Corpus-based Content Construction

Proceedings of the 27th International Conference on Computational Linguistics, pages 3505–3515 Santa Fe, New Mexico, USA, August 20-26, 2018. 3505 Corpus-based Content Construction Balaji Vasan Srinivasan, Pranav Ravindra Maneriker, Kundan Krishna, Natwar Modani Adobe Research Big data Experience Lab Bangalore, India [balsrini,pmanerik,kkrishna,nmodani]@adobe.com Abstract Enterprise content writers are engaged in writing textual content for various purposes. Often, the text being written may already be present in the enterprise corpus in the form of past articles and can be re-purposed for the current needs. In the absence of suitable tools, authors manually curate/create such content (sometimes from scratch) which reduces their productivity. To address this, we propose an automatic approach to generate an initial version of the author’s intended text based on an input content snippet. Starting with a set of extracted textual fragments related to the snippet based on the query words in it, the proposed approach builds the desired text from these fragment by simultaneously optimizing the information coverage, relevance, diversity and co- herence in the generated content. Evaluations on standard datasets shows improved performance against existing baselines on several metrics. 1 Introduction Due to the increase in the number of channels and the need for more customized content for different audience segments, content authors need to produce content at a very rapid pace. On the other hand, a lot of such content has already been produced and is available in enterprise content repositories. Ideally, the author looks for such content and re-purposes them for her current need. Due to the tedium of this task, the author may resort to creating the content from scratch. Automating this task can greatly aid the author in improving their overall productivity. Given an enterprise author’s requirement in the form of a snippet describing the content she is try- ing to build, standard querying mechanisms can be utilized to fetch the right set of articles that can cater to her needs. However, constructing an initial piece of ‘article’ that can cater to her needs would still be a non-trivial task. Existing work in this direction focuses on generating content for standard platforms like Wikipedia (Banerjee and Mitra, 2016), where the target content is divided into different ‘categories/section’ and text is generated for each section. While this could be useful, the absence of a predefined structure in an enterprise setting makes the extension of such an approach challenging. This calls for an approach that can automatically build an author’s content covering various aspects of the target article. The content style in an enterprise corpus is fairly consistent. However, there could be multiple rep- resentations of the same information across the repository. This necessitates the content generation to reduce the redundancy in the constructed content and adding topical diversity to the article. Finally, the content generated should be coherent with a cohesive flow of information. To address all these requirements, we propose an approach that automatically generates a coherent and non-redundant article based on an author’s input snippet simultaneously optimizing the relevance of the generated text …

Content truncated to 500 words