Avishek Anand | Explainable Information Retrieval

9 JUNE 2022

Information prioritization is a crucial objective of information retrieval (IR) systems to tackle problems of information overload. An example of information prioritization is document ranking, a fundamental IR task that searches large text collections given an underspecified query. Document ranking not only taps into the world knowledge to satisfy user information needs but also supports Web tasks that we call knowledge-intensive tasks. It is perhaps unsurprising that most search and knowledge-intensive tasks are now solved using data-driven over parameterized learning models. On the one hand, these models have furthered the state-of-the-art on many IR tasks. But on the other hand, such learning systems tend to be opaque and non-interpretable, creating apprehensions of models being "right but for the wrong reasons". In this talk, I will first outline the landscape of interpretability of machine learning models in IR. Following this, I will talk about some challenges in explaining machine learning models in IR and some of our initial results. Finally, I will outline our vision toward a Neural and interpretable search engine that we are building from the results of our research.