tax2vec: Constructing Interpretable Features from Taxonomies for Short Text Classification

The use of background knowledge is largely unexploited in text classification tasks. This paper explores word taxonomies as means for constructing new semantic features, which may improve the performance and robustness learned classifiers. We propose tax2vec, a parallel algorithm taxonomy-based demonstrate its on six short problems: prediction gender, personality type, age, news topics, drug side effects effectiveness. constructed combination with fast linear classifiers, tested against strong baselines such hierarchical attention neural networks, achieves comparable results documents. algorithm’s also few-shot learning setting, indicating that inclusion features can data-scarce situations. tax2vec capability to extract corpus-specific keywords demonstrated. Finally, we investigate space potential where observe similarity well known Zipf’s law.