Keyword Extraction and Semantic Tag Prediction

Content on the web is often organized through user generated tags for intuitive search and retrieval. Such tags convey meta-information about the subject matter of the texts they represent. For this project, we applied machine learning (Bayesian co-occurrence, k-NN, SVM, NNS) to predict tags of StackExchange posts obtained from Kaggle: “Facebook Recruiting Keyword Extraction III.” Using our non-parametric, fuzzy Nearest Neighbor Search algorithm, we achieved a F1-score of 0.471 on a testing set with unseen data and 0.773 on the Kaggle test set (containing many duplicate data points). Furthermore, when predicting a single tag per post, our algorithm attained approximately 71.1% accuracy (on unseen data), surpassing the 0.65 accuracy attained by Stanley & Byrne (2013). Our keyword-tag co-occurrence model and fuzzy NNS proved to be fast and practical for large-scale subject and tag prediction problems with tens-of-thousands of tags and training documents.