Multilingual information extraction for job market analysis

In today’s digital societies, language remains the mechanism for sharing information. As languages other than English make their way into big data, the need for language understanding technology able to robustly extract information from the rich diversity of languages keeps growing. If achieved, technology can make valuable information more accessible to people.
Natural Language Understanding (NLU), an important field within AI, has made promising advances towards this objective. What is crucially missing is an integrated approach. Traditional NLU techniques are limited to a certain source X (X=a particular language, text domain, a particular task). When moving from X to a new source Y, we start from scratch and need a substantial amount of annotated training data to develop Y-specific extraction capabilities. I propose a new paradigm that learns from multiple Xs simultaneously, and combines the merits of transfer learning and weak supervision, by using neural networks that learn their own representations. I propose to exploit such representations to guide learning and broaden the scope of NLU: to anticipate performance to make the big search space feasible, to select few examples for human labeling, and to acquire multilingual meaning representations that transfer better to new applications. Besides existing benchmarks, the designed technology is tested on a use case provided by Styrelsen for Arbejdsmarked og Rekruttering to derive insights into demanded job skills.

The project will combine my expertise in transfer learning and weak supervision. MultiVaLUE will make prototypes publicly available to accelerate innovation in the use of linguistic data for knowledge discovery.