Invited Speakers

Data-to-text Natural Language Generation: Traditional, Novel and Future Methods

Data-to-text Natural Language Generation (NLG) is a consolidated field of research which normally combines Computational Linguistics, Software Engineering and Artificial Intelligence methods to generate natural language from non-linguistic representations. Traditionally, most data-to-text applications have been designed using a modular pipeline architecture, in which the non-linguistic input data is converted into natural language through several intermediate transformations. In contrast, influenced by the phenomenon of deep learning, recent neural models for data-to-text generation have been proposed as end-to-end approaches, where the non-linguistic input is rendered in natural language with much less explicit intermediate representations in-between. Theoretically, we know that pipeline approaches are more transparent and their modules can be reused across applications, whereas neural end-to-end approaches may demand less manual labor and has registered state-of-the-art results in other text generation tasks, such as Machine Translation. Although we know these pros and cons in theory, the question about which kind of model empirically generates the most fluent and semantic texts from non-linguistic representations still remains unanswered. This lack of an empirical comparison is partially caused by the fact that traditional benchmarks for the task only consist of raw non-linguistic representations in parallel with their textual realizations, benefiting the evaluation of end-to-end approaches but not of pipeline ones, since explicit intermediate representations are missing for the study of particular modules of the latter architecture. In this presentation, I will introduce an annotation framework to enrich popular benchmarks with explicit intermediate representations, which will help the development and evaluation of particular pipeline modules as Discourse Ordering, Aggregation, Lexicalization, Referring Expressions Generation and Textual Realization. Based on a version of a popular data-to-text benchmark enriched with our framework, I will also present the results of a comparison between pipeline and end-to-end approaches. Finally, based on the findings of this project, I will project future challenges in the research of data-to-text NLG in terms of data, applications and evaluation.

Hierarchical Structure in Story Generation

We explore the task of story generation: creative systems that can build coherent and fluent passages of text about a topic. Using a dataset of 300k human-written stories paired with writing prompts, we investigate hierarchical story generation. Our models first generate a premise and then transform it into a passage of text. We develop models that improve the relevance of the story to the premise using a novel form of model fusion and present improvements to self-attention that better capture long-range context. Then, we build upon this work by proposing a coarse to fine mechanism for story generation, decomposing the task into several steps. We first explicitly generate logical verb sequences to model action in stories, then form these into sentences, and finally fill-in character names. We show that such decompositions improve the consistency and diversity of generated stories.