Creating classification models from textual descriptions of companies using crunchbase

Abstract

This paper compares different models for multilabel text classification, using information collected from Crunchbase, a large database that holds information about more than 600000 companies. Each company is labeled with one or more categories, from a subset of 46 possible categories, and the proposed models predict the categories based solely on the company textual description. A number of natural language processing strategies have been tested for feature extraction, including stemming, lemmatization, and part-of-speech tags. This is a highly unbalanced dataset, where the frequency of each category ranges from 0.7% to 28%. Our findings reveal that the description text of each company contain features that allow to predict its area of activity, expressed by its corresponding categories, with about 70% precision, and 42% recall. In a second set of experiments, a multiclass problem that attempts to find the most probable category, we obtained about 67% accuracy using SVM and Fuzzy Fingerprints. The resulting models may constitute an important asset for automatic classification of texts, not only consisting of company descriptions, but also other texts, such as web pages, text blogs, news pages, etc.