Data-to-text Generation with Macro Planning Ratish Puduppully and Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB
[email protected] [email protected] Abstract content to talk about and how it might be orga- nized in discourse), sentence planning (aggregat- Recent approaches to data-to-text generation ing content into sentences, deciding specific words have adopted the very successful encoder- to describe concepts and relations, and generat- decoder architecture or variants thereof. ing referring expressions), and linguistic realisa- These models generate text which is flu- tion (applying the rules of syntax, morphology ent (but often imprecise) and perform quite and orthographic processing to generate surface poorly at selecting appropriate content and ordering it coherently. To overcome some forms). Recent neural network-based approaches of these issues, we propose a neural model (Lebret et al., 2016; Mei et al., 2016; Wiseman with a macro planning stage followed by a et al., 2017) make use of the encoder-decoder ar- generation stage reminiscent of traditional chitecture (Sutskever et al., 2014), are trained end- methods which embrace separate modules to-end, and have no special-purpose modules for for planning and surface realization. Macro how to best generate a text, aside from generic plans represent high level organization of im- mechanisms such as attention and copy (Bahdanau portant content such as entities, events and et al., 2015; Gu et al., 2016). The popularity of their interactions; they are learnt from data and given as input to the generator.