Counterfactual Learning for Controlled Data-to-text – ACDC

The project ACDC builds on recent advances in neural langage generation to handle data-to-text objectives. We focus on the textual synthesis of information given as tabular data. One main focus of the project will be given on the search for invariance in the input data, the extraction of high-level compression operators and the customization of the produced outputs. We propose to rely on deep and reinforcement learning techniques, involving inference, manipulation and decoding of representations of synthsis operators in a contunous space. The goal is to produce smooth representation spaces, encoding various kinds of semantic symetries of operators, enabling the control of the way contents are summarized according to the input data tables. The explicit inference of operators that we envision in this project will allow to build interpretable models, easying then the analysis of the produced synthesis, and the planing of consistently organized textual reports, detailing various salients aspects of the data.