Data-driven approaches to content selection for data-to-text generation

Abstract

Data-to-text systems are powerful in generating reports from data automatically and thus they simplify the presentation of complex data. Rather than presenting data using visualisation techniques, data-to-text systems use human language, which is the most common way for human-human communication. In addition, data-to-text systems can adapt their output content to users’ preferences, background or interests and therefore they can be pleasant for users to interact with. Content selection is an important part of every data-to-text system, because it is the module that decides which from the available information should be conveyed to the user. This thesis makes three important contributions. Firstly, it investigates data-driven approaches to content selection with respect to users’ preferences. It develops, compares and evaluates two novel content selection methods. The first method treats content selection as a Markov Decision Process (MDP), where the content selection decisions are made sequentially, i.e. given the already chosen content, decide what to talk about next. The MDP is solved using Reinforcement Learning (RL) and is optimised with respect to a cumulative reward function. The second approach considers all content selection decisions simultaneously by taking into account data relationships and treats content selection as a multi-label classification task. The evaluation shows that the users significantly prefer the output produced by the RL framework, whereas the multi-label classification approach scores significantly higher than the RL method in automatic metrics. The results also show that the end users’ preferences should be taken into account when developing Natural Language Generation (NLG) systems. NLG systems are developed with the assistance of domain experts, however the end users are normally non-experts. Consider for instance a student feedback generation system, where the system imitates the teachers. The system will produce feedback based on the lecturers’ rather than the students’ preferences although students are the end users. Therefore, the second contribution of this thesis is an approach that adapts the content to “speakers” and “hearers” simultaneously. It considers initially two types of known stakeholders; lecturers and students. It develops a novel approach that analyses the preferences of the two groups using Principal Component Regression and uses the derived knowledge to hand-craft a reward function that is then optimised using RL. The results show that the end users prefer the output generated by this system, rather than the output that is generated by a system that mimics the experts. Therefore, it is possible to model the middle ground of the preferences of different known stakeholders. In most real world applications however, first-time users are generally unknown, which is a common problem for NLG and interactive systems: the system cannot adapt to user preferences without prior knowledge. This thesis contributes a novel framework for addressing unknown stakeholders such as first time users, using Multi-objective Optimisation to minimise regret for multiple possible user types. In this framework, the content preferences of potential users are modelled as objective functions, which are simultaneously optimised using Multi-objective Optimisation. This approach outperforms two meaningful baselines and minimises regret for unknown users.