I am working on an automated insights generation use case where I want to generate meaningful sentences from given aggregated data.

For example,

Student = John
Total_Marks = 96
Class_Average = 85

NLG model-generated insights:

1. You did an excellent job, John! Your score is 96!  
2. You have scored 11 marks above the class average.

When I look at classic NLG, they are sentence generation approaches given a starting letter or word. This might be more of a Neural Machine Transition use case.

What do you think my approach should be?

  • $\begingroup$ Can you add details about the domain and scope of kinds of data that you want to generate sentences from e.g. is it all summaries of numerical data, is it all related to school reports? Also, how variable you need the output to be. These are important details, because a simple randomly-chosen template with placeholders would be suitable NLG for a restricted domain with limited variability, and has benefits of not making grammatical errors etc. $\endgroup$ May 21 '19 at 6:58