On online content platforms, users often aim to explore the catalog and discover new, personalized content through exploratory searches-such as "new releases for me." Traditional search systems, which prioritize lexical and semantic matching over personalized retrieval, have historically struggled to support this type of intent.In contrast, recommendation services that leverage user-item and item-item signals tend to be more effective for addressing exploratory queries.Agentic technologies offer a promising opportunity to enhance exploratory search by harnessing large language models (LLMs) to interpret complex query intents and route them to the most suitable downstream services.However, deploying such

Listeners of long-form talk-audio content, such as podcast episodes, often find it challenging to understand the overall structure and locate relevant sections. A practical solution is to divide episodes into chapters--semantically coherent segments labeled with titles and timestamps. Since most episodes on our platform at Spotify currently lack creator-provided chapters, automating the creation of chapters is essential. Scaling the chapterization of podcast episodes presents unique challenges. First, episodes tend to be less structured than written texts, featuring spontaneous discussions with nuanced transitions. Second, the transcripts are usually lengthy, averaging about 16,000 tokens, which necessitates efficient processing that can preserve context. To address these challenges, we introduce PODTILE, a fine-tuned encoder-decoder transformer to segment conversational data. The model simultaneously generates chapter transitions and titles for the input transcript. To preserve context, each input text is augmented with global context, including the episode's title, description, and previous chapter titles. In our intrinsic evaluation, PODTILE achieved an 11% improvement in ROUGE score over the strongest baseline. Additionally, we provide insights into the practical benefits of auto-generated chapters for listeners navigating episode content. Our findings indicate that auto-generated chapters serve as a useful tool for engaging with less popular podcasts. Finally, we present empirical evidence that using chapter titles can enhance effectiveness of sparse retrieval in search tasks.

In this paper we improve the zero-shot generalization ability of language models via Mixture-Of-Memory Augmentation (MoMA), a mechanism that retrieves augmentation documents from multiple information corpora ("external memories"), with the option to "plug in" unseen memory at inference time.We develop a joint learning mechanism that trains the augmentation component with latent labels derived from the end retrieval task, paired with hard negatives from the memory mixture.We instantiate the model in a zero-shot dense retrieval setting by augmenting strong T5-based retrievers with MoMA.With only T5-base, our model obtains strong zero-shot retrieval accuracy on the eighteen tasks included in the standard BEIR benchmark, outperforming some systems with larger model sizes.As a plug-inplay model, our model can efficiently generalize to any unseen corpus, meanwhile achieving comparable or even better performance than methods relying on target-specific pretraining.Our analysis further illustrates the necessity of augmenting with mixture-of-memory for robust generalization, the benefits of augmentation learning, and how MoMA utilizes the plugin memory at inference time without changing its parameters.

The remarkable abilities of large language models (LLMs) like GPT-4 partially stem from posttraining processes like Reinforcement Learning from Human Feedback (RLHF) involving human preferences encoded in a reward model.However, these reward models (RMs) often lack direct knowledge of why, or under what principles, the preferences annotations were made.In this study, we identify principles that guide RMs to better align with human preferences, and then develop an axiomatic framework to generate a rich variety of preference signals to uphold them.We use these axiomatic signals to train a model for scoring answers to longform questions.Our approach yields a Preference Model with only about 220M parameters that agrees with gold humanannotated preference labels more often than GPT-4.The contributions of this work include: training a standalone preference model that can score human-and LLM-generated answers on the same scale; developing an axiomatic framework for generating training data pairs tailored to certain principles; and showing that a small amount of axiomatic signals can help small models outperform GPT-4 in preference scoring.We intend to release our model.

Abstract In the summarization domain, a key requirement for summaries is to be factually consistent with the input document. Previous work has found that natural language inference (NLI) models do not perform competitively when applied to inconsistency detection. In this work, we revisit the use of NLI for inconsistency detection, finding that past work suffered from a mismatch in input granularity between NLI datasets (sentence-level), and inconsistency detection (document level). We provide a highly effective and light-weight method called SummaCConv that enables NLI models to be successfully used for this task by segmenting documents into sentence units and aggregating scores between pairs of sentences. We furthermore introduce a new benchmark called SummaC (Summary Consistency) which consists of six large inconsistency detection datasets. On this dataset, SummaCConv obtains state-of-the-art results with a balanced accuracy of 74.4%, a 5% improvement compared with prior work.

Philippe Laban, Tobias Schnabel, Paul Bennett, Marti A. Hearst. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Si Sun, Yingzhuo Qian, Zhenghao Liu, Chenyan Xiong, Kaitao Zhang, Jie Bao, Zhiyuan Liu, Paul Bennett. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

People often capture photos or notes from their phones to integrate later into a document. But current mobile capture tools can make this hard, with the captured information ending up fragmented and decontextualized. This paper explores how to help document authors capture, contextualize, and use document-related information. A survey of 66 information workers reveals that document-focused information capture differs from other types of mobile information capture, and that while people capture a broad range of information types while mobile, most document-related capture comes in the form of photos, notes, and bookmarks. Based on this survey we built Scraps, which consists of two parts: 1) a mobile app that makes it easy for people to capture and add context to information from their phone, and 2) a Word sidebar that helps them later link that information to a document on their desktop. In a field study with 11 information workers, we find that Scraps streamlined the process of capturing and using document-related information, and enabled people to focus on writing over integrating captured information.

Shuqi Lu, Di He, Chenyan Xiong, Guolin Ke, Waleed Malik, Zhicheng Dou, Paul Bennett, Tie-Yan Liu, Arnold Overwijk. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Information overload is a prevalent challenge in many high-value domains. A prominent case in point is the explosion of the biomedical literature on COVID-19, which swelled to hundreds of thousands of papers in a matter of months. In general, biomedical literature expands by two papers every minute, totalling over a million new papers every year. Search in the biomedical realm, and many other vertical domains is challenging due to the scarcity of direct supervision from click logs. Self-supervised learning has emerged as a promising direction to overcome the annotation bottleneck. We propose a general approach for vertical search based on domain-specific pretraining and present a case study for the biomedical domain. Despite being substantially simpler and not using any relevance labels for training or development, our method performs comparably or better than the best systems in the official TREC-COVID evaluation, a COVID-related biomedical search competition. Using distributed computing in modern cloud infrastructure, our system can scale to tens of millions of articles on PubMed and has been deployed as Microsoft Biomedical Search, a new search experience for biomedical literature: https://aka.ms/biomedsearch.

We present a self-supervised learning framework, COCO-LM, that pretrains Language Models by COrrecting and COntrasting corrupted text sequences. Following ELECTRA-style pretraining, COCO-LM employs an auxiliary language model to corrupt text sequences, upon which it constructs two new tasks for pretraining the main model. The first token-level task, Corrective Language Modeling, is to detect and correct tokens replaced by the auxiliary model, in order to better capture token-level semantics. The second sequence-level task, Sequence Contrastive Learning, is to align text sequences originated from the same source input while ensuring uniformity in the representation space. Experiments on GLUE and SQuAD demonstrate that COCO-LM not only outperforms recent state-of-the-art pretrained models in accuracy, but also improves pretraining efficiency. It achieves the MNLI accuracy of ELECTRA with 50% of its pretraining GPU hours. With the same pretraining steps of standard base/large-sized models, COCO-LM outperforms the previous best models by 1+ GLUE average points.

In many domains, user preferences are similar locally within like-minded subgroups of users, but typically differ globally between those subgroups. Local recommendation models were shown to substantially improve top-K recommendation performance in such settings. However, existing local models do not scale to large-scale datasets with an increasing number of subgroups and do not support inductive recommendations for users not appearing in the training set. Key reasons for this are that subgroup detection and recommendation get implemented as separate steps in the model or that local models are explicitly instantiated for each subgroup. In this paper, we propose an End-to-end Local Factor Model (Elfm) which overcomes these limitations by combining both steps and incorporating local structures through an inductive bias. Our model can be optimized end-to-end and supports incremental inference, does not require a full separate model for each subgroup, and has overall small memory and computational costs for incorporating local structures. Empirical results show that our method substantially improves recommendation performance on large-scale datasets with millions of users and items with considerably smaller model size. Our user study also shows that our approach produces coherent item subgroups which could aid in the generation of explainable recommendations.

We introduce the novel research problem of task recognition in daily life. We recognize tasks such as project management, planning, meal-breaks, communication, documentation, and family care. We capture Cyber, Physical, and Social (CPS) activities of 17 participants over four weeks using device-based sensing, app activity logging, and an experience sampling methodology. Our cohort includes students, casual workers, and professionals, forming the first real-world context-rich task behaviour dataset. We model CPS activities across different task categories, results highlight the importance of considering the CPS feature sets in modelling, especially work-related tasks.

Item-to-item recommendation (e.g., “People who like this also like...”) is a ubiquitous and important type of recommendation in real-world systems. Observational data from historical interaction logs abound in these settings. However, since virtually all observational data exhibit biases, such as time-in-inventory or interface biases, it is crucial that recommender algorithms account for these biases. In this paper, we develop a principled approach for item-to-item recommendation based on causal inference and present a practical and highly effective method for estimating the causal parameters from a small annotated dataset. Empirically, we find that our approach substantially improves upon existing methods while requiring only small amounts of annotated data.

Recent progress in deep learning has brought tremendous improvements in conversational AI, leading to a plethora of commercial conversational services that allow naturally spoken interactions, increasing the need for more human-centric interactions in IR. As a result, we have witnessed a resurgent interest in developing modern CIR systems in research communities and industry. This tutorial presents recent advances in CIR, focusing mainly on neural approaches and new applications developed in the past five years. Our goal is to provide a thorough and in-depth overview of the general definition of CIR, the components of CIR systems, new applications raised for its conversational aspects, and the (neural) techniques recently developed for it.

In this work, we focus on the contextual document ranking task, which deals with the challenge of user interaction modeling for conversational search. Given a history of user feedback behaviors, such as issuing a query, clicking a document, and skipping a document, we propose to introduce behavior awareness to a neural ranker, resulting in a Hierarchical Behavior Aware Transformers (HBA-Transformers) model. The hierarchy is composed of an intra-behavior attention layer and an inter-behavior attention layer to let the system effectively distinguish and model different user behaviors. Our extensive experiments on the AOL session dataset demonstrate that the hierarchical behavior aware architecture is more powerful than a simple combination of history behaviors. Besides, we analyze the conversational property of queries. We show that coherent sessions tend to be more conversational and thus are more demanding in terms of considering history user behaviors.

Many interactive online systems, such as social media platforms or news sites, provide personalized experiences through recommendations or news feed customization based on people's feedback and engagement on individual items (e.g., liking items). In this paper, we investigate how we can support a greater degree of user control in such systems by changing the way the system allows people to gauge the consequences of their feedback actions. To this end, we consider two important aspects of how the system responds to feedback actions: (i) immediacy, i.e., how quickly the system responds with an update, and (ii) visibility, i.e., whether or not changes will get highlighted. We used both an in-lab qualitative study and a large-scale crowd-sourced study to examine the impact of these factors on people's reported preferences and observed behavioral metrics. We demonstrate that UX design which enables people to preview the impact of their actions and highlights changes results in a higher reported transparency, an overall preference for this design, and a greater selectivity in which items are liked.

Conversational query rewriting aims to reformulate a concise conversational query to a fully specified, context-independent query that can be effectively handled by existing information retrieval systems. This paper presents a few-shot generative approach to conversational query rewriting. We develop two methods, based on rules and self-supervised learning, to generate weak supervision data using large amounts of ad hoc search sessions, and to fine-tune GPT-2 to rewrite conversational queries. On the TREC Conversational Assistance Track, our weakly supervised GPT-2 rewriter improves the state-of-the-art ranking accuracy by 12%, only using very limited amounts of manual query rewrites. In the zero-shot learning setting, the rewriter still gives a comparable result to previous state-of-the-art systems. Our analyses reveal that GPT-2 effectively picks up the task syntax and learns to capture context dependencies, even for hard cases that involve group references and long-turn dependencies.

Asking clarifying questions in response to search queries has been recognized as a useful technique for revealing the underlying intent of the query. Clarification has applications in retrieval systems with different interfaces, from the traditional web search interfaces to the limited bandwidth interfaces as in speech-only and small screen devices. Generation and evaluation of clarifying questions have been recently studied in the literature. However, user interaction with clarifying questions is relatively unexplored. In this paper, we conduct a comprehensive study by analyzing large-scale user interactions with clarifying questions in a major web search engine. In more detail, we analyze the user engagements received by clarifying questions based on different properties of search queries, clarifying questions, and their candidate answers. We further study click bias in the data, and show that even though reading clarifying questions and candidate answers does not take significant efforts, there still exist some position and presentation biases in the data. We also propose a model for learning representation for clarifying questions based on the user interaction data as implicit feedback. The model is used for re-ranking a number of automatically generated clarifying questions for a given query. Evaluation on both click data and human labeled data demonstrates the high quality of the proposed method.

Individuals' personal information collections (their emails, files, appointments, web searches, contacts, etc) offer a wealth of insights into the organization and structure of their everyday lives. In this paper we address the task of learning representations of personal information items to capture individuals' ongoing activities, such as projects and tasks: Such representations can be used in activity-centric applications like personal assistants, email clients, and productivity tools to help people better manage their data and time. We propose a graph-based approach that leverages the inherent interconnected structure of personal information collections, and derive efficient, exact techniques to incrementally update representations as new data arrive. We demonstrate the strengths of our graph-based representations against competitive baselines in a novel intrinsic rating task and an extrinsic recommendation task.

This paper studies a new scenario in conversational search, conversational question suggestion, which leads search engine users to more engaging experiences by suggesting interesting, informative, and useful follow-up questions. We first establish a novel evaluation metric, usefulness, which goes beyond relevance and measures whether the suggestions provide valuable information for the next step of a user’s journey, and construct a public benchmark for useful question suggestion. Then we develop two suggestion systems, a BERT based ranker and a GPT-2 based generator, both trained with novel weak supervision signals that convey past users’ search behaviors in search sessions. The weak supervision signals help ground the suggestions to users’ information-seeking trajectories: we identify more coherent and informative sessions using encodings, and then weakly supervise our models to imitate how users transition to the next state of search. Our offline experiments demonstrate the crucial role our “next-turn” inductive training plays in improving usefulness over a strong online system. Our online A/B test in Bing shows that our more useful question suggestions receive 8% more user clicks than the previous system.

As AI technology advances, it offers promising opportunities to improve educational outcomes when integrated with an overall learning experience. We investigate forward-looking interactive reading experiences that leverage both automatic question generation and analysis of attention signals, such as gaze tracking, to improve short- and long-term learning outcomes. We aim to expand the known pedagogical benefits of adjunct questions to more general reading scenarios, by investigating the benefits of adjunct questions generated after participants attend to passages in an article, based on their gaze behavior. We also compare the effectiveness of manually-written questions with those produced by Automatic Question Generation (AQG). We further investigate gaze and reading patterns indicative of low vs. high learning in both short- and long-term scenarios (one-week followup). We show AQG-generated adjunct questions have promise as a way to scale to a wide variety of reading material where the cost of manually curating questions may be prohibitive.

Search queries are often short, and the underlying user intent may be ambiguous. This makes it challenging for search engines to predict possible intents, only one of which may pertain to the current user. To address this issue, search engines often diversify the result list and present documents relevant to multiple intents of the query. An alternative approach is to ask the user a question to clarify her information need. Asking clarifying questions is particularly important for scenarios with “limited bandwidth” interfaces, such as speech-only and small-screen devices. In addition, our user studies and large-scale online experiments show that asking clarifying questions is also useful in web search. Although some recent studies have pointed out the importance of asking clarifying questions, generating them for open-domain search tasks remains unstudied and is the focus of this paper. Lack of training data even within major search engines for this task makes it challenging. To mitigate this issue, we first identify a taxonomy of clarification for open-domain search queries by analyzing large-scale query reformulation data sampled from Bing search logs. This taxonomy leads us to a set of question templates and a simple yet effective slot filling algorithm. We further use this model as a source of weak supervision to automatically generate clarifying questions for training. Furthermore, we propose supervised and reinforcement learning models for generating clarifying questions learned from weak supervision data. We also investigate methods for generating candidate answers for each clarifying question, so users can select from a set of pre-defined answers. Human evaluation of the clarifying questions and candidate answers for hundreds of search queries demonstrates the effectiveness of the proposed solutions.

Advances in artificial intelligence (AI) frame opportunities and challenges for user interface design. Principles for human-AI interaction have been discussed in the human-computer interaction community for over two decades, but more study and innovation are needed in light of advances in AI and the growing uses of AI technologies in human-facing applications. We propose 18 generally applicable design guidelines for human-AI interaction. These guidelines are validated through multiple rounds of evaluation including a user study with 49 design practitioners who tested the guidelines against 20 popular AI-infused products. The results verify the relevance of the guidelines over a spectrum of interaction scenarios and reveal gaps in our knowledge, highlighting opportunities for further research. Based on the evaluations, we believe the set of design guidelines can serve as a resource to practitioners working on the design of applications and features that harness AI technologies, and to researchers interested in the further development of human-AI interaction design principles.

AI technologies have been incorporated into many end-user applications. However, expectations of the capabilities of such systems vary among people. Furthermore, bloated expectations have been identified as negatively affecting perception and acceptance of such systems. Although the intelligibility of ML algorithms has been well studied, there has been little work on methods for setting appropriate expectations before the initial use of an AI-based system. In this work, we use a Scheduling Assistant - an AI system for automated meeting request detection in free-text email - to study the impact of several methods of expectation setting. We explore two versions of this system with the same 50% level of accuracy of the AI component but each designed with a different focus on the types of errors to avoid (avoiding False Positives vs. False Negatives). We show that such different focus can lead to vastly different subjective perceptions of accuracy and acceptance. Further, we design expectation adjustment techniques that prepare users for AI imperfections and result in a significant increase in acceptance.

Human activity is one of the most important pieces of context affecting an individual's information needs. Understanding the relationship between activities, time, location, and other contextual features can improve the quality of various intelligent systems, including contextual search engines, task managers, digital personal assistants, chat bots, and recommender systems.

Intelligent assistants can serve many purposes, including entertainment (e.g. playing music), home automation, and task management (e.g. timers, reminders). The role of these assistants is evolving to also support people engaged in work tasks, in workplaces and beyond. To design truly useful intelligent assistants for work, it is important to better understand the work tasks that people are performing. Based on a survey of 401 respondents' daily tasks and activities in a work setting, we present a classification of work-related tasks, and analyze their key characteristics, including the frequency of their self-reported tasks, the environment in which they undertake the tasks, and which, if any, electronic devices are used. We also investigate the cyber, physical, and social aspects of tasks. Finally, we reflect on how intelligent assistants could influence and help people in a work environment to complete their tasks, and synthesize our findings to provide insight on the future of intelligent assistants in support of amplifying personal productivity.

Email continues to be one of the most important means of online communication. People spend a significant amount of time sending, reading, searching and responding to email in order to manage tasks, exchange information, etc. In this paper, we study intent identification in workplace email. We use a large scale publicly available email dataset to characterize intents in enterprise email and propose methods for improving intent identification in email conversations. Previous work focused on classifying email messages into broad topical categories or detecting sentences that contain action items or follow certain speech acts. In this work, we focus on sentence-level intent identification and study how incorporating more context (such as the full message body and other metadata) could improve the performance of the intent identification models. We experiment with several models for leveraging context including both classical machine learning and deep learning approaches. We show that modeling the interaction between sentence and context can significantly improve the performance.

This paper presents GEneric iNtent Encoder (GEN Encoder) which learns a distributed representation space for user intent in search. Leveraging large scale user clicks from Bing search logs as weak supervision of user intent, GEN Encoder learns to map queries with shared clicks into similar embeddings end-to-end and then fine-tunes on multiple paraphrase tasks. Experimental results on an intrinsic evaluation task - query intent similarity modeling - demonstrate GEN Encoder's robust and significant advantages over previous representation methods. Ablation studies reveal the crucial role of learning from implicit user feedback in representing user intent and the contributions of multi-task learning in representation generality. We also demonstrate that GEN Encoder alleviates the sparsity of tail search traffic and cuts down half of the unseen queries by using an efficient approximate nearest neighbor search to effectively identify previous queries with the same search intent. Finally, we demonstrate distances between GEN encodings reflect certain information seeking behaviors in search sessions.

Recommender systems rely heavily on the predictive accuracy of the learning algorithm. Most work on improving accuracy has focused on the learning algorithm itself. We argue that this algorithmic focus is myopic. In particular, since learning algorithms generally improve with more and better data, we propose shaping the feedback generation process as an alternate and complementary route to improving accuracy. To this effect, we explore how changes to the user interface can impact the quality and quantity of feedback data -- and therefore the learning accuracy. Motivated by information foraging theory, we study how feedback quality and quantity are influenced by interface design choices along two axes: information scent and information access cost. We present a user study of these interface factors for the common task of picking a movie to watch, showing that these factors can effectively shape and improve the implicit feedback data that is generated while maintaining the user experience.

We study the problem of representing and recommending products for grocery shopping. We carefully investigate grocery transaction data and observe three important patterns: products within the same basket complement each other in terms of functionality (complementarity); users tend to purchase products that match their preferences (compatibility); and a significant fraction of users repeatedly purchase the same products over time (loyalty). Unlike conventional e-commerce settings, complementarity and loyalty are particularly predominant in the grocery shopping domain. This motivates a new representation learning approach to leverage complementarity and compatibility holistically, as well as a new recommendation approach to explicitly account for users' 'must-buy' purchases in addition to their overall preferences and needs. Doing so not only improves product classification and recommendation performance on both public and proprietary transaction data covering various grocery store types, but also reveals interesting findings about the relationships between preferences, necessity, and loyalty in consumer purchases.

In this paper, we study how to leverage calendar information to help with email re-finding using a zero-query prototype, Calendar-Aware Proactive Email Recommender System (CAPERS). CAPERS proactively selects and displays potentially useful emails to users based on their upcoming calendar events with a particular focus on meeting preparation. We approach this problem domain through a survey, a task-based experiment, and a field experiment comparing multiple email recommenders in a large technology company. We first show that a large proportion of email access is related to meetings and then study the effects of four email recommenders on user perception and engagement taking into account four categories of factors: the amount of email content, email recency, calendar-email content match, and calendar-email people match. We demonstrate that these factors all positively predict the usefulness of emails to meeting preparation and that calendar-email content match is the most important. We study the effects of different machine learning models for predicting usefulness and find that an online-learned linear model doubles user engagement compared with the baselines, which suggests the benefit of continuous online learning.

Any learning algorithm for recommendation faces a fundamental trade-off between exploiting partial knowledge of a user»s interests to maximize satisfaction in the short term and discovering additional user interests to maximize satisfaction in the long term. To enable discovery, a machine learning algorithm typically elicits feedback on items it is uncertain about, which is termed algorithmic exploration in machine learning. This exploration comes with a cost to the user, since the items an algorithm chooses for exploration frequently turn out to not match the user»s interests. In this paper, we study how users tolerate such exploration and how presentation strategies can mitigate the exploration cost. To this end, we conduct a behavioral study with over 600 people, where we vary how algorithmic exploration is mixed into the set of recommendations. We find that users respond non-linearly to the amount of exploration, where some exploration mixed into the set of recommendations has little effect on short-term satisfaction and behavior. For long-term satisfaction, the overall goal is to learn via exploration about the items presented. We therefore also analyze the quantity and quality of implicit feedback signals such as clicks and hovers, and how they vary with different amounts of mix-in exploration. Our findings provide insights into how to design presentation strategies for algorithmic exploration in interactive recommender systems, mitigating the short-term costs of algorithmic exploration while aiming to elicit informative feedback data for learning.

"Remind me to get milk later this afternoon." In communications and planning, people often express uncertainty about time using imprecise temporal expressions (ITEs). Unfortunately, modern virtual assistants often lack system support to capture the intents behind these expressions. This can result in unnatural interactions and undesirable interruptions (e.g., having a work reminder delivered at 12pm when out at lunch, because the user said "this afternoon"). In this paper we explore existing practices, expectations, and preferences surrounding the use of ITEs. Our mixed methods approach employs surveys, interviews, and an analysis of a large corpus of written communications. We find that people frequently use a diverse set of ITEs in both communication and planning. These uses reflect a variety of motivations, such as conveying uncertainty or task priority. In addition, we find that people have a variety of expectations about time input and management when interacting with virtual assistants. We conclude with design implications for future virtual assistants.

Email has been central to online communication for the past two decades. Through constant use, new information flows are being defined around users' interactions with emails. Alongside traditional messages, the email inbox is an always-available repository of to-do lists, reminders, files and notes. In this paper, we investigate the use of self-addressed emails (self-Es) as an information management tool, by analysing both: (i) responses to a survey about email use; and (ii) a collection of user donated self-addressed emails. Our results show that sending self-Es is a frequent behaviour among the users we questioned. In addition, we find that to-dos and reminders are the most popular type of information contained in emails-to-self. Our findings have direct implications for the development of systems that support novel interactions with information flows centred around email.

Email is still among the most popular online activities. People spend a significant amount of time sending, reading and responding to email in order to communicate with others, manage tasks and archive personal information. Most previous research on email is based on either relatively small data samples from user surveys and interviews, or on consumer email accounts such as those from Yahoo! Mail or Gmail. Much less has been published on how people interact with enterprise email even though it contains less automatically generated commercial email and involves more organizational behavior than is evident in personal accounts. In this paper, we extend previous work on predicting email reply behavior by looking at enterprise settings and considering more than dyadic communications. We characterize the influence of various factors such as email content and metadata, historical interaction features and temporal features on email reply behavior. We also develop models to predict whether a recipient will reply to an email and how long it will take to do so. Experiments with the publicly-available Avocado email collection show that our methods outperform all baselines with large gains. We also analyze the importance of different features on reply behavior predictions. Our findings provide new insights about how people interact with enterprise email and have implications for the design of the next generation of email clients.

Software applications often use classification models to trigger specialized experiences for users. Search engines, for example, use query classifiers to trigger specialized "instant answer" experiences where information satisfying the user query is shown directly on the result page, and email applications use classification models to automatically move messages to a spam folder. When such applications have acceptable default (i.e., non-specialized) behavior, users are often more sensitive to failures in model precision than failures in model recall. In this paper, we consider model-selection algorithms for these precision-constrained scenarios. We develop adaptive model-selection algorithms to identify, using as few samples as possible, the best classifier from among a set of (precision) qualifying classifiers. We provide statistical correctness and sample complexity guarantees for our algorithms. We show with an empirical validation that our algorithms work well in practice.

Automated personal assistants such as Siri, Cortana, and Google Now provide services to help users accomplish tasks, including tools to set reminders. We study how people specify and use reminders. Our study analyzes a sample of six months of logs of user-specified reminders from Cortana (Microsoft's intelligent personal assistant), the first large-scale analysis of such reminders. We focus our analyses on time-based reminders, the most common type of reminder found in the logs. We perform a data-driven analysis to identify common categories of tasks that give rise to these reminders across a large number of users, and we arrange these tasks into a taxonomy. We identify temporal patterns linked to the type of task, time of creation, and terms in the reminder text. Finally, we show that these patterns generalize by addressing a prediction task. Specifically, we show that a reminder's creation time is a strong feature in predicting the notification time, and that including the reminder text further improves prediction accuracy. The results have implications for the design of systems aimed at helping people to complete tasks and to plan future activities.

In this paper, we study shortlists as an interface component for recommender systems with the dual goal of supporting the user's decision process, as well as improving implicit feedback elicitation for increased recommendation quality. A shortlist is a temporary list of candidates that the user is currently considering, e.g., a list of a few movies the user is currently considering for viewing. From a cognitive perspective, shortlists serve as digital short-term memory where users can off-load the items under consideration -- thereby decreasing their cognitive load. From a machine learning perspective, adding items to the shortlist generates a new implicit feedback signal as a by-product of exploration and decision making which can improve recommendation quality. Shortlisting therefore provides additional data for training recommendation systems without the increases in cognitive load that requesting explicit feedback would incur.

Query-based triggers play a crucial role in modern search systems, e.g., in deciding when to display direct answers on result pages. We address a common scenario in designing such triggers for real-world settings where positives are rare and search providers possess only a small seed set of positive examples to learn query classification models. We choose the critical domain of self-harm intent detection to demonstrate how such small seed sets can be expanded to create meaningful training data with a sizable fraction of positive examples. Our results show that with our method, substantially more positive queries can be found compared to plain random sampling. Additionally, we explored the effectiveness of traditional active learning approaches on classification performance and found that maximum uncertainty performs the best among several other techniques that we considered.

Considerable work in web page classification has focused on incorporating the topical structure of the web (e.g., the hyperlink graph) to improve prediction accuracy. However, the majority of work has primarily focused on relational or graph-based methods that are impractical to run at scale or in an online environment. This raises the question of whether it is possible to leverage the topical structure of the web while incurring nearly no additional prediction-time cost. To this end, we introduce an approach which adjusts a page content-only classification from that obtained with a global prior to the posterior obtained by incorporating a prior which reflects the topic cohesion of the site. Using ODP data, we empirically demonstrate that our approach yields significant performance increases over a range of topics.

It has been suggested that online search and retrieval contributes to the intellectual isolation of users within their preexisting ideologies, where people's prior views are strengthened and alternative viewpoints are infrequently encountered. This so-called "filter bubble" phenomenon has been called out as especially detrimental when it comes to dialog among people on controversial, emotionally charged topics, such as the labeling of genetically modified food, the right to bear arms, the death penalty, and online privacy. We seek to identify and study information-seeking behavior and access to alternative versus reinforcing viewpoints following shocking, emotional, and large-scale news events. We choose for a case study to analyze search and browsing on gun control/rights, a strongly polarizing topic for both citizens and leaders of the United States. We study the period of time preceding and following a mass shooting to understand how its occurrence, follow-on discussions, and debate may have been linked to changes in the patterns of searching and browsing. We employ information-theoretic measures to quantify the diversity of Web domains of interest to users and understand the browsing patterns of users. We use these measures to characterize the influence of news events on these web search and browsing patterns.

Many individuals on social networking sites provide traits about themselves, such as interests or demographics. Social networking sites can use this information to provide better content to match their users' interests, such as recommending scheduled events or various relevant products. These tasks require accurate probability estimates to determine the correct answer to return. Relational machine learning (RML) is an excellent framework for these problems as it jointly models the user labels given their attributes and the relational structure. Further, semi-supervised learning methods could enable RML methods to exploit the large amount of unlabeled data in networks.

Queries issued to a search engine are often under-specified or ambiguous. The user's search context or background may provide information that disambiguates their information need in order to automatically predict and issue a more effective query. The disambiguation can take place at different stages of the retrieval process. For instance, contextual query suggestions may be computed and recommended to users on the result page when appropriate, an approach that does not require modifying the original query's results. Alternatively, the search engine can attempt to provide efficient access to new relevant documents by injecting these documents directly into search results based on the user's context.

editorial Free AccessOverview of the Special Issue on Contextual Search and Recommendation Editors: Paul N. Bennett View Profile , Kevyn Collins-Thompson View Profile , Diane Kelly View Profile , Ryen W. White View Profile , Yi Zhang View Profile Authors Info & Claims ACM Transactions on Information SystemsVolume 33Issue 1March 2015 Article No.: 1epp 1–7https://doi.org/10.1145/2691351Published:17 March 2015Publication History 12citation543DownloadsMetricsTotal Citations12Total Downloads543Last 12 Months32Last 6 weeks10 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteeReaderPDF

Many interesting domains in machine learning can be viewed as networks, with relationships (e.g., friendships) connecting items (e.g., individuals). The Active Exploration (AE) task is to identify all items in a network with a desired trait (i.e., positive labels) given only partial information about the network. The AE process iteratively queries for labels or network structure within a limited budget; thus, accurate predictions prior to making each query is critical to maximizing the number of positives gathered. However, the targeted AE query process produces partially observed networks that can create difficulties for predictive modeling. In particular, we demonstrate that these partial networks can exhibit extreme label correlation bias, which makes it difficult for conventional relational learning methods to accurately estimate relational parameters. To overcome this issue, we model the joint distribution of possible edges and labels to improve learning and inference. Our proposed method, Probabilistic Relational Expectation Maximization (PR-EM), is the first AE approach to accurately learn the complex dependencies between attributes, labels, and structure to improve predictions. PR-EM utilizes collective inference over the missing relationships in the partial network to jointly infer unknown item traits. Further, we develop a linear inference algorithm to facilitate efficient use of PR-EM in large networks. We test our approach on four real world networks, showing that AE with PR-EM gathers significantly more positive items compared to state-of-the-art methods.

The prevalence of datasets that can be represented as networks has recently fueled a great deal of work in the area of Relational Machine Learning (RML). Due to the statistical correlations between linked nodes in the network, many RML methods focus on predicting node features (i.e., labels) using the network relationships. However, many domains are comprised of a single, partially-labeled network. Thus, relational versions of Expectation Maximization (i.e., R-EM), which jointly learn parameters and infer the missing labels, can outperform methods that learn parameters from the labeled data and apply them for inference on the unlabeled nodes. Although R-EM methods can significantly improve predictive performance in networks that are densely labeled, they do not achieve the same gains in sparsely labeled networks and can perform worse than RML methods. In this work, we show the fixed-point methods that R-EM uses for approximate learning and inference result in errors that prevent convergence in sparsely labeled networks. We then propose two methods that do not experience this problem. First, we develop a Relational Stochastic EM (R-SEM) method, which uses stochastic parameters that are not as susceptible to approximation errors. Then we develop a Relational Data Augmentation (R-DA) method, which integrates over a range of stochastic parameter values for inference. R-SEM and R-DA can use any collective RML algorithm for learning and inference in partially labeled networks. We analyze their performance with two RML learners over four real world datasets, and show that they outperform independent learning, RML and R-EM -- particularly in sparsely labeled networks.

Current research on Web search has focused on optimizing and evaluating single queries. However, a significant fraction of user queries are part of more complex tasks [Jones and Klinkner 2008] which span multiple queries across one or more search sessions [Liu and Belkin 2010; Kotov et al. 2011]. An ideal search engine would not only retrieve relevant results for a user's particular query but also be able to identify when the user is engaged in a more complex task and aid the user in completing that task [Morris et al. 2008; Agichtein et al. 2012]. Toward optimizing whole-session or task relevance, we characterize and address the problem of intrinsic diversity (ID) in retrieval [Radlinski et al. 2009], a type of complex task that requires multiple interactions with current search engines. Unlike existing work on extrinsic diversity [Carbonell and Goldstein 1998; Zhai et al. 2003; Chen and Karger 2006] that deals with ambiguity in intent across multiple users, ID queries often have little ambiguity in intent but seek content covering a variety of aspects on a shared theme. In such scenarios, the underlying needs are typically exploratory, comparative, or breadth-oriented in nature. We identify and address three key problems for ID retrieval: identifying authentic examples of ID tasks from post-hoc analysis of behavioral signals in search logs; learning to identify initiator queries that mark the start of an ID search task; and given an initiator query, predicting which content to prefetch and rank.

With massive amounts of data being generated and stored ubiquitously in every discipline and every aspect of our daily life, how to handle such big data poses many challenging issues to researchers in data and information systems. The participants of CIKM 2013 are active researchers on large scale data, information and knowledge management, from multiple disciplines, including database systems, data mining, information retrieval, human-computer interaction, and knowledge or information management.

There is an increasing amount of structure on the web as a result of modern web languages, user tagging and annotation, emerging robust NLP tools, and an ever growing volume of linked data. These meaningful, semantic, annotations hold the promise to significantly enhance information access, by enhancing the depth of analysis of today's systems. Currently, we have only started exploring the possibilities and only begin to understand how these valuable semantic cues can be put to fruitful use. ESAIR'13 focuses on two of the most challenging aspects to address in the coming years. First, there is a need to include the currently emerging knowledge resources (such as DBpedia, Freebase) as underlying semantic model giving access to an unprecedented scope and detail of factual information. Second, there is a need to include annotations beyond the topical dimension (think of sentiment, reading level, prerequisite level, etc) that contain vital cues for matching the specific needs and profile of the searcher at hand.

Current research on web search has focused on optimizing and evaluating single queries. However, a significant fraction of user queries are part of more complex tasks [20] which span multiple queries across one or more search sessions [26,24]. An ideal search engine would not only retrieve relevant results for a user's particular query but also be able to identify when the user is engaged in a more complex task and aid the user in completing that task [29,1]. Toward optimizing whole-session or task relevance, we characterize and address the problem of intrinsic diversity (ID) in retrieval [30], a type of complex task that requires multiple interactions with current search engines. Unlike existing work on extrinsic diversity [30] that deals with ambiguity in intent across multiple users, ID queries often have little ambiguity in intent but seek content covering a variety of aspects on a shared theme. In such scenarios, the underlying needs are typically exploratory, comparative, or breadth-oriented in nature. We identify and address three key problems for ID retrieval: identifying authentic examples of ID tasks from post-hoc analysis of behavioral signals in search logs; learning to identify initiator queries that mark the start of an ID search task; and given an initiator query, predicting which content to prefetch and rank.

Web search engines frequently show the same documents repeatedly for different queries within the same search session, in essence forgetting when the same documents were already shown to users. Depending on previous user interaction with the repeated results, and the details of the session, we show that sometimes the repeated results should be promoted, while some other times they should be demoted.

Inferring rankings over elements of a set of objects, such as documents or images, is a key learning problem for such important applications as Web search and recommender systems. Crowdsourcing services provide an inexpensive and efficient means to acquire preferences over objects via labeling by sets of annotators. We propose a new model to predict a gold-standard ranking that hinges on combining pairwise comparisons via crowdsourcing. In contrast to traditional ranking aggregation methods, the approach learns about and folds into consideration the quality of contributions of each annotator. In addition, we minimize the cost of assessment by introducing a generalization of the traditional active learning scenario to jointly select the annotator and pair to assess while taking into account the annotator quality, the uncertainty over ordering of the pair, and the current model uncertainty. We formalize this as an active learning strategy that incorporates an exploration-exploitation tradeoff and implement it using an efficient online Bayesian updating scheme. Using simulated and real-world data, we demonstrate that the active learning strategy achieves significant reductions in labeling cost while maintaining accuracy.

Most research in Web search personalization models users as static or slowly evolving entities with a given set of preferences defined by their past behavior. However, recent publications as well as empirical evidence suggest that for a significant number of search sessions, users diverge from their regular search profiles in order to satisfy atypical, limited-duration information needs. In this work, we conduct a large-scale inspection of real-life search sessions to further understand this scenario. Subsequently, we design an automatic means of detecting and supporting such atypical sessions. We demonstrate significant improvements over state-of-the-art Web search personalization techniques by accounting for the typicality of search sessions. The proposed method is evaluated based on Web-scale search session data spanning several months of user activity.

Accurate prediction of changing web page content improves a variety of retrieval and web related components. For example, given such a prediction algorithm one can both design a better crawling strategy that only recrawls pages when necessary as well as a proactive mechanism for personalization that pushes content associated with user revisitation directly to the user. While many techniques for modeling change have focused simply on past change frequency, our work goes beyond that by additionally studying the usefulness in page change prediction of: the page's content; the degree and relationship among the prediction page's observed changes; the relatedness to other pages and the similarity in the types of changes they undergo. We present an expert prediction framework that incorporates the information from these other signals more effectively than standard ensemble or basic relational learning techniques. In an empirical analysis, we find that using page content as well as related pages significantly improves prediction accuracy and compare it to common approaches. We present numerous similarity metrics to identify related pages and focus specifically on measures of temporal content similarity. We observe that the different metrics yield related pages that are qualitatively different in nature and have different effects on the prediction performance.

Many important search tasks require multiple search sessions to complete. Tasks such as travel planning, large purchases, or job searches can span hours, days, or even weeks. Inevitably, life interferes, requiring the searcher either to recover the "state" of the search manually (most common), or plan for interruption in advance (unlikely). The goal of this work is to better understand, characterize, and automatically detect search tasks that will be continued in the near future. To this end, we analyze a query log from the Bing Web search engine to identify the types of intents, topics, and search behavior patterns associated with long-running tasks that are likely to be continued. Using our insights, we develop an effective prediction algorithm that significantly outperforms both the previous state-of-the-art method, and even the ability of human judges, to predict future task continuation. Potential applications of our techniques would allow a search engine to pre-emptively "save state" for a searcher (e.g., by caching search results), perform more targeted personalization, and otherwise better support the searcher experience for interrupted search tasks.

User behavior provides many cues to improve the relevance of search results through personalization. One aspect of user behavior that provides especially strong signals for delivering better relevance is an individual's history of queries and clicked documents. Previous studies have explored how short-term behavior or long-term behavior can be predictive of relevance. Ours is the first study to assess how short-term (session) behavior and long-term (historic) behavior interact, and how each may be used in isolation or in combination to optimally contribute to gains in relevance through search personalization. Our key findings include: historic behavior provides substantial benefits at the start of a search session; short-term session behavior contributes the majority of gains in an extended search session; and the combination of session and historic behavior out-performs using either alone. We also characterize how the relative contribution of each model changes throughout the duration of a session. Our findings have implications for the design of search systems that leverage user behavior to personalize the search experience.

Methods that reduce the amount of labeled data needed for training have focused more on selecting which documents to label than on which queries should be labeled. One exception to this (Long et al. 2010) uses expected loss optimization (ELO) to estimate which queries should be selected but is limited to rankers that predict absolute graded relevance. In this work, we demonstrate how to easily adapt ELO to work with any ranker and show that estimating expected loss in DCG is more robust than NDCG even when the final performance measure is NDCG.

Identifying content for which a user may search has a variety of applications, including ranking and recommendation. In this poster, we examine how pre-search context can be used to predict content that the user will seek before they have even specified a search query. We call this anticipatory search. Using a log-based approach, we compare different methods for predicting the content to be searched using different attributes of the pre-query context and behavioral signals from previous visitors to the most recent browse URL. Each method covers different cases and shows promise for query-free anticipatory search on the Web.

Many techniques for improving search result quality have been proposed. Typically, these techniques increase average effectiveness by devising advanced ranking features and/or by developing sophisticated learning to rank algorithms. However, while these approaches typically improve average performance of search results relative to simple baselines, they often ignore the important issue of robustness. That is, although achieving an average gain overall, the new models often hurt performance on many queries. This limits their application in real-world retrieval scenarios. Given that robustness is an important measure that can negatively impact user satisfaction, we present a unified framework for jointly optimizing effectiveness and robustness. We propose an objective that captures the tradeoff between these two competing measures and demonstrate how we can jointly optimize for these two measures in a principled learning framework. Experiments indicate that ranking models learned this way significantly decreased the worst ranking failures while maintaining strong average effectiveness on par with current state-of-the-art models.

A user's expertise or ability to understand a document on a given topic is an important aspect of that document's relevance. However, this aspect has not been well-explored in information retrieval systems, especially those at Web scale where the great diversity of content, users, and tasks presents an especially challenging search problem. To help improve our modeling and understanding of this diversity, we apply automatic text classifiers, based on reading difficulty and topic prediction, to estimate a novel type of profile for important entities in Web search -- users, websites, and queries. These profiles capture topic and reading level distributions, which we then use in conjunction with search log data to characterize and compare different entities.

We present a new approach for personalizing Web search results to a specific user. Ranking functions for Web search engines are typically trained by machine learning algorithms using either direct human relevance judgments or indirect judgments obtained from click-through data from millions of users. The rankings are thus optimized to this generic population of users, not to any specific user. We propose a generative model of relevance which can be used to infer the relevance of a document to a specific user for a search query. The user-specific parameters of this generative model constitute a compact user profile. We show how to learn these profiles from a user's long-term search history. Our algorithm for computing the personalized ranking is simple and has little computational overhead. We evaluate our personalization approach using historical search data from thousands of users of a major Web search engine. Our findings demonstrate gains in retrieval performance for queries with high ambiguity, with particularly large improvements for acronym queries.

Traditionally, search engines have ignored the reading difficulty of documents and the reading proficiency of users in computing a document ranking. This is one reason why Web search engines do a poor job of serving an important segment of the population: children. While there are many important problems in interface design, content filtering, and results presentation related to addressing children's search needs, perhaps the most fundamental challenge is simply that of providing relevant results at the right level of reading difficulty. At the opposite end of the proficiency spectrum, it may also be valuable for technical users to find more advanced material or to filter out material at lower levels of difficulty, such as tutorials and introductory texts. We show how reading level can provide a valuable new relevance signal for both general and personalized Web search. We describe models and algorithms to address the three key problems in improving relevance for search using reading difficulty: estimating user proficiency, estimating result difficulty, and re-ranking based on the difference between user and result reading level profiles. We evaluate our methods on a large volume of Web query traffic and provide a large-scale log analysis that highlights the importance of finding results at an appropriate reading level for the user.

Personalization of search results offers the potential for significant improvements in Web search. Among the many observable user attributes, approximate user location is particularly simple for search engines to obtain and allows personalization even for a first-time Web search user. However, acting on user location information is difficult, since few Web documents include an address that can be interpreted as constraining the locations where the document is relevant. Furthermore, many Web documents -- such as local news stories, lottery results, and sports team fan pages -- may not correspond to physical addresses, but the location of the user still plays an important role in document relevance. In this paper, we show how to infer a more general location relevance which uses not only physical location but a more general notion of locations of interest for Web pages. We compute this information using implicit user behavioral data, characterize the most location-centric pages, and show how location information can be incorporated into Web search ranking. Our results show that a substantial fraction of Web search queries can be significantly improved by incorporating location-based features.

The information needs of search engine users vary in complexity, depending on the task they are trying to accomplish. Some simple needs can be satisfied with a single query, whereas others require a series of queries issued over a longer period of time. While search engines effectively satisfy many simple needs, searchers receive little support when their information needs span session boundaries. In this work, we propose methods for modeling and analyzing user search behavior that extends over multiple search sessions. We focus on two problems: (i) given a user query, identify all of the related queries from previous sessions that the same user has issued, and (ii) given a multi-query task for a user, predict whether the user will return to this task in the future. We model both problems within a classification framework that uses features of individual queries and long-term user search behavior at different granularity. Experimental evaluation of the proposed models for both tasks indicates that it is possible to effectively model and analyze cross-session search behavior. Our findings have implications for improving search for complex information needs and designing search engine features to support cross-session search tasks.

The emergence of human computation systems, including Mechanical Turk and games with a purpose, has made it feasible to distribute relevance judgment tasks to workers over the Web. Most human computation systems assign tasks to individuals randomly, and such assignments may match workers with tasks that they may be unqualified or unmotivated to perform. We compare two groups of workers, those given a choice of queries to judge versus those who are not, in terms of their self-rated competence and their actual performance. Results show that when given a choice of task, workers choose ones for which they have greater expertise, interests, confidence, and understanding.

The web contains many duplicate and near-duplicate documents. Given that user satisfaction is negatively affected by redundant information in search results, a significant amount of research has been devoted to developing duplicate detection algorithms. However, most such algorithms rely solely on document content to detect duplication, ignoring the fact that a primary goal of duplicate detection is to identify documents that contain redundant information with respect to a particular user query. Similarly, although query-dependent result diversification algorithms compute a query-dependent ranking, they tend to do so on the basis of a query-independent content similarity score.

Deploying a classifier to large-scale systems such as the web requires careful feature design and performance evaluation. Evaluation is particularly challenging because these large collections frequently change. In this paper we adapt stratified sampling techniques to evaluate the precision of classifiers deployed in large-scale systems. We investigate different types of stratification strategies, and then we derive a new online sampling algorithm that incrementally approximates the theoretical optimal disproportionate sampling strategy. In experiments, the proposed algorithm significantly outperforms both simple random sampling as well as other types of stratified sampling, with an average reduction of about 20% in labeling effort to reach the same confidence and interval-bounds on precision

A query considered in isolation offers limited information about a searcher's intent. Query context that considers pre-query activity (e.g., previous queries and page visits), can provide richer information about search intentions. In this paper, we describe a study in which we developed and evaluated user interest models for the current query, its context (from pre-query session activity), and their combination, which we refer to as intent. Using large-scale logs, we evaluate how accurately each model predicts the user's short-term interests under various experimental conditions. In our study we: (i) determine the extent of opportunity for using context to model intent; (ii) compare the utility of different sources of behavioral evidence (queries, search result clicks, and Web page visits) for building predictive interest models, and; (iii) investigate optimally combining the query and its context by learning a model that predicts the context weight for each query. Our findings demonstrate significant opportunity in leveraging contextual information, show that context and source influence predictive accuracy, and show that we can learn a near-optimal combination of the query and context for each query. The findings can inform the design of search systems that leverage contextual information to better understand, model, and serve searchers' information needs.

Many have speculated that classifying web pages can improve a search engine's ranking of results. Intuitively results should be more relevant when they match the class of a query. We present a simple framework for classification-enhanced ranking that uses clicks in combination with the classification of web pages to derive a class distribution for the query. We then go on to define a variety of features that capture the match between the class distributions of a web page and a query, the ambiguity of a query, and the coverage of a retrieved result relative to a query's set of classes. Experimental results demonstrate that a ranker learned with these features significantly improves ranking over a competitive baseline. Furthermore, our methodology is agnostic with respect to the classification space and can be used to derive query classes for a variety of different taxonomies.

No abstract available.

While large-scale taxonomies--especially for web pages--have been in existence for some time, approaches to automatically classify documents into these taxonomies have met with limited success compared to the more general progress made in text classification. We argue that this stems from three causes: increasing sparsity of training data at deeper nodes in the taxonomy, error propagation where a mistake made high in the hierarchy cannot be recovered, and increasingly complex decision surfaces in higher nodes in the hierarchy. While prior research has focused on the first problem, we introduce methods that target the latter two problems--first by biasing the training distribution to reduce error propagation and second by propagating up "first-guess" expert information in a bottom-up manner before making a refined top down choice. Finally, we present an empirical study demonstrating that the suggested changes lead to 10--30% improvements in F1 scores versus an accepted competitive baseline, hierarchical SVMs.

We investigate using topic prediction data, as a summary of document content, to compute measures of search result quality. Unlike existing quality measures such as query clarity that require the entire content of the top-ranked results, class-based statistics can be computed efficiently online, because class information is compact enough to precompute and store in the index. In an empirical study we compare the performance of class-based statistics to their language-model counterparts for predicting two measures: query difficulty and expansion risk. Our findings suggest that using class predictions can offer comparable performance to full language models while reducing computation overhead.

We consider the problem of identifying the consensus ranking for the results of a query, given preferences among those results from a set of individual users. Once consensus rankings are identified for a set of queries, these rankings can serve for both evaluation and training of retrieval and learning systems. We present a novel approach to collecting the individual user preferences over image-search results: we use a collaborative game in which players are rewarded for agreeing on which image result is best for a query. Our approach is distinct from other labeling games because we are able to elicit directly the preferences of interest with respect to image queries extracted from query logs. As a source of relevance judgments, this data provides a useful complement to click data. Furthermore, the data is free of positional biases and is collected by the game without the risk of frustrating users with non-relevant results; this risk is prevalent in standard mechanisms for debiasing clicks. We describe data collected over 34 days from a deployed version of this game that amounts to about 18 million expressed preferences between pairs. Finally, we present several approaches to modeling this data in order to extract the consensus rankings from the preferences and better sort the search results for targeted queries.

There has been recent interest in collecting user or assessor preferences, rather than absolute judgments of relevance, for the evaluation or learning of ranking algorithms. Since measures like precision, recall, and DCG are defined over absolute judgments, evaluation over preferences will require new evaluation measures that explicitly model them. We describe a class of such measures and compare absolute and preference measures over a large TREC collection.

No abstract available.

Many corpus-based Machine Translation (MT) systems generate a number of partial translations which are then pieced together rather than immediately producing one overall translation. While this makes them more robust to ill-formed input, they are subject to disfluencies at phrasal translation boundaries even for well-formed input. We address this “boundary friction” problem by introducing a method that exploits overlapping phrasal translations and the increased confidence in translation accuracy they imply. We specify an efficient algorithm for producing translations using overlap. Finally, our empirical analysis indicates that this approach produces higher quality translations than the standard method of combining non-overlapping fragments generated by our Example-Based MT (EBMT) system in a peak-to-peak comparison.

Text classifiers that give probability estimates are more readily applicable in a variety of scenarios. For example, rather than choosing one set decision threshold, they can be used in a Bayesian risk model to issue a run-time decision which minimizes a user-specified cost function dynamically chosen at prediction time. However, the quality of the probability estimates is crucial. We review a variety of standard approaches to converting scores (and poor probability estimates) from text classifiers to high quality estimates and introduce new models motivated by the intuition that the empirical score distribution for the "extremely irrelevant", "hard to discriminate", and "obviously relevant" items are often significantly different. Finally, we analyze the experimental performance of these models over the outputs of two text classifiers. The analysis demonstrates that one of these models is theoretically attractive (introducing few new parameters while increasing flexibility), computationally efficient, and empirically preferable.

The intuition that different text classifiers behave in qualitatively different ways has long motivated attempts to build a better metaclassifier via some combination of classifiers. We introduce a probabilistic method for combining classifiers that considers the context-sensitive reliabilities of contributing classifiers. The method harnesses reliability indicators---variables that provide a valuable signal about the performance of classifiers in different situations. We provide background, present procedures for building metaclassifiers that take into consideration both reliability indicators and classifier outputs, and review a set of comparative studies undertaken to evaluate the methodology.