Giuseppe Carenini

With the exponential growth of the Web and the rapid adoption of mobile devices, information technology is permeating through almost all sectors of our society. As a result, more and more people analyze data and make important decisions by interacting with computer systems.  Supporting these users is however a challenging task because it requires computer systems to model rather complex communicative and inferential processes. In my research, I study several aspects of this task inspired by the following four guiding principles:

• Users need efficient access to the huge amount of information that is currently expressed in natural language, both in self-contained documents and in human conversations -- I believe a key part of the solution is text summarization, which has become a major thrust of my current research program. More specifically, I have focused on summarizing corpora of evaluative documents, as well as on email thread summarization and its generalization to summarizing human conversations in various modalities (e.g., blogs and meetings).
• Sophisticated interactive systems, supporting analysis and decision-making, need to be able to interpret and generate language beyond single sentences.  On the one hand, most human knowledge is expressed in natural language and extracting useful information form such body of text requires the ability to uncover how clauses and sentences are related to each other in a document. On the other hand, as systems become more intelligent and autonomous, they need to be able to explain their conclusions to humans, which requires the ability to generate explanations in the form of coherent multi-sentential text. The process of interpreting multi-sentential text is called discourse parsing and it is currently one of my most successful lines of research. As for generation of coherent text, although did not work on it for quite some time, I am now re-exploring this area with more modern concepts and techniques.
• To deal with highly heterogeneous user groups, interactive computer-systems should be user-adaptive. They should be able to model properties of their users (user modeling), and personalize accordingly different aspects of the interaction (e.g., content presented). User-adaptive interaction has been shown to be beneficial for a variety of tasks and applications; for instance, in my own work several years ago on generating user tailored text.  More recently, extending these benefits to provide personalization in Infovis has become a very successful pillar of my current research program. 
• Infovis is a critical component of effective data analysis and decision-making, even more so when it is synergistically combined with natural language. One of my key research goal in this space has been to study the integration of  NLP and InfoVis techniques to support users in exploring conversational data. This involved developing new visual encodings, new topic modeling techniques as well as new methods for interactive topic modeling. With respect to decision-making, I have been working for more than a decade on a visual analytic framework to support preferential choice, namely, the fundamental decision process of selecting the best option out of a set of alternatives.