Learning a taxonomy of job market skills from textual resources

点击翻译按钮获取中文摘要

The objective of this project consists of designing tools, using natural language processing and machine learning algorithms, to build automatically a taxonomy using the various skills accumulated in the databases of our industrial partner. This refers to the problem of taxonomy learning (also called taxonomy extraction), which consists of determining the concepts that supersede some others based on a corpus of text providing examples of how concepts might be related together. We will propose models, based on deep learning networks, to incorporate pattern-based features as well as distributional features to determine if two skills are related together. To determine candidate hypernyms that should be considered for a more detailed analysis, morphological analysis as well as statistical analysis (ex. co-occurrence analysis) will be applied. We will construct, train and compare different recurrent models to determine the architectures that are more suited to detect skills that are interrelated together. Distributional representations (i.e. word embeddings) for most of the skills present in the corporate database will be built. We will also extend current approaches to represent out of vocabulary words and compound terms. ********