Skip to content

Latest commit

 

History

History
39 lines (30 loc) · 1.72 KB

overall.md

File metadata and controls

39 lines (30 loc) · 1.72 KB
Raw

Overall

Parse tree structure

Span-finder treats the parse as a tree. Every node in the tree corresponds to a span in the text, and is associated with a type.

For the shallow semantic parsing problem (e.g. FrameNet), the tree starts with a special node of type @@VIRTUAL_ROOT@@, and corresponds to a dummy span in the text (the begin-of-sentence token). The children of virtual root form the second layer of the tree, and their types are frames (or "events" in SRL), whose corresponding spans are lexical units. Each frame may also have some children, and they are frame elements, whose corresponding spans are their trigger spans.

Not that although span-finder can work with arbitrarily numbers of layers theoretically, we restrict it to predict at most 3 layers for FrameNet parsing. For other tasks that may involve fewer or more layers, you may change that setting in the configuration file.

Ontology

Ontology defines what types can be the children of what types. For example, a frame can be a child of virtual root, while a frame element can only be the child of its corresponding frames. To inform the model this, an ontology file should be specified.

Each line in the ontology defines a rule for a type. The first line should be like

@@VIRTUAL_ROOT@@<TAB>frame1<TAB>frame2<TAB>frame3...

and it defines what types can be used as "frames". The following lines should be like

frame1<TAB>fe1<TAB>fe2<TAB>fe3...
frame2<TAB>fe2<TAB>fe4<TAB>fe5...

Do not forget to replace <TAB> with \t. You may find this file in the provided dataset.

The ontology file is only needed for training. After training, the ontology will be stored in the checkpoint and will be automatically used during inference.