SpkAtt-2023

This repository contains the data and supplementary materials for Task 2 of the

2023 Shared Task on Speaker Attribution (SpkAtt-2023),

co-located with KONVENS 2023.

The Shared Task competition is run on CodaLab

News: Test data (blind) is now available!

Important Dates:

• February, 2023 - Trial data release
• April 1, 2023 - Training and development data release
• June 8, 2023 - Evaluation enabled on the dev set
• June 15, 2023 - Test data release (blind)
• July 1, 2023 - Submissions open
• July 31, 2023 - Submissions close
• August 14, 2023 - System descriptions due
• September 7, 2023 - Camera-ready system paper deadline
• September 18-22, 2023 - Workshop at KONVENS 2023

Task 2 Data

Data Format:

The data is available as JSON files where document (news article) is a individual JSON file.

Each JSON file has the following format:

{
    "Annotations": [
        {
            "Addr": [],
            "Cue": ["0:1"],
            "Form": "Indirect",
            "Frame": ["0:0", "0:1"],
            "HasNested": false,
            "IsNested": false,
            "Message": ["0:2", "0:3", "0:4", "0:5"],
            "STWR": "Speech",
            "Source": ["0:0"]
        },
    ],
    "DocumentName": "Title of the document",
    "Sentences": [
        {
            "SentenceId": 0,
            "Tokens": ["They", "said", "this", "is", "a", "sentence", "."]
        },
    ]
}

where

• Addr, Cue, Frame, Message and Source are an array of SentenceID:TokenID offsets.
• Form describes the form of message (what kind of quote it is); one of Direct|Indirect|FreeIndirect|IndirectFreeIndirect|Reported
• STWR describes how the message was uttered; one of Speech|Thought|Writing|ST|SW|TW. ST is Speech+Thought, SW is Speech+Writing, TW is Thought+Writing
• HasNested is a boolean indicating if there is another annotation nested within this annotation
• IsNested is a boolean indication if this annotation is nested within some other annotation. The subtask 2 (simplified) only considers annotations with "IsNested": false

Trial Data

Trial data is available under data/trial in this repository. Besides the machine readable JSON, we also provide pretty printed files (ending with .pretty.json in the folder data/trial/pretty) which contain additional information to make them easily human-readable (namely the text of each annotation in addition to the sentence/token offsets).

Data Source

The data for includes news articles from the German WIKINEWS, extracted from the articles XML dump. The entire dump from April 2022 consists of 13,001 published articles, from which we sampled 1000 articles to annotate. These annotated articles contain almost 250,000 tokens.

License

The data is provided under the CC BY-NC-SA 4.0 license.

Supplementary Materials

A code skeleton to read and write the data format can be found in the starter_kit directory. The annotation guidelines (in German) can be found in doc directory.