First experimental result on Revues.org corpus level 1- Part I
Part I: Introduction
We present the first experimental result on Revues.org reference corpus. This corpus constitutes the first level of our reference data extracted from the site Revues.org. Objective here is to automatically label the article references in Revues.org site. That is, in the part of bibliographical references at the end of an article, fields such as author, title, date etc. are automatically detected.
This automatic detection can be realised by building a probabilistic model. A Conditional Random Field (CRF) is a probabilistic learning model typically used for this objective. Conditional Random Fields (CRFs) proposed by Laffery et al. (Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data, 2001) have been developed to label sequential data such as natural language text data. One of the first studies that used a CRF model for the detection of bibliographical information is the work of Peng and McCallum (Accurate Information Extraction from Research Papers using Conditional Random Fields, 2004).
Our first experiments are based on this major technique, CRFs. We build a CRF model on Revues.org corpus level 1 (extracted from the bibliography part at the end of article) using an existing language processing toolkit, MALLET software, developed by Andrew McCallum and his team members of Umass Amherst.
To apply a probabilistic model on a corpus, it is indispensable to well determine a learning data format. As our corpus contains relatively complex information compared to other data used in similar works and output labels are not exactly determined, we need a learning data extraction process before applying a CRF model on the corpus. Our first experiments especially concentrate on the preparation of an appropriate learning data. Then we learn a CRF model on this newly prepared data and verify its general accuracy.
In the next parts, we explain the process of preparation of learning data. It includes the detailed analysis of the characteristics of Revues.org reference corpus. Then we apply a CRF model using MALLET software on the leaning data. We temporarily evaluate a learned CRF model on a test set having identical format with learning data.