Deep models of Semantic Knowledge (DemoSem)
The project aims to design optimal deep models of assembled pieces of knowledge from corpora (syntagmatic relations) and lexicons (paradigmatic relations) with the help of the mathematical state-of-the-art methods, which include graph-based approaches and deep neural networks. The application environment is Word Sense Disambiguation (WSD) task, since it awaits for its break-through performance that would enhance real-world applications.
Most of the relatively successful methods rely on a specific domain and on the usage of pre-annotated data when creating a knowledge graph (such as, the supervised methods).
The methods, based on graphs, handle dense networks of connected relations, but they suffer from flexibility with respect to the relevance of the presented semantic knowledge. For that reason we will explore methods, based on deep neural networks, which are at the heart of all successful methods for Machine Learning. However, these methods are black boxes with respect to the processes, controlled by humans.
Our previous experience is on: a) language technologies and b) graph-based methods.
Thus, we would like to put our efforts into:
- preparation of data with explicated semantic information (on the base of the senses and valencies);
- modelling of the semantic information, encoded in the language resources, through the interaction between knowledge graphs and deep neural networks;
- training of automatic tools for Word sense Disambiguation and
- testing of the automatic tools on a general domain and large data.
Our main tasks are as follows:
- Creation of language resources and language models of semantic knowledge.
- Application of approaches with deep neural networks to the semantic knowledge.
- Integration of the graph-based methods within the deep neural networks.
- Creation of appropriate models and automatic tools for improving the Word Sense Disambigation task.
Organization: Institute of Information and Communication Technologies (IICT-BAS)
Coordinator of the research team: Associate professor, Dr. Kiril Simov
Natonal Science Fund – 2016