The aim of the SLING project is to learn to read and understand Wikipedia articles in many languages for the purpose of knowledge base completion, e.g. adding facts mentioned in Wikipedia (and other sources) to the Wikidata knowledge base. We use frame semantics as a common representation for both knowledge representation and document annotation. The SLING parser can be trained to produce frame semantic representations of text directly without any explicit intervening linguistic representation.
The SLING project is still work in progress. We do not yet have a full system that can extract facts from arbitrary text, but we have built a number of the subsystems needed for such a system. The SLING frame store is our basic framework for building and manipulating frame semantic graph structures. The Wiki flow pipeline can take a raw dump of Wikidata and convert this into one big frame graph. This can be loaded into memory so we can do fast graph traversal for inference and reasoning over the knowledge base. The Wiki flow pipeline can also take raw Wikipedia dumps and convert these into a set of documents with structured annotations extracted from the Wiki markup. This also produces phrase tables that are used for mapping names to entities. There is a SLING Python API for accessing all this information and we also have a bot for uploading extracted facts to Wikidata.
The SLING parser is used for annotating text with frame semantic annotations. It is a general transition-based frame semantic parser using bi-directional LSTMs for input encoding and a Transition Based Recurrent Unit (TBRU) for output decoding. It is a jointly trained model using only the text tokens as input and the transition system has been designed to output frame graphs directly without any intervening symbolic representation.
The SLING framework includes an efficient and scalable frame store implementation as well as a neural network JIT compiler for fast training and parsing.
A more detailed description of the SLING parser can be found in this paper:
- Michael Ringgaard, Rahul Gupta, and Fernando C. N. Pereira. 2017. SLING: A framework for frame semantic parsing. http://arxiv.org/abs/1710.07032.
SLING is also the foundation for KnolCase. KnolCase is a distributed case-based knowledge management tool for gathering information about subjects of interest and organizing these into case files, which can be shared with others and published to public knowledge bases.
- Installation and building
- Wikipedia and Wikidata processing
- SLING Python API
- CASPAR frame semantics parser
- Myelin neural network JIT compiler
- Semantic frames
- SLINGDB persistent key-value store
Original authors of the code in this package include:
- Michael Ringgaard
- Rahul Gupta
- Anders Sandholm
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent