N L P Tasks

Natural Language Processing

 

 

 

Task Description

Tasks 

SeeDev consists of two subtasks. Participants can choose to participate in either (or both) of these tasks. 

  1. Binary relation extraction task (SeeDev-binary) 

Systems are evaluated for their ability to extract relations involving two entity arguments. Entity annotations are provided. 

  1. Full event extraction task (SeeDev-full) 

Systems are evaluated for their ability to extract all types of events, including events involving more than two entities. There is no trigger word in the SeeDev event representation. Entity annotations are provided. 

Entities and events are specified in the information representation scheme described below. 

Results will be evaluated using precision, recall and F-measure. We plan to focus on Interaction and Regulation relations because the knowledge they carry is more valuable to plant scientists. 

 

Information representation scheme 

The SeeDev representation scheme defines 16 entity types, which can be grouped into higher-level classes: 

Scheme

 

These entities participates in 21 types of events that can be grouped into five categories: 

Where and When

Presence_In_Genotype

Occurrence_In_Genotype

Presence_At_Stage

• Occurrence_During

Localization

 

Function

Involvement_In_Process

Transcription_Or_Translation

Functional_Equivalence

Regulation

• Regulation_Of_Accumulation

• Regulation_Of_Development_Phase

• Regulation_Of_Expression

 Regulation_Of_Molecule_Activity

 Regulation_Of_Process

 Regulation_Of_Tissue_Development


 


 

Composition and Membership

Primary_Structure_Composition

Protein_Complex_Composition

Protein_Domain_Composition

Family_Membership

Sequence_Identity

 

 Interaction

• Interaction

• Binding

 

Each event type can be associated with the Negation modality. Events can have between two and eight arguments. 

The binary event extraction subtask uses a simplified version of the representation scheme including only binary relations. In this version, n-ary events are decomposed into binary relations according to specific rules. Event labels are the same as for the complete version, except for the Is_Linked_To relation, which is specific to the binary framework. 

The detailed formal representation including the role names of the event arguments can be found here. 

Event arguments are strongly typed, which means that all types of entities are not possible as event arguments. The possible combinations of entity types per event, i.e. event signatures, are specified here.