Finding middle ground? Multi-objective Natural Language Generation from time-series data

Abstract

A Natural Language Generation (NLG) system is able to generate text from nonlinguistic data, ideally personalising the content to a user's specific needs. In some cases, however, there are multiple stakeholders with their own individual goals, needs and preferences. In this paper, we explore the feasibility of combining the preferences of two different user groups, lecturers and students, when generating summaries in the context of student feedback generation. The preferences of each user group are modelled as a multivariate optimisation function, therefore the task of generation is seen as a multi-objective (MO) optimisation task, where the two functions are combined into one. This initial study shows that treating the preferences of each user group equally smooths the weights of the MO function, in a way that preferred content of the user groups is not presented in the generated summary.

How to talk to strangers: Generating medical reports for first-time users

“…However, our model can also account for content selection and information ordering simultaneously, as the fitness function is based on probabilities of content to be chosen, given previously selected content. In this regard, our approach is an improvement of a previously used approach [5], which only considers content selection.…”

“…The multi-objective optimisation approach by Gkatzia et al [5] is similar to our proposed method in that we frame content selection as an optimisation task with two objective functions. The two methods differ in two ways.…”

“…b) Evolutionary Algorithms and Multi-objective Optimisation for NLG: As discussed above, Gkatzia et al [5] suggest a multi-objective approach to Natural Language Generation using Reinforcement Learning, where the reward function was the aggregated sum of two objective functions. However, the standard way of solving a multi-objective task is through generating a set of optimal solutions using genetic algorithms [8].…”

“…In this work, we concentrate on content selection. There are several approaches to content selection which have previously studied including supervised learning [14], [16], unsupervised learning [12], Reinforcement Learning [21], multi-objective optimisation [5], Gricean Maxims [22], Integer Linear Programming [23], interest scores assigned to content [24], statistical approaches [25], a combination of statistical and template-based approaches [26] and many more. None of the aforementioned approaches uses multi-objective optimisation, with the exception of Gkatzia et al [5].…”

“…Previous studies have demonstrated that text descriptions can be more effective, understandable and helpful in decision making than graphical representations of data [2] [3]. In addition, previous work has shown that it is important for these types of NLG systems to adapt their output to specific users or user groups, such as nurses and patients [4], or lecturers and students [5]. Handling first-time users becomes problematic when no prior information about a user exists.…”