The dictionary-construction process took three years and produced a dictionary of over 20,000 highly accurate terms mapping to more than 90 topics.

Google

To classify reviews into topics that matter to managers and researchers, the Culture 500 team built a custom dictionary from the ground up by combining machine learning tools with human expertise. This approach dramatically improved the accuracy of their models.

Custom Culture 500 Dictionary Combining Machine Learning and Human Expertise

MIT Sloan Management Review project page on measuring culture in leading companies, including the Culture 500 dataset. https://sloanreview.mit.edu/projects/measuring-culture-in-leading-companies/

MIT Sloan Management Review

Construction of the Culture 500 dictionary began by selecting topics that companies frequently included in their corporate values or that had been researched extensively.

Topic Selection for the Culture 500 Dictionary

For each selected topic, researchers built a preliminary dictionary of high-fidelity words and phrases, then had two team members analyze a random sample of at least 20 reviews containing each term. The bootstrapping process generated tens of thousands of terms, which human experts reviewed and refined.

Bootstrapping Process for Culture 500 Dictionary Terms

For each term in the Culture 500 dictionary, the methodology specified its component words, close variants, sequence and distance between words in a phrase, exclusion terms, and common misspellings.

Learn Before

Related