Combining intent modeling and visual user interfaces can help users discover novel information and dramatically improve their information-exploration performance.
Current-generation search engines serve billions of requests each day, returning responses to search queries in fractions of a second. They are great tools for checking facts and looking up information for which users can easily create queries (such as "Find the closest restaurants" or "Find reviews of a book"). What search engines are not good at is supporting complex information-exploration and discovery tasks that go beyond simple keyword queries. In information exploration and discovery, often called "exploratory search," users may have difficulty expressing their information needs, and new search intents may emerge and be discovered only as they learn by reflecting on the acquired information.8,9,18 This finding roots back to the "vocabulary mismatch problem"13 that was identified in the 1980s but has remained difficult to tackle in operational information retrieval (IR) systems (see the sidebar "Background"). In essence, the problem refers to human communication behavior in which the humans writing the documents to be retrieved and the humans searching for them are likely to use very different vocabularies to encode and decode their intended meaning.8,21
Assisting users in the search process is increasingly important, as everyday search behavior ranges from simple look-ups to a spectrum of search tasks23 in which search behavior is more exploratory and information needs and search intents uncertain and evolving over time.
We introduce interactive intent modeling, an approach promoting resourceful interaction between humans and IR systems to enable information discovery that goes beyond search. It addresses the vocabulary mismatch problem by giving users potential intents to explore, visualizing them as directions in the information space around the user's present position, and allowing interaction to improve estimates of the user's search intents.
Interactive intent modeling is based on two scientific principles (see Figure 1):
Visualization. Visualizing the current search intent and directions in the information space; and
Adaptation. Interactive adaptation of the intent model, balancing exploration of the information space and exploitation of user feedback; the intent model must be able to rigorously handle uncertainty due to limited, possibly suboptimal, user feedback.
By visualizing query and data elements (such as keywords), this approach enables the system to show its understanding of user search intent to the user and also provide a view of available search directions around the user's current position in the information space. The initial evidence concerning user search intent is often limited. The intent model is thus manageable for the user only if the system is able to predict a sufficient subset of the potentially relevant intents. Given the visualization of the intent model and its relation to the information space, the user is able to provide feedback for the intent model, allowing the system to improve intent estimates on subsequent iterations, retrieve and rank data, and update the visualization of directions in the information space.
The SciNet system for scientific literature search (http://augmentedresearch.hiit.fi/) is an example of the two principles in interactive intent modeling (see Figure 2).20 The system currently indexes more than 50 million scientific articles and is designed to assist users exploring information related to a particular research topic through rapid feedback loops and in making sense of the available information around the initial query context.14,20
In the Figure 2 scenario, a user is trying to learn about "3D gestures" and types in the corresponding query. The user is visualized with an estimate of his or her present search intents, as well as potential intents, and directions in the information space on a radar screen. The user then navigates by directly manipulating the estimated intents on the display.
Figure 2a is the system's response to the initial query on 3D gestures, offering directions to, say, "video games," "user interfaces," "gesture recognition," and "virtual reality." In Figure 2b, the user has selected "gesture recognition" and is offered further options to continue the exploration to more specific topics (such as "nearest neighbor approach" and "hidden Markov models") but also to more general topics (such as "pointing gestures" and "spatial interaction") estimated to be relevant for the user's interaction history. The modeling is based on a fast online regression model that estimates task-level search intents.20 The model estimates relevance related to potential search intent and uncertainty related to these estimates based on user feedback. Search intents are visualized as keywords, and selection of which intents to visualize is determined through the exploration-exploitation paradigm. The trick is to present to the user not only choices estimated to be most relevant but those for which the upper confidence bound is the greatest. The user decides whether to explore or exploit, as both relevant and uncertain keywords are visualized; for example, if the user first selects "gesture recognition" and then "hidden Markov models," the system would then suggest specific hidden Markov model applications in gesture recognition, which would be exploitative, as they are estimated to be most relevant and also allow the user to continue to explore more uncertain directions (such as other computational techniques in gesture recognition).
While the idea of providing the user uncertain interaction options may be counterintuitive from a conventional IR perspective, which is based on the principle of maximizing relevance, our interactive intent modeling helps users overcome the vocabulary-mismatch problem, as the system provides them interaction resources to continuously direct the search and actively explore relevant, though not initially obvious, information. Experiments show user-task-level performance can be improved significantly.
To deliver such support, the interface provides a nonintrusive relevance-feedback mechanism in which the user pulls keywords closer to the center of the radar screen to increase their importance and pushes keywords away from the center of the radar screen to decrease their importance. The keywords can be enlarged with a fisheye lens that follows the mouse cursor anywhere on the radar screen. In response, the system updates intent visualization and search results. The radar screen's radial layout represents good balance between the amount of information shown and comprehensibility compared to alternative visualizations with lower or higher degrees of freedom that could make interaction with the visualization more difficult.
The SciNet example demonstrates how visualization can be used to elicit feedback. Feedback can be targeted directly to the intent model (the inner circle of the radar screen in Figure 2) or to possible future directions (the outer rim of the radar screen in Figure 2). Due to the vocabulary-mismatch problem, users often have trouble expressing their needs as written queries and are likely to start their search with imprecise queries. Hence, interaction and feedback mechanisms that engage users to provide feedback on how to direct their search in the subsequent iterations are crucial. This is grounded in a well-known cognitive-science theory stating users find recognition easier than recall.3 It is usually easier for humans to recognize something they see than describe it from scratch.
However, increasing evidence from IR research supports the finding that while relevance feedback is useful in enabling systems to better serve user search intent, in most cases users do not in fact use feedback mechanisms.18 This observation is related to two other cognitive science findings: users find it easier to recognize smaller units than more complex ones, and it is easier for them to make sense of information relative to a reference point than in isolation.6 Assessing the relevance of a full document may be an even more demanding task than formulating a new query.
Information visualization can turn laborious relevance assessment into a more fluent recognition task; for example, visualizing essential document content can be faster for recognizing important directions toward finding relevant information than forcing users to read this information from an original document.1,17
Recent visualizations applied in search tasks support sense making of bibliographic data by, for example, incrementally interactively exploring networks of data.10 While these systems show the importance of visualization for sense making, they are limited by not allowing users to negotiate the intent model with the system, allowing them only to explore information through direct links already present in the network data.1
By visualizing potential intents, an IR system can give users a spectrum of choices in a form suitable for the human visual system to process rapidly, even when the data changes dynamically as interaction occurs. Such an interface requires advanced data-driven visualizations that can be computed online. Moreover, a visualization should not contain only information already familiar to the user that would be good for recognition but lead to the intent model getting stuck in a "context bubble." Instead, unseen parts of the information space must be offered to the user, facilitating sense making through the relation of these parts to already familiar information when possible.
Given the evolutionary nature of search, as demonstrated in our example search scenario involving SciNet, it is important to not only exploit the feedback elicited from the user but balance it with exploration. Users must be able to focus on a specific location in the information space (exploit) and be able to broaden their search through more general areas (explore).
This insight is particularly important for users exploring information with which they are not familiar. Users often suffer from what psychologists call "anchoring," or the tendency to make insufficient adjustments to initial values when judging under conditions of uncertainty.22 Users may thus tend to refrain from abandoning their initial expression of their information needs or from adjusting them very much, causing subsequent expression of information needs to be biased toward their current knowledge. This bias reduces the likelihood they will discover something novel.
This behavioral finding has consequences for machine-learning approaches to modeling search intent. A promising direction for predicting intent while still allowing users to be in control of the search process comes from machine-learning methods that learn online. Online learning methods are able to update models one observation at a time so future predictions can be made immediately when feedback is received. The goal in online learning for search is to predict the relevance of content interactively, meaning that soon after prediction is made, the judgment of its usefulness is received from user feedback. This information can then be used by the IR system to refine the prediction hypothesis used by the method.
Online learning methods are able to update models one observation at a time so future predictions can be made immediately when feedback is received.
Standard machine learning for online prediction does not solve the problem of discovering what interaction options are most useful in allowing the learning method to improve its estimates and therefore create visualizations for the user. Straightforward exploitation by choosing the directions currently estimated to be most relevant could lead to converging to suboptimal goals and locking users in context bubbles predetermined by the user's limited prior knowledge; for example, a user searching for "3D gestures" might never explore "pointing gestures," as the initial query scope could already be too specific to allow such exploration. A promising solution for collecting feedback that also allows exploration is the "exploration-exploitation" paradigm of reinforcement learning.5 In it, the model and its environment (the user) form an online loop, and learning involves finding a balance between exploration (showing items from uncharted information space for feedback) and exploitation (showing items most likely to be relevant, given the current user intent model).
Users can thus be assisted to direct their searches under conditions of uncertainty by learning intent models online based on feedback they give about the models. Due to the limited and imperfect feedback available, the amount of uncertainty about user intent can be substantial. It is therefore important for an IR system to use models capable of handling uncertainty in a rigorous manner. Probabilistic online learning models can be used in the exploitation part of the exploration-exploitation dilemma; moreover, as the models are probabilistic, they quantify uncertainty related to the estimates, using it to determine the optimal alternative choices to be visualized for users.5
When studying IR systems designed to negotiate user search intent, it is important to realize the utility these systems improve is not retrieval effectiveness at a query-response level but task-level performance. Within interactive IR systems, users are often required to do more work to complete their tasks, and, while some of it may be wasted, they may be more successful in correcting their initially suboptimal actions. To get a complete picture of performance, two aspects of IR systems should be measured simultaneously: IR system effectiveness, given a complete description of an information need; and human task performance, given the system's interaction modes.
The SciNet system was recently studied in task-based experiments in which users were given 30 minutes to solve research tasks using IR systems operating on a database of more than 50 million scholarly articles. We compared a system setup with interactive intent modeling against a conventional IR system with a list-based visualization and interaction with typed queries. We quantified the quality of retrieved information, adoption of the visualization, and feedback mechanisms separately.14,20 We found interactive search intent modeling significantly improved users task performance. We also found the task outcomes graded more highly by experts, and the search user interface enhanced interaction without compromising task execution time. We attributed the improved task performance to the improved quality of retrieved information and to the improved visualizations and interaction modes offered by the system. In particular, interactive intent modeling increased recall of novel information without losing precision. This performance demonstrates the power of the interactive intent modeling technique in supporting exploration and discovery of information that can be difficult to find with systems that rely on conventional search user interfaces (see the table here).
Engaging users to interact with IR systems is crucial for such systems to be able to offer better interaction modes and reduce uncertainty related to user expression of search intent. Despite significant improvement in user task performance on the example SciNet system discussed earlier, we are only scratching the surface of human-centered computing as part of the search activity. Intent-aware IR systems can benefit from ubiquitous computing in at least two ways, as discussed next.
IR systems can be extended by augmenting a real scene with predictions of what the user might find useful, shown as augmented reality on head-mounted displays (HMDs). Users' implicit and explicit reactions to visualized content can reveal their intent and help improve the user intent model contextualized to the immediate setting. Figure 3a shows how suitable information (such as topics, research group, and publications) the user can recognize and act upon can be visualized on a HMD superimposed on the real scene;2 for example, augmenting a user's environment when visiting a poster session at a conference with visual cues and information can help the system collect information about the user's intent even when the user is not actively engaged with a search engine.
Implicit feedback from physiological computing. Recent advances in wearable computing have facilitated capturing users' affective and cognitive states (such as wearable electro-encephalography, or EEG, systems). Moreover, other physiological sensors (such as galvanic skin response and heart-rate sensors) are being integrated into wrist-wearable products like smartwatches. Such physiological signals give researchers additional sources of feedback information not previously available.
It has been shown that affective state information can be used for relevance judgment prediction,4 and affective and psychophysiological signals are being employed in multimedia search systems with encouraging results.19 Figure 3b shows an example of an EEG sensor setup used to demonstrate term-relevance prediction from brain signals. The experiment shows it is possible to automatically detect the relevance of text information visualized for the user directly from brain signals by analyzing neural activity of participants while providing relevance judgments to text terms for a given topic.11 Employing such physiology-based relevance detection for implicit relevance feedback on visualized information can be used by IR system developers to improve the prediction of the intent model to complement or substitute explicit user relevance ratings.
Recent work demonstrates there is significant room for improving the support provided to users involved in exploratory forms of search. Overall, researchers recognize the need for search and information-exploration systems that combine the information-processing capabilities of humans and computers.
Interactive intent modeling is a theoretically motivated, empirically proven way to support information exploration and discovery. It can increase users' capacity for information processing and discovery through computing technologies that assist users navigating complex information spaces.
Interactive intent modeling provides additional resources for users to better learn about the information space and give increased feedback for the system so it can efficiently adapt its understanding of user-search intent.
Engaging users to adopt interactive feedback mechanisms for information exploration and sense making requires user-interface techniques that go beyond search boxes and lists of links to enable them to better interact with the system and have control over their findings. Modeling user intent online as interaction occurs and even in situations where user feedback is noisy and suboptimal requires machine-learning models that learn online and are able to explore, not just exploit. IR system design must ultimately integrate interactive visualizations, intent prediction, multimodal feedback, and a higher-level context of tasks and goals.
IR systems must be able to help users solve tasks, not just retrieve documents. Users need search engines and user interfaces that adapt to their capabilities and search behavior, rather than require them to adapt to them.
This work was partly supported by the Academy of Finland (278090, Multivire 255725 and the COIN Center of Excellence 251170), TEKES (D2I and Re:KnoW), and the European Commission through the FP7 Project MindSee 611570. Certain data included here is derived from the Web of Science prepared by Thomson Reuters, Inc., Philadelphia, PA. Data is also included from the Digital Libraries of the ACM, IEEE, and Springer.
2. Ajanki, A., Billinghurst, M., Gamper, H., Järvenpää, T., Kandemir, M., Kaski, S., Koskela, M., Kurimo, M., Laaksonen, J., Puolamäki, K., Ruokolainen, T., and Tossavainen, T. An augmented reality interface to contextual information. Virtual Reality 15, 2-3 (June 2011), 161–173.
4. Arapakis, I., Athanasakos, K., and Jose, J.M. A comparison of general vs. personalised affective models for the prediction of topical relevance. In Proceedings of the 33rd International ACM SIGIR Conference on Research and Development in Information Retrieval (Geneva, Switzerland, July 19–23). ACM Press, New York, 2010, 371–378.
9. Belkin, N.J., Oddy, R.N., and Brooks, H.M. Ask for information retrieval: Part I: Background and theory. In Readings in Information Retrieval. Morgan Kaufmann Publishers Inc., San Francisco, CA, 1997, 299–304.
10. Chau, D.H., Kittur, A., Hong, J.I., and Faloutsos, C. Apolo: Making sense of large network data by combining rich user interaction and machine learning. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (Vancouver, BC, Canada, May 7–12). ACM Press, New York, 2011, 167–176.
11. Eugster, M.J.A., Ruotsalo, T., Spapé, M.M., Kosunen, I., Barral, O., Ravaja, N., Jacucci, G., and Kaski, S. Predicting term-relevance from brain signals. In Proceedings of the 37th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (Gold Coast, Queensland, Australia, July 6–11). ACM Press, New York, 2014, 425–434.
14. Glowacka, D., Ruotsalo, T., Konuyshkova, K., Athukorala, K., Kaski, S., and Jacucci, G. Directing exploratory search: Reinforcement learning from user interactions with keywords. In Proceedings of the 2013 International Conference on Intelligent User Interfaces (Santa Monica, CA, Mar. 19–22). ACM, Press, New York, 2013, 117–128.
15. Hearst, M.A. Tilebars: Visualization of term distribution information in full text information access. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (Denver, CO, May 7–11). ACM Press/Addison-Wesley Publishing Co., New York, 1995, 59–66.
16. Hearst, M.A. and Pedersen, J.O. Reexamining the cluster hypothesis: Scatter/gather on retrieval results. In Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (Zurich, Switzerland, Aug. 18–22). ACM Press, New York, 1996, 76–84.
19. Moshfeghi, Y. and Jose, J.M. An effective implicit relevance feedback technique using affective, physiological and behavioural features. In Proceedings of the 36th International ACM Conference on Research and Development in Information Retrieval (Dublin, Ireland, July 28–Aug. 1). ACM Press, New York, 2013, 133–142.
20. Ruotsalo, T., Peltonen, J., Eugster, M., Glowacka, D., Konyushkova, K., Athukorala, K., Kosunen, I., Reijonen, A., Myllymäki, P., Jacucci, G., and Kaski, S. Directing exploratory search with interactive intent modeling. In Proceedings of the 22nd ACM International Conference on Information & Knowledge Management (San Francisco, CA, Oct. 27–Nov. 1). ACM Press, New York, 2013, 1759–1764.
21. Saracevic, T., Kantor, P., Chamis, A. Y., and Trivison, D. A study of information seeking and retrieving. I. Background and methodology. Journal of the American Society for Information Science 39, 3 (May 1988), 161–176.
23. Wilson, M.L., Kules, B., Schraefel, M.C., and Shneiderman, B. From keyword search to exploration: Designing future search interfaces for the Web. Foundations and Trends in Web Science 2, 1 (Jan. 2010), 1–97.
24. Yee, K.-P., Swearingen, K., Li, K., and Hearst, M. Faceted metadata for image search and browsing. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (Ft. Lauderdale, FL, Apr. 5–10). ACM Press, New York, 2003, 401–408.
25. Zhao, L. and Callan, J. Term necessity prediction. In Proceedings of the 19th ACM International Conference on Information and Knowledge Management (Toronto, Canada, Oct. 26–30). ACM Press, New York, 2010, 259–268.
Figure 1. Exploring information with interactive intent modeling is based on two principles: visualizing current search intent and direction; and balancing exploration and exploitation of user feedback. The user's cognitive effort is thus reduced, as it is easier to recognize items instead of having to remember them when reformulating queries.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2015 ACM, Inc.