Domain-Adaptable Hybrid Generation of RDF Entity Descriptions

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 306–315, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Domain-Adaptable Hybrid Generation of RDF Entity Descriptions Or Biran∗ n-Join or@n-join.com Kathleen McKeown Columbia University kathy@cs.columbia.edu Abstract RDF ontologies provide structured data on entities in many domains and continue to grow in size and diversity. While they can be useful as a starting point for generating descriptions of entities, they often miss important information about an entity that cannot be captured as simple relations. In addition, generic approaches to generation from RDF cannot capture the unique style and content of specific domains. We des- cribe a framework for hybrid generation of entity descriptions, which combines gene- ration from RDF data with text extracted from a corpus, and extracts unique aspects of the domain from the corpus to create domain-specific generation systems. We show that each component of our appro- ach significantly increases the satisfaction of readers with the text across multiple ap- plications and domains. 1 Introduction RDF ontologies are a wonderful source for genera- tion: they feature standardized structure, are con- stantly expending and span many interesting dom- ains. However, generation from RDF introduces two major difficulties. First, RDF contains relati- onships between entities but often lacks other im- portant information about an entity (e.g., historical background and context) which is hard to capture with simple relations. Second, RDF data spans many domains, and presents the difficulty of hand- ling specific domains in generation. Generally speaking, there are three approaches: domain-specific approaches (with hand-written or other rules relevant to each domain), which are not scalable; generic approaches (generating in ∗ Work done while at Columbia University exactly the same way for all domains) which re- sult in unnatural text and miss important content; and domain adaptation, which attempts to automa- tically transfer an approach from one domain to another. Our approach aims to leverage the advantages of all three. We present a generic framework of generation meta-systems for RDF applications, which uses domain adaptation to create domain- specific systems. Biography and Company Des- cription are examples of applications (an applica- tion is the description of RDF entities of a particu- lar type), while Politician and Model are examples of domains within the Biography application. The reason our framework is able to adapt to new domains automatically is that it relies on hy- brid concept-to-text (C2T) and text-to-text (T2T) generation: part of the generated text consists of messages that are created from structured data ac- cording to a generic recipe, while another part co- mes from messages extracted from a domain cor- pus. In addition, we use existing methods to ex- tract paraphrases and discourse models from the domain corpus, which further refines how text is generated differently for each domain. 2 Related Work Generation from RDF data is not a new to- pic. Duboue and McKeown (2003) described a content selection approach for generation from RDF data. Sun and Mellish (2006) present a domain-independent approach …

Content truncated to 500 words