Explainable deep neural network for Skills prediction from resumes

Abstract

The automatic identification of skills in text documents (CVs, job offers, articles, etc.) is a task of Natural Language Processing (NLP) that finds its application in the construction of job recommender systems or in the automatic identification of the professional qualifications of researchers, employees or job seekers; this, in order to bridge the "skills gap" that the ATD (Association for Talent Development) defines as a significant gap between the skills held by an organization’s human resources and those it needs for its development. Several researchers have proposed methods for automatic identification of skills in text documents. But those methods, for some of them, only allow the identification of skills explicitly mentioned in the documents, and for others lack of explainability. The main objective of this thesis is to extend the scope of existing approaches by proposing an artificial intelligence model based on Convolutional Neural Networks (CNN) capable of identifying in resumes, a set of skills thatwe will refer to as high-level skills insofar as they can be explained by more basic skills. High level skills as perceived in this work are generally professional qualifications such as "web developer", "programmer analyst", "network administrator", and so forth. In addition, we propose to explain the decisions of the CNN model by illustrating the high-level skills predicted by the terms contained in the resume and that characterize those skills. The first contribution of this thesis is thus, the design of a multi-label classification architecture based on CNN and which use the "binary relevance" approach to multi-label classification of resumes according to skills they contain. Input resumes are transformed into matrices using a self-trained word embedding model and the resulting matrices are submitted as inputs to the CNN. Experiments carried out on a corpus of 30,000 IT resumes collected from the Internet have demonstrated the effectiveness of the model which reaches 98.79% of recall and 91.34% of precision. The second major contribution is at the level of the explainability of the CNN models. Globally, we propose a method to explain the predictions of CNN models built for any text classification problem. More precisely, we describe amethod based on the principle of the LRP (Layer-wise Relevance Backpropagation) algorithm to compute the contributions of the terms selected by the convolution filters to the values predicted by the model. In addition, we show the limitations of the base LRP method and propose an adaptation of the formula that computes the contributions of input features. Finally, propose to identify sufficient and necessary features to simplify the explanations provided to users. The distribution of the relevance obtained with our explanation method is similar to that of LIME, a well-known state of the art model; and the evaluation of the complexity of both methods shows that ours is significantly better than LIME. Furthermore, we show how LIME sometimes assigns a score to terms that have no influence on the output. However, LIME has the advantage to bemodel-agnostic.