Semantica: Algorithms and Tools for Multilingual Knowledge Extraction

Text is the main repository of human knowledge and automatic knowledge extraction is now made possible at unprecedented levels as a gigantic number of documents has been digitized.

The goal of this project is to create deep-learning algorithms to carry out a large-scale multilingual knowledge extraction from text.

In previous projects, our group developed algorithms to identify real-world entities, events in texts, semantic roles, and temporal relations. We used supervised machine-learning techniques that we trained on semantically annotated corpora. However, such corpora are a scarce resource for languages like Swedish, but even for German or French.

The proposed project will build on the results from earlier projects and create algorithms that require less, possibly no, manual annotation.

We will explore distant supervision with the mapping of entities in loosely parallel texts such as the Swedish and English versions of Wikipedia. We will investigate algorithms linking entities to existing knowledge bases and multilingual embeddings. We will apply these components to much larger corpora that cover substantial portions of human knowledge.

We will process the documents and store the results with frameworks we designed at Lund. Knowledge extraction is a key element to many advanced applications in text processing.

Generic multilingual tools will make it easier to design semantic applications that we believe will be instrumental in the development of the information society.