Large language models (LLMs) have shown superb capability of modeling multimodal signals including audio and text, allowing the model to generate spoken or textual response given a speech input. However, it remains a challenge for the model to recognize personal named entities, such as contacts in a phone book, when the input modality is speech. In this work, we start with a speech recognition task and propose a retrievalbased solution to contextualize the LLM: we first let the LLM detect named entities in speech without any context, then use this named entity as a query to retrieve phonetically similar named entities from a personal database and feed them to the LLM, and finally run context-aware LLM decoding. In a voice assistant task, our solution achieved up to 30.2% relative word error rate reduction and 73.6% relative named entity error rate reduction compared to a baseline system without contextualization. Notably, our solution by design avoids prompting the LLM with the full named entity database, making it highly efficient and applicable to large named entity databases.

Virtual Assistants (VAs) are important Information Retrieval platforms that help users accomplish various tasks through spoken commands. The speech recognition system (speech-to-text) uses query priors, trained solely on text, to distinguish between phonetically confusing alternatives. Hence, the generation of synthetic queries that are similar to existing VA usage can greatly improve upon the VA's abilities---especially for use-cases that do not (yet) occur in paired audio/text data. In this paper, we provide a preliminary exploration of the use of Large Language Models (LLMs) to generate synthetic queries that are complementary to template-based methods. We investigate whether the methods (a) generate queries that are similar to randomly sampled, representative, and anonymized user queries from a popular VA, and (b) whether the generated queries are specific. We find that LLMs generate more verbose queries, compared to template-based methods, and reference aspects specific to the entity. However, the generated queries are similar to VA user queries, and are specific enough to retrieve the relevant entity. We conclude that queries generated by LLMs and templates are complementary.

Virtual assistants are becoming increasingly important speech-driven Information Retrieval platforms that assist users with various tasks. We discuss open problems and challenges with respect to modeling spoken information queries for virtual assistants, and list opportunities where Information Retrieval methods and research can be applied to improve the quality of virtual assistant speech recognition. We discuss how query domain classification, knowledge graphs and user interaction data, and query personalization can be helpful to improve the accurate recognition of spoken information domain queries. Finally, we also provide a brief overview of current problems and challenges in speech recognition.

Virtual assistants make use of automatic speech recognition (ASR) to help users answer entity-centric queries.However, spoken entity recognition is a difficult problem, due to the large number of frequently-changing named entities.In addition, resources available for recognition are constrained when ASR is performed on-device.In this work, we investigate the use of probabilistic grammars as language models within the finite-state transducer (FST) framework.We introduce a deterministic approximation to probabilistic grammars that avoids the explicit expansion of non-terminals at model creation time, integrates directly with the FST framework, and is complementary to n-gram models.We obtain a 10% relative word error rate improvement on long tail entity queries compared to when a similarly-sized n-gram model is used without our method.

Language models (LMs) for virtual assistants (VAs) are typically trained on large amounts of data, resulting in prohibitively large models which require excessive memory and/or cannot be used to serve user requests in real-time. Entropy pruning results in smaller models but with significant degradation of effectiveness in the tail of the user request distribution. We customize entropy pruning by allowing for a keep list of infrequent n-grams that require a more relaxed pruning threshold, and propose three methods to construct the keep list. Each method has its own advantages and disadvantages with respect to LM size, ASR accuracy and cost of constructing the keep list. Our best LM gives 8% average Word Error Rate (WER) reduction on a targeted test set, but is 3 times larger than the baseline. We also propose discriminative methods to reduce the size of the LM while retaining the majority of the WER gains achieved by the largest LM.

High-quality automatic speech recognition (ASR) is essential for virtual assistants (VAs) to work well. However, ASR often performs poorly on VA requests containing named entities. In this work, we start from the observation that many ASR errors on named entities are inconsistent with real-world knowledge. We extend previous discriminative n-gram language modeling approaches to incorporate real-world knowledge from a Knowledge Graph (KG), using features that capture entity type-entity and entity-entity relationships. We apply our model through an efficient lattice rescoring process, achieving relative sentence error rate reductions of more than 25% on some synthesized test sets covering less popular entities, with minimal degradation on a uniformly sampled VA test set.

We focus on improving the effectiveness of a Virtual Assistant (VA) in recognizing emerging entities in spoken queries. We introduce a method that uses historical user interactions to forecast which entities will gain in popularity and become trending, and it subsequently integrates the predictions within the Automated Speech Recognition (ASR) component of the VA. Experiments show that our proposed approach results in a 20% relative reduction in errors on emerging entity name utterances without degrading the overall recognition quality of the system.

In this work, we uncover a theoretical connection between two language model interpolation techniques, count merging and Bayesian interpolation. We compare these techniques as well as linear interpolation in three scenarios with abundant training data per component model. Consistent with prior work, we show that both count merging and Bayesian interpolation outperform linear interpolation. We include the first (to our knowledge) published comparison of count merging and Bayesian interpolation, showing that the two techniques perform similarly. Finally, we argue that other considerations will make Bayesian interpolation the preferred approach in most circumstances.

Two products are substitutes if both can satisfy the same consumer need. Intrinsic incorporation of product substitutability - where substitutability is integrated within latent vector space models - is in contrast to the extrinsic re-ranking of result lists. The fusion of text matching and product substitutability objectives allows latent vector space models to mix and match regularities contained within text descriptions and substitution relations. We introduce a method for intrinsically incorporating product substitutability within latent vector space models for product search that are estimated using gradient descent; it integrates flawlessly with state-of-the-art vector space models. We compare our method to existing methods for incorporating structural entity relations, where product substitutability is incorporated extrinsically by re-ranking. Our method outperforms the best extrinsic method on four benchmarks. We investigate the effect of different levels of text matching and product similarity objectives, and provide an analysis of the effect of incorporating product substitutability on product search ranking diversity. Incorporating product substitutability information improves search relevance at the cost of diversity.

Information Retrieval systems rely on large test collections to measure their effectiveness in retrieving relevant documents. While the demand is high, the task of creating such test collections is laborious due to the large amounts of data that need to be annotated, and due to the intrinsic subjectivity of the task itself. In this paper we study the topical relevance from a user perspective by addressing the problems of subjectivity and ambiguity. We compare our approach and results with the established TREC annotation guidelines and results. The comparison is based on a series of crowdsourcing pilots experimenting with variables, such as relevance scale, document granularity, annotation template and the number of workers. Our results show correlation between relevance assessment accuracy and smaller document granularity, i.e., aggregation of relevance on paragraph level results in a better relevance accuracy, compared to assessment done at the level of the full document. As expected, our results also show that collecting binary relevance judgments results in a higher accuracy compared to the ternary scale used in the TREC annotation guidelines. Finally, the crowdsourced annotation tasks provided a more accurate document relevance ranking than a single assessor relevance label. This work resulted is a reliable test collection around the TREC Common Core track.

We introduce pytrec_eval, a Python interface to the trec_eval information retrieval evaluation toolkit. pytrec_eval exposes the reference implementations of trec_eval within Python as a native extension. We show that pytrec_eval is around one order of magnitude faster than invoking trec_eval as a sub process from within Python. Compared to a native Python implementation of NDCG, pytrec_eval is twice as fast for practically-sized rankings. Finally, we demonstrate its effectiveness in an application where pytrec_eval is combined with Pyndri and the OpenAI Gym where query expansion is learned using Q-learning.

We propose the Neural Vector Space Model (NVSM), a method that learns representations of documents in an unsupervised manner for news article retrieval. In the NVSM paradigm, we learn low-dimensional representations of words and documents from scratch using gradient descent and rank documents according to their similarity with query representations that are composed from word representations. We show that NVSM performs better at document ranking than existing latent semantic vector space methods. The addition of NVSM to a mixture of lexical language models and a state-of-the-art baseline vector space model yields a statistically significant increase in retrieval effectiveness. Consequently, NVSM adds a complementary relevance signal. Next to semantic matching, we find that NVSM performs well in cases where lexical matching is needed. NVSM learns a notion of term specificity directly from the document collection without feature engineering. We also show that NVSM learns regularities related to Luhn significance. Finally, we give advice on how to deploy NVSM in situations where model selection (e.g., cross-validation) is infeasible. We find that an unsupervised ensemble of multiple models trained with different hyperparameter values performs better than a single cross-validated model. Therefore, NVSM can safely be used for ranking documents without supervised relevance judgments.

Email responses often contain items---such as a file or a hyperlink to an external document---that are attached to or included inline in the body of the message. Analysis of an enterprise email corpus reveals that 35% of the time when users include these items as part of their response, the attachable item is already present in their inbox or sent folder. A modern email client can proactively retrieve relevant attachable items from the user's past emails based on the context of the current conversation, and recommend them for inclusion, to reduce the time and effort involved in composing the response. In this paper, we propose a weakly supervised learning framework for recommending attachable items to the user. As email search systems are commonly available, we constrain the recommendation task to formulating effective search queries from the context of the conversations. The query is submitted to an existing IR system to retrieve relevant items for attachment. We also present a novel strategy for generating labels from an email corpus---without the need for manual annotations---that can be used to train and evaluate the query formulation model. In addition, we describe a deep convolutional neural network that demonstrates satisfactory performance on this query formulation task when evaluated on the publicly available Avocado dataset and a proprietary dataset of internal emails obtained through an employee participation program.

Machine learning plays a role in many aspects of modern IR systems, and deep learning is applied in all of them. The fast pace of modern-day research has given rise to many approaches to many IR problems. The amount of information available can be overwhelming both for junior students and for experienced researchers looking for new research topics and directions. The aim of this full- day tutorial is to give a clear overview of current tried-and-trusted neural methods in IR and how they benefit IR.

We introduce a novel latent vector space model that jointly learns the latent representations of words, e-commerce products and a mapping between the two without the need for explicit annotations. The power of the model lies in its ability to directly model the discriminative relation between products and a particular word. We compare our method to existing latent vector space models (LSI, LDA and word2vec) and evaluate it as a feature in a learning to rank setting. Our latent vector space model achieves its enhanced performance as it learns better product representations. Furthermore, the mapping from words to products and the representations of words benefit directly from the errors propagated back from the product representations during parameter estimation. We provide an in-depth analysis of the performance of our model and analyze the structure of the learned representations.

We introduce an unsupervised discriminative model for the task of retrieving experts in online document collections. We exclusively employ textual evidence and avoid explicit feature engineering by learning distributed word representations in an unsupervised way. We compare our model to state-of-the-art unsupervised statistical vector space and probabilistic generative approaches. Our proposed log-linear model achieves the retrieval performance levels of state-of-the-art document-centric methods with the low inference cost of so-called profile-centric approaches. It yields a statistically significant improved ranking over vector space and generative models in most cases, matching the performance of supervised methods on various benchmarks. That is, by using solely text we can do as well as methods that work with external evidence and/or relevance feedback. A contrastive analysis of rankings produced by discriminative and generative approaches shows that they have complementary strengths due to the ability of the unsupervised discriminative model to perform semantic matching.

User interactions with mobile devices increasingly depend on voice as a primary input modality. Due to the disadvantages of sending audio across potentially spotty network connections for speech recognition, in recent years there has been growing attention to performing recognition on-device. The limited computational resources, however, typically require additional model constraints. In this work, we explore the task of on-device utterance verification, wherein the recognizer must transcribe an utterance if it is in a target set or reject it as being out of domain. We present a data-driven methodology for mining tens of thousands of target phrases from an existing corpus. We then compare two common garbage-modeling approaches to utterance verification: a sub-word rejection model and a white-listed n-gram model. We examine a deficiency of the sub-word modeling approach and introduce a novel modification that makes use of common prefixes between targeted phrases and non-targeted phrases. We show good performance in the trade-off between recall and word error rate using both the prefix and white-listed n-gram approaches. Finally, we evaluate the prefix-based approach in a hybrid setting where rejected instances are sent to a server-side recognizer.