Learning-to-rank with Prior Knowledge as Global Constraints

A good ranking function is the core of any Information Retrieval system. The ranking function can be a simple cosine similarity or, more likely for any advanced IR system like a Web search engine, a function processing a high number of signals which typically requires a lot of hand-tuning to deal with the large variability of queries. The result is a sub-optimal and highly complex ranking function which, in spite of having been crafted by human experts, is hard to control and debug. In the last few years, learning-to-rank from examples has emerged as a more flexible approach to design ranking functions. While learning-to-rank approaches have been proved to significantly outperform hand-tuned solutions, they still feature many disadvantages. First, they rely on a large number of training examples to model the high variability of the input query stream. Unfortunately, constructing a training set is more complicated than labeling examples in classical supervised classification tasks. Indeed, the labeling process for learning-to-rank tasks is inherently error-prone and incomplete. Secondly, learning-to-rank schemas usually do not account for the explicit knowledge that human experts have built over the years. It would be nice to integrate this knowledge without having to rely on a large set of examples to infer it. Finally, the proposed approach opens new ways to integrate unlabeled data into the learning process, as the rules must be respected also by the unlabeled data. This paper presents a general framework to convert prior knowledge in form of First Order Logic (FOL) clauses into a set of continuous constrains and shows how these constraints can be integrated into any learning-to-rank approach which is optimized via gradient descent.