be non-trivial. We describe selected algorithmic and engineering problems encountered, and the solutions we found for them. Our measurements indicate that we have built a competitive system.

The publish/subscribe (or pub/sub) paradigm is an increasingly popular model for interconnecting applications in a distributed environment. Many existing pub/sub systems are based on pre-defined subjects, and hence are able to exploit multicast technologies to provide scalability and availability. An emerging alternative to subject-based systems, known as content-based systems, allow information consumers to request events based on the content of published events. This model is considerably more flexible than subject-based pub/sub. However, it was previously not known how to efficiently multicast published events to interested content-based subscribers within a large and geographically distributed network of broker (or router) machines. We develop and evaluate a novel and efficient distributed algorithm for this purpose, called -link matching". Link matching performs just enough computation at each node to determine the subset of links to which an event should be forwarded. We show via simulations that: link matching yields higher throughput than flooding when subscriptions are selective; and the overall CPU utilization of link matching is comparable to that of centralized matching.

Process groups in distributed applications and services rely on failure detectors to detect process failures completely, and as quickly, accurately, and scalably as possible, even in the face of unreliable message deliveries. In this paper, we look at quantifying the optimal scalability, in terms of network load, (in messages per second, with messages having a size limit) of distributed, complete failure detectors as a function of application-specified requirements. These requirements are 1) quick failure detection by some non-faulty process, and 2) accuracy of failure detection. We assume a crash-recovery (non-Byzantine) failure model, and a network model that is probabilistically unreliable (w.r.t. message deliveries and process failures). First, we characterize, under certain independence assumptions, the optimum worst-case network load imposed by any failure detector that achieves an application's requirements. We then discuss why traditional heart beating schemes are inherently unscalable according to the optimal load. We also present a randomized, distributed, failure detector algorithm that imposes an equal expected load per group member. This protocol satisfies the application defined constraints of completeness and accuracy, and speed of detection on an average. It imposes a network load that differs frown the optimal by a sub-optimality factor that is much lower than that for traditional distributed heartbeating schemes. Moreover, this sub-optimality factor does not vary with group size (for large groups).

Content-based subscription systems are an emerging alternative to traditional publish-subscribe systems, because they permit more flexible subscriptions along multiple dimensions. In these systems, each subscription is a predicate which may test arbitrary attributes within an event. However, the matching problem for content-based systems — determining for each event the subset of all subscriptions whose predicates match the event — is still an open problem. We present an efficient, scalable solution to the matching problem. Our solution has an expected time complexity that is sub-linear in the number of subscriptions, and it has a space complexity that is linear. Specifically, we prove that for predicates reducible to conjunctions of elementary tests, the expected time to match a random event is no greater than O(N 1 ) where N is the number of subscriptions, and is a closed-form expression that depends on the number and type of attributes (in some cases, 1=2). We present some optimizations to our algorithms that improve the search time. We also present the results of simulations that validate the theoretical bounds and that show acceptable performance levels for tens of thousands of subscriptions. Department of Computer Science, Cornell University, Ithaca, N.Y. 14853-7501, aguilera@cs.cornell.edu IBM T.J. Watson Research Center, Yorktown Heights, N.Y. 10598, fstrom, sturman, tusharg@watson.ibm.com Department of Computer Science, University of Illinois at Urbana-Champaign, 1304 W. Springfield Ave, Urbana, I.L. 61801, astley@cs.uiuc.edu

With the emergence of the internet, independent applications are starting to be integrated with each other. This has created a need for technology for glueing together applications both within and across organizations, without having to re-engineer individual components. We propose an approach for developing this glue technology based on message flows and discuss the open research problems in realizing this approach.

This paper describes the first Reliable Broadcast algorithms that are simultaneously efficient in both time and messages. These algorithms tolerate crash and omission failures. Each Reliable Broadcast takes O(f) time and O(fn) messages, where f is the number of processes that actually fail during this broadcast and n is the total number of processes. In other words, each additional process that fails during a broadcast can increase the broadcast time by at most a constant, and the number of messages by at most O(n). The algorithm tolerant of crash failures requires f+2 rounds. The one for general-omission failures requires 2f+3 rounds.