Broader Impact. As information retrieval systems (search engines, recommender systems, conversational agents) touch on every aspect of our life, the technology that we help develop should be informed by, and inform, the world around us. This starts with understanding and replying to immediate stakeholders — users, decision makers, and engineers — across applications such as web search, information discovery in social networks, HR analytics, medical search, and e-commerce. But there are broader concerns. Studies suggest that online platforms have impacted soci- ety by leading to increasing polarization; changing the metrics can help to begin to assess and understand such issues. The IR field needs to recruit more diverse people, and not just as collabo- rators, but also students. These changes might help IR researchers better understand the methods (and communicate about them), which in turn could lead to additional insights and theoretical developments. Finally, a stronger emphasis on transparency will likely force us to document our methods and experiments in a better and more systematic ways. This, in turn, will positively impact teaching and reproducibility in IR.
Broader Impact. The proposed research will make users better informed and more aware of information quality and its broader context, by providing a broader, more balanced view of the information space and meta-cognition for the further process of information seeking or decision making. It also connects a user’s information seeking behavior to a growing understanding of the cognitive processes that underlie a person’s searching, learning, and decision-making. This should lead to users being more confident and efficient in their learning and decision-making, to improvement in the overarching task of connecting people with the right information, to support for complex matching tasks such as expert-finding or peer matching, and to enabling people to learn more and to learn more effectively.
Broader Impact. Online evaluation is not just the domain of a few global search brands, it is an industry. For exam- ple, the online evaluation/optimization company Optimizely has gone from 0 to 500+ employees in about six years. Such companies enable smaller companies to perform online evaluation and test changes. Work in online evaluation of search will have a substantial impact on search as well as related topics, such as recommender systems.
Broader Impact. A better understanding and a more insightful explanation of IR system performance opens up new possibilities in terms of reproducibility, external validity, and generalizability of experimental results since it provides the means to understand what succeeded or failed, especially if linked to failure analysis. Better analytic tools are also an indispensable basis for moving IR toward becoming a predictive science.
Broader Impact. This research presents an incredible opportunity to broaden the community, because it will open a wide range of research questions, which have been largely ignored, yet are of central concern to the evaluation of, for instance, support for complete search sessions, or of personalization of search.
Broader Impact. The development of machine learning and AI “algorithms” for new products and services, and the success of deep learning methods in high visibility competitions, have received enormous attention in the media. New courses on machine learning and neural models are proliferating and many new papers related to these topics are appearing every day on arXiv. Given that search is nearly ubiquitous and that search engines are the best example of IR and NLP in action, it is not surprising that there has been a significant upsurge in interest in applying neural models to many aspects of IR and search, both by graduate students and by companies working on a broad range of applications. Although this research is still at an early stage and much remains to be done to demonstrate the levels of improvement in effectiveness seen in other fields, the potential impact of new retrieval approaches developed using learning techniques is enormous. If we are successful, for example, we will make significant progress towards achieving one of the long-standing goals of IR - an “intelligent” conversational search assistant similar to those seen in the early Star Trek episodes that inspired so many future computer scientists. On a more mundane level, we should also be able to develop search techniques that are substantially more portable and effective across tasks, domains, applications, and languages. Search will be substantially more effective and available to all segments of society.
Broader Impact. Several sub-communities in information retrieval have covered various aspects of the GIO frame- work. We think that this conceptual framework can facilitate a spread of best practices across the field. New indexing approaches and search models will likely arise from applications that use generated information objects to the fullest. As the GIO framework requires us to rethink many moving parts of typical IR systems, it bears the potential for more collaboration across typical IR subfields. Additionally, by opening up the definition of IR to explicitly include summarization, natural language generation, and computer vision, it enhances the ties between the IR community and communities on information extraction, knowledge graph construction, text and vision, spoken dialogs, and databases.
Broader Impact. The goal of efficiency is to make it possible to process (in the general sense) more data with fewer resources. For search this has the effect of reducing the overall quantity of hardware (and associated infrastructure). In a data center with 100,000 machines, a reduction in execution time by 5% is the equivalent of 5,000 machines, together with the power, cooling, and maintenance of those machines. This is a green computing contribution (aka green IR). The released resources could be used to absorb more user growth without requiring additional resources, could be directed towards other workloads, or could be simply turned off until needed. A secondary effect is an increase in the amount of data that can be processed on a single resource, lowering the barrier to entry for researchers in the broader field of IR, and for companies entering the marketplace. A cluster of machines would not be needed if we make the work efficient enough such that it can be carried out on a laptop. In the longer term, the quest for efficient IR has the potential to impact the design of CPUs and other components in the computing system. CPUs with vector processors (AVX, etc) are already in desktops – and the instructions are already being used in IR applications. FPGAs are now being integrated into CPUs, and we know that Microsoft is already using FPGAs for search. Future directions include GPGPU instruction integration into the main core. Decreased costs might be achieved by offloading some of the work to the user’s device (edge computing). For example, the last phase of reranking using an already learned (or adaptive) ranking function might even be performed on the user’s mobile phone. Such a change might be disruptive to the standard model of internet or other search, and might also enable new privacy- preserving mechanisms.
Broader Impact. The most obvious, immediate, impact is the reduction in the cost of (re-)finding and (re-)using one’s own data which exists in many information silos (e.g., is the SWIRL document on my laptop, on a shared drive (which one?), or in my email (which account?)). Improved access to personal information will also reduce the frustration in finding and thus reduce the friction in working with personal data. This impacts everyone with a connected device - virtually every human on the planet. Improvements in this area will also translate to finding in the enterprise, where information workers spend large amounts of their time (re-)finding existing organizational knowledge within their own corporate repository and across all the repositories within the organization.
