| Title | : | Natural Language Generation from Structured Data |
| Speaker | : | Shreyas Shetty M (IITM) |
| Details | : | Thu, 12 Oct, 2017 2:00 PM @ A M Turing Hall |
| Abstract: | : | Neural models for Natural language generation (NLG) have been successfully used in a wide variety of tasks such as machine translation, summarization, dialog generation, etc. In this work, we focus on one such NLG task which involves rendering natural language descriptions from a structured table of facts. A typical example of such a task is generating product descriptions from structured catalog tables. Unlike other NLG tasks, the input here has a specific structure which can be exploited for building richer neural models. We experiment with a recently released dataset WikiBio which contains fact tables about people and their corresponding one line biographical descriptions in English. We propose two models (i) Flat model and (iii) Hierarchical model. Our models leverage the seq2seq framework, and encode the structural properties of the input data. Our experiments show that the proposed models give (i) 10% relative improvement (Flat model) and (ii) 18.8% relative improvement (Hierarchical model) over a recently proposed state of the art method. Qualitative observations demonstrate the ability of the proposed models to capture interesting patterns in sentence generation. |
