An Evaluation of Layout Features for Information Extraction from Calls for Papers

We describe a feature-rich conditional random field model for the extraction of conference and workshop information (e.g. name, date, location, deadline) from calls for papers (CFPs). This has applications in the automatic construction of a conference knowledge base from a collection of CFPs. Relevant information in CFPs is often contained in regions that do not contain complete, grammatical sentences, but can be distinguished visually from other parts of the text by their formatting. We show that in this situation layout features, i.e. features that measure physical layout properties of a text, improve extraction accuracy considerably. On a corpus of CFPs we observe a 30 % gain in F1 through the use of layout features.