The FLEXILOG project was funded as a Starting Grant from the European Research Council (ERC). It was a 5 year project led by Steven Schockaert, running from May 2015 until April 2020, and is hosted by Cardiff University. Its objectives were:
- to learn interpretable vector space representations of entities and their relationships
- to use these vector space representations as a basis for plausible reasoning about ontologies, rule bases, and other forms of structured knowledge
- to use these vector space representations as a basis for learning structured domain theories from the web
Main aims
Vector space embeddings have become a popular representation framework in many areas of natural language processing and knowledge representation. In the context of knowledge base completion, for example, their ability to capture important statistical dependencies in relational data has proven remarkably powerful. These vector space models, however, are typically not interpretable, which can be problematic for at least two reasons. First, in applications it is often important that we can provide an intuitive justification to the end user as to why a given statement is believed, and such justifications are moreover invaluable for debugging or assessing the performance of a system. Second, the black box nature of these representations makes it difficult to integrate them with other sources of information, such as statements derived from natural language, or from structured domain theories. Symbolic representations, on the other hand, are easy to interpret, but classical inference is not sufficiently robust (e.g. in case of inconsistency) and too inflexible (e.g. in case of missing knowledge) for most applications.
The overall aim of the FLEXILOG project was to develop novel forms of reasoning that combine the transparency of symbolic methods with the flexibility and robustness of vector space representations. For example, symbolic inference can be augmented with inductive reasoning patterns (based on cognitive models of human commonsense reasoning), by relying on fine-grained semantic relationships that are derived from vector space representations. Conversely, logical formulas can be interpreted as spatial constraints on vector space representations. This duality between logical theories and vector space representations opens up various new possibilities for learning interpretable domain theories from data, which will enable new ways of tackling applications such as recognising textual entailment, automated knowledge base completion, or zero-shot learning.