Creating Relational Data from Unstructured and Ungrammatical Data Sources

An Intelligent System for Identifying Influential Words in Real-Estate Classifieds

This paper focuses on the problem of quantifying how certain words in a text affect, positively or negatively, some numeric signal. These words can lead to important decisions for significant applications such as E-commerce. For example, consider the corpus of real-estate classifieds, which we developed as a case study. Each classified has a description of a real-estate property, along with simple features such as the location and the number of bedrooms. The problem then is to identify which keywords influence the price of the property. Such identification is complicated due to the existence of simple features (numeric and nominal attributes) that also affect the price. In this research, we propose a two-stage regression model to solve this problem. To assess our contribution, we analyze, as a case study, four corpora of real-estate classifieds. The analysis shows that our model predicts the price of a real-estate unit more accurately using the accompanying text, compared to the prediction relying only on simple features. We also demonstrate the capability of our model to annotate (automatically) words that affect the price positively or negatively.