/AI2D-RST

Repository for the conference article "Enhancing the AI2 Diagrams dataset using Rhetorical Structure Theory", published in the Proceedings of the 11th International Language Resources and Evaluation Conference.

Primary LanguagePython

AI2D-RST

This repository contains code for the conference article "Enhancing the AI2 Diagrams dataset using Rhetorical Structure Theory" by Tuomo Hiippala and Serafina Orekhova, published in the Proceedings of the 11th Language Resources and Evaluation Conference.

Caveat

Note that this repository is very much a work in progress, as the scripts in the utils directory relate to unpublished and ongoing work.

If you came here for the LREC paper, see the section below.

LREC paper

Please find the conference paper here.

Application of Rhetorical Structure Theory to the AI2D dataset

This section documents common RST relations encountered during the annotation and presents examples of their application to the AI2D dataset. These examples are intended to illustrate how we applied RST to the AI2D diagrams, particularly during the experiments measuring agreement between annotators, as reported in the conference paper.

Relation Description Example
restatement A relationship in which two diagram elements could stand in for each other in the context of the entire diagram. Due to their independent status, both elements are annotated as nucleus. The food web in 318.png would remain understandable even if either illustrations or their labels removed, because they restate each other.
identification A relationship in which an element identifies another diagram element or its part. The identifying element is marked as a satellite while the identified element is annotated as nucleus. The labels in 2891.png identify specific parts of the cell.
effect A relationship in which one element affects another. The source of the effect is marked as satellite, while the affected element is marked as nucleus. Effect is a common relation in graph-like diagrams, such as the food web in 519.png, indicating that gulls sustain foxes.
sequence A temporal or spatial sequence between diagram elements. Both elements are marked as nucleus. The life cycle of a mosquito illustrated in 108.png shows a sequence of its various stages of development.
property-ascription A relationship in which one element describes the properties of another element. The element providing the properties is marked satellite, whereas the one being described annotated as nucleus. 1031.png shows groups of different types of plant leaves. Their properties (shape) are described using written labels.
title A relationship in which one element acts as a title for the entire diagram or its part. The element providing the title is marked as satellite, while the entire diagram is annotated as nucleus. The upper part of 555.png features a title, which describes the entire diagram. 108.png, in turn, features a title right in the middle of the diagram.
none A relationship used during to flag problematic relations during the annotation process: also used if one of the bounding boxes marking the elements is missing. This relation is not included in the final AI2D-RST corpus. The copyright information in 564.png is not a part of the rhetorical structure of the diagram.

The confusion matrix below shows how we used the categories above to describe the discourse relations extracted from the dataset.

drawing