Case Study: Bitvore achieves 93.5% F1 market taxonomy classification with John Snow Labs Spark NLP

Bitvore achieves 93.5% F1 score with John Snow Labs for market taxonomy classification

Bitvore faced the challenge of classifying companies into market segments using the Thomson Reuters Business Classification taxonomy. The primary hurdles included a lack of reliable, subjectively-labeled training data, errors from a keyword-based batch labeling process, and dealing with low-quality or insufficient data for many of the 61 categories. To address this, they partnered with John Snow Labs, utilizing their Spark NLP for a natural language processing solution.

John Snow Labs implemented a solution that included advanced data preprocessing, an outlier detection algorithm to filter mislabeled texts, and a semi-supervised algorithm to automate some labeling. This resulted in a highly accurate multilabel classifier. The final model achieved a 93.5% f1 score on the test set, enabling Bitvore's analytics team to reliably classify companies into market segments and easily monitor model performance.

Bitvore