Superframes

Superframes: A Schema And Data for Universal Semantic Role Annotation is a DFG-funded research project led by Kilian Evang that is being conducted at Heinrich Heine University Düsseldorf from 10/2025 to 09/2028.

Resources

Project description

Existing schemas for semantic role labeling (SRL) rely on large lexicons, which makes annotation slow and hampers the adoption of SRL as an everyday tool for corpus linguists and NLP practitioners, the way syntactic parsing has become. I argue that decompositional representations of the meaning of predicates as proposed in some linguistic theories offer a way out of this problem, enabling semantic role annotation and prediction for all languages without the need for language-specific lexicons. But such representations first have to be made practical for the use in large-scale annotation and NLP. In particular, the number of primitive relations used in decompositional representations as well as their structural complexity has to be constrained. In the proposed project, I venture to define such a practical schema as an extension of dependency syntax and use it to annotate existing lexicons as well as large quantities of text in different languages. Based on this data, I propose to develop machine learning models for SRL as well as downstream tasks, and to show that SRL based on decompositional meaning representations improves upon traditional SRL in terms of inter-annotator agreement and machine learning performance.

Aims and objectives

WP1: Creation of annotated corpora

The goal is to create a large multilingual corpus annotated with both Universal Dependencies and Superframes in order to train and evaluate SRL systems with a lexicon-free output format and to enable corpus linguistic studies.

WP2: Creation of lexical resources

The goal is to map PropBank to Superframes in order to achieve interoperability and enable data augmentation.

WP3: Intrinsic evaluation

The goal is to test the suitability of Superframes as an SRL scheme by training and evaluating SRL systems on the produced data.

WP4: Extrinsic evaluation

The goal is to test the suitability of automatically produced Superframes annotation as input to semantic tasks like relation extraction, dialogue systems, and AMR parsing./p>

People

Publications

TBD

Preliminary work

Legal