Understanding user intents in information access scenarios can help us provide more relevant and personalized search results and recommendations. However, analyzing user intents is not easy, especially for emerging forms of Web search such as Artificial Intelligence (AI)-driven chat. To understand user intents from retrospective log data, we need a way to label them with meaningful categories that capture their diversity and dynamics. Existing methods rely on manual or Machine-Learned (ML) labeling, which is either expensive or inflexible for large and dynamic datasets. Large Language Models (LLMs) could generate rich and relevant concepts, descriptions, and examples for user intents using log data of user interactions. However, using LLMs to generate a user intent taxonomy and applying it for a given Information Retrieval (IR) application can be problematic for two main reasons: (1) such a taxonomy is not externally validated; and (2) there may be an undesirable feedback loop if an LLM does both these tasks without external validation. To address this, we propose a new methodology with human experts and assessors to verify the quality of the LLM-generated taxonomy. We also present an end-to-end pipeline that uses an LLM with Human-in-the-Loop (HITL) to produce, refine, and apply labels for user intent analysis in log data. We demonstrate its effectiveness by uncovering new insights into user intents from search and chat logs from the Microsoft Bing Web search engine. The novelty in this research stems from the method for generating purpose-driven user intent taxonomies with strong validation. Our approach not only helps remove methodological and practical bottlenecks from intent-focused research, but also provides a new framework for generating, validating, and applying other kinds of taxonomies in a scalable and adaptable way, with reasonable human effort.

Ying-Chun Lin, Jennifer Neville, Jack Stokes, Longqi Yang, Tara Safavi, Mengting Wan, Scott Counts, Siddharth Suri, Reid Andersen, Xiaofeng Xu, Deepak Gupta, Sujay Kumar Jauhar, Xia Song, Georg Buscher, Saurabh Tiwary, Brent Hecht, Jaime Teevan. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Transforming unstructured text into structured and meaningful forms, organized by useful category labels, is a fundamental step in text mining for downstream analysis and application. However, most existing methods for producing label taxonomies and building text-based label classifiers still rely heavily on domain expertise and manual curation, making the process expensive and time-consuming. This is particularly challenging when the label space is under-specified and large-scale data annotations are unavailable. In this paper, we address these challenges with Large Language Models (LLMs), whose prompt-based interface facilitates the induction and use of large-scale pseudo labels. We propose TnT-LLM, a two-phase framework that employs LLMs to automate the process of end-to-end label generation and assignment with minimal human effort for any given use-case. In the first phase, we introduce a zero-shot, multi-stage reasoning approach which enables LLMs to produce and refine a label taxonomy iteratively. In the second phase, LLMs are used as data labelers that yield training samples so that lightweight supervised classifiers can be reliably built, deployed, and served at scale. We apply TnT-LLM to the analysis of user intent and conversational domain for Bing Copilot (formerly Bing Chat), an open-domain chat-based search engine. Extensive experiments using both human and automatic evaluation metrics demonstrate that TnT-LLM generates more accurate and relevant label taxonomies when compared against state-of-the-art baselines, and achieves a favorable balance between accuracy and efficiency for classification at scale.

User satisfaction metrics are an integral part of search engine development as they help system developers to understand and evaluate the quality of the user experience. Research to date has mostly focused on predicting success or frustration as a proxy for satisfaction. However, users' search experience is more complex than merely being either successful or not. As such, using success rate as a measure of satisfaction can be limiting. In this work, we propose the use of utility as a measure of searcher satisfaction. This concept represents the fulfillment a user receives from con-suming a service and explains how users aim to gain optimal overall satisfaction. Our utility metrics measure the user satisfac-tion by aggregating all their interaction with the search engine. These interactions are represented as a timeline of actions and their dwelltimes, where each action is classified as having a posi-tive or negative effect on the user. We examine sessions mined from Bing logs, with multi-point scale assessment of searcher satisfaction and show that utility is a better proxy for satisfaction compared to success. Leveraging that data, we design metrics of searcher satisfaction that assess the overall utility accumulated by a user during her search session. We use real user traffic from millions of users in an A/B setting to compare utility metrics to success rate metrics. We show that utility is a better metric for evaluating searcher satisfaction with the search engine, and a more sensitive and accurate metric when compared to predicting success. These metrics are currently adopted as the top-level met-ric for evaluating the thousands of A/B experiments that are run on Bing each year.

In this paper, we describe principles for designing metrics in the context of A/B experiments. We share some issues that comes up in designing such experiments and provide solutions to avoid such pitfalls.

Understanding how users examine result pages across a broad range of information needs is critical for search engine design. Cursor movements can be used to estimate visual attention on search engine results page (SERP) components, including traditional snippets, aggregated results, and advertisements. However, these signals can only be leveraged for SERPs where cursor tracking was enabled, limiting their utility for informing the design of new SERPs. In this work, we develop robust, log-based mouse movement models capable of estimating searcher attention on novel SERP arrangements. These models can help improve SERP design by anticipating searchers' engagement patterns given a proposed arrangement. We demonstrate the efficacy of our method using a large set of mouse-tracking data collected from two independent commercial search engines.

Modern search engines make extensive use of people's contextual information to finesse result rankings. Using a searcher's location provides an especially strong signal for adjusting results for certain classes of queries where people may have clear preference for local results, without explicitly specifying the location in the query direct-ly. However, if the location estimate is inaccurate or searchers want to obtain many results from a particular location, they have limited control on the location focus in the search results returned. In this paper we describe a user study that examines the effect of offering searchers more control over how local preferences are gathered and used. We studied providing users with functionality to offer explicit relevance feedback (ERF) adjacent to results automatically identi-fied as location-dependent (i.e., more from this location). They can use this functionality to indicate whether they are interested in a particular search result and desire more results from that result's location. We compared the ERF system against a baseline (NoERF) that used the same underlying mechanisms to retrieve and rank results, but did not offer ERF support. User performance was as-sessed across 12 experimental participants over 12 location-sensitive topics, in a fully counter-balanced design. We found that participants interacted with ERF frequently, and there were signs that ERF has the potential to improve success rates and lead to more efficient searching for location-sensitive search tasks than NoERF.

Past studies of user behavior in Web search have correlated eye-gaze and mouse cursor positions, and other lines of research have found cursor interactions to be useful in determining user intent and relevant parts of Web pages. However, cursor interactions are not all the same; different types of cursor behavior patterns exist, such as reading, hesitating, scrolling and clicking, each of which has a different meaning. We conduct a search study with 36 subjects and 32 search tasks to determine when gaze and cursor are aligned, and thus when the cursor position is a good proxy for gaze position. We study the effect of time, behavior patterns, user, and search task on the gaze-cursor alignment, findings which lead us to question the maxim that "gaze is well approximated by cursor." These lessons inform an experiment in which we predict the gaze position with better accuracy than simply using the cursor position, improving the state-of-the-art technique for approximating visual attention with the cursor. Our new technique can help make better use of large-scale cursor data in identifying how users examine Web search pages.

When searching for destinations and activities, the interests and knowledge of locals and non-locals may vary. In this paper, we compare and contrast the search-related interests of these two groups, and when they share a common interest (in our case, for restaurants), we analyze the quality of the venues they intend to visit. We find differences in interests depending on local knowledge, and that locals generally select higher-quality venues than non-locals. These findings have implications for search and recommendation systems that can personalize results based on local knowledge and leverage that knowledge to benefit non-locals.

Users of search engines often abandon their searches. Despite the high frequency of Web search abandonment and its importance to Web search engines, little is known about why searchers abandon beyond that it can be for good or bad reasons. In this paper, we ex-tend previous work by studying search abandonment using both a retrospective survey and an in-situ method that captures aban-donment rationales at abandonment time. We show that although satisfaction is a common motivator for abandonment, one-in-five abandonment instances does not relate to satisfaction. We also studied the automatic prediction of the underlying reason for ob-served abandonment. We used features of the query and the results, interaction with the result page (e.g., cursor movements, scrolling, clicks), and the full search session. We show that our classifiers can learn to accurately predict the reasons for observed search abandonment. Such accurate predictions help search providers estimate user satisfaction for queries without clicks, affording a more complete understanding of search engine performance.

While lots of reading happens on mobile devices, little research has been performed on how the reading-interaction actually takes place. Therefore we describe our findings on a study conducted with 18 users which were asked to read a number of texts while their touch and gaze data was being recorded. We found three reader types and identified their preferred alignment of text on the screen. Based on our findings we are able to computationally estimate the reading area with an approximate .81 precision and .89 recall. Our computed reading speed estimate has an average 10.9% wpm error in contrast to the measured speed, and combining both techniques we can pinpoint the reading location at a given time with an overall word error of 9.26 words, or about three lines of text on our device.

An increasing amount of text is being read digitally. In this paper we explore how eye tracking devices can be used to aggregate reading data of many readers in order to provide authors and editors with objective and implicitly gathered quality feedback. We present a robust way to jointly evaluate the gaze data of multiple readers, with respect to various reading-related features. We conducted an experiment in which a group of high school students composed essays subsequently read and rated by a group of seven other students. Analyzing the recorded data, we find that the amount of regression targets, the reading-to-skimming ratio, reading speed and reading count are the most discriminative features to distinguish very comprehensible from barely comprehensible text passages. By employing machine learning techniques, we are able to classify the comprehensibility of text automatically with an overall accuracy of 62%.

Distinguishing whether eye tracking data reflects reading or skimming already proved to be of high analytical value. But with a potentially more widespread usage of eye tracking systems at home, in the office or on the road the amount of environmental and experimental control tends to decrease. This in turn leads to an increase in eye tracking noise and inaccuracies which are difficult to address with current reading detection algorithms. In this paper we propose a method for constructing and training a classifier that is able to robustly distinguish reading from skimming patterns. It operates in real time, considering a window of saccades and computing features such as the average forward speed and angularity. The algorithm inherently deals with distorted eye tracking data and provides a robust, linear classification into the two classes read and skimmed. It facilitates reaction times of 750ms on average, is adjustable in its horizontal sensitivity and provides confidence values for its classification results; it is also straightforward to implement. Trained on a set of six users and evaluated on an independent test set of six different users it achieved a 86% classification accuracy and it outperformed two other methods.

Web search components such as ranking and query suggestions analyze the user data provided in query and click logs. While this data is easy to collect and provides information about user behavior, it omits user interactions with the search engine that do not hit the server; these logs omit search data such as users' cursor movements. Just as clicks provide signals for relevance in search results, cursor hovering and scrolling can be additional implicit signals. In this work, we demonstrate a technique to extend models of the user's search result examination state to infer document relevance. We start by exploring recorded user interactions with the search results, both qualitatively and quantitatively. We find that cursor hovering and scrolling are signals telling us which search results were examined, and we use these interactions to reveal latent variables in searcher models to more accurately compute document attractiveness and satisfaction. Accuracy is evaluated by computing how well our model using these parameters can predict future clicks for a particular query. We are able to improve the click predictions compared to a basic searcher model for higher ranked search results using the additional log data.

Users' search activity has been used as implicit feedback to model search interests and improve the performance of search systems. In search engines, this behavior usually takes the form of queries and result clicks. However, richer data on how people engage with search results can now be captured at scale, creating new opportu-nities to enhance search. In this poster we focus on one type of newly-observable behavior: text selection events on search-result captions. We show that we can use text selections as implicit feedback to significantly improve search result relevance.

Understanding the impact of individual and task differences on search result page examination strategies is important in developing improved search engines. Characterizing these effects using query and click data alone is common but insufficient since they provide an incomplete picture of result examination behavior. Cursor- or gaze-tracking studies reveal richer interaction patterns but are often done in small-scale laboratory settings. In this paper we leverage large-scale rich behavioral log data in a naturalistic setting. We examine queries, clicks, cursor movements, scrolling, and text highlighting for millions of queries on the Bing commercial search engine to better understand the impact of user, task, and user-task interactions on user behavior on search result pages (SERPs). By clustering users based on cursor features, we identify individual, task, and user-task differences in how users examine results which are similar to those observed in small-scale studies. Our findings have implications for developing search support for behaviorally-similar searcher cohorts, modeling search behavior, and designing search systems that leverage implicit feedback.

Reading is one of the most frequent activities of knowledge workers. Eye tracking can provide information on what document parts users read, and how they were read. This article aims at generating implicit relevance feedback from eye movements that can be used for information retrieval personalization and further applications. We report the findings from two studies which examine the relation between several eye movement measures and user-perceived relevance of read text passages. The results show that the measures are generally noisy, but after personalizing them we find clear relations between the measures and relevance. In addition, the second study demonstrates the effect of using reading behavior as implicit relevance feedback for personalizing search. The results indicate that gaze-based feedback is very useful and can greatly improve the quality of Web search. The article concludes with an outlook introducing attentive documents keeping track of how users consume them. Based on eye movement feedback, we describe a number of possible applications to make working with documents more effective.

We investigate how people interact with Web search engine result pages using eye-tracking. While previous research has focused on the visual attention devoted to the 10 organic search results, this paper examines other components of contemporary search engines, such as ads and related searches. We systematically varied the type of task (informational or navigational), the quality of the ads (relevant or irrelevant to the query), and the sequence in which ads of different quality were presented. We measured the effects of these variables on the distribution of visual attention and on task performance. Our results show significant effects of each variable. The amount of visual attention that people devote to organic results depends on both task type and ad quality. The amount of visual attention that people devote to ads depends on their quality, but not the type of task. Interestingly, the sequence and predictability of ad quality is also an important factor in determining how much people attend to ads. When the quality of ads varied randomly from task to task, people paid little attention to the ads, even when they were good. These results further our understanding of how attention devoted to search results is influenced by other page elements, and how previous search experiences influence how people attend to the current page.

An understanding of how people allocate their visual attention when viewing Web pages is very important for Web authors, interface designers, advertisers and others. Such knowledge opens the door to a variety of innovations, ranging from improved Web page design to the creation of compact, yet recognizable, visual representations of long pages. We present an eye-tracking study in which 20 users viewed 361 Web pages while engaged in information foraging and page recognition tasks. From this data, we describe general location-based characteristics of visual attention for Web pages dependent on different tasks and demographics, and generate a model for predicting the visual attention that individual page elements may receive. Finally, we introduce the concept of fixation impact, a new method for mapping gaze data to visual scenes that is motivated by findings in vision research.

We examine two basic sources for implicit relevance feedback on the segment level for search personalization: eye tracking and display time. A controlled study has been conducted where 32 participants had to view documents in front of an eye tracker, query a search engine, and give explicit relevance ratings for the results. We examined the performance of the basic implicit feedback methods with respect to improved ranking and compared their performance to a pseudo relevance feedback baseline on the segment level and the original ranking of a Web search engine.

We describe an approach for creating precise personalized document classifiers based on the user's attention. The general idea is to observe which parts of a document the user was interested in just before he or she comes to a classification decision. Having information about this manual classification decision and the document parts the decision was based on, we can learn precise classifiers. For observing the user's focus point of attention we use an unobtrusive eye tracking device and apply an algorithm for reading behavior detection. On this basis, we can extract terms characterizing the text parts interesting to the user and employ them for describing the class the document was assigned to by the user. Having learned classifiers in that way, new documents can be classified automatically using techniques of passage-based retrieval. We prove the very strong improvement of incorporating the user's visual attention by a case study that evaluates an attention-based term extraction method.

This paper presents the knowledge-system SONOCONSULT – an intelligent system in the medical domain. We evaluated the accuracy, acceptance and impact of SONOCONSULT, which has been used in clinical routine since 2002. The system was well accepted and had a significant clinical impact. In contrast to our original expectations, the diagnostic conclusions, although inferred with high accuracy, were less important especially for experienced physicians.

We examine the effect of incorporating gaze-based attention feedback from the user on personalizing the search process. Employing eye tracking data, we keep track of document parts the user read in some way. We use this information on the subdocument level as implicit feedback for query expansion and reranking.

In the proposed PhD thesis, it will be examined how attention data from the user, especially generated by an eye tracker, can be exploited in order to enhance and personalize information retrieval methods.

User-centered consultation systems can be viewed as systems with trust in the competence of the user, whereas conventional knowledge systems are often based on some mistrust. By offering a variety of formal and informal diagnostic services mediated by flexible agents, the user solves his or her problem by asking the agents for specific information complementary to his or her knowledge. The development process is flexible enough to build minimalist Wiki-like informal systems very quickly, which can be incrementally refined and formalized in a distributed manner. The approach is demonstrated with two case studies.