News

News and their relation with the project, including a mapping to the technical tasks.

27/09/24: I attended INLG in Tokyo, and had lots of talks and fun with the great INLG community. Our demo 'Filling Gaps in Wikipedia: Leveraging Data-to-Text Generation to Improve Encyclopedic Coverage of Underrepresented Groups' got the Best Demo Paper award there!

31/08/24: The project will be continued until end of November 2024. 

15/07/24: Our two DCU-NLG system demonstrations submitted to INLG were accepted: 'Filling Gaps in Wikipedia: Leveraging Data-to-Text Generation to Improve Encyclopedic Coverage of Underrepresented Groups', and 'QCET: An Interactive Taxonomy of Quality Criteria for Comparable and Repeatable Evaluation of NLP Systems'. We will present live demos in Tokyo!

• Task 1.3: In-parallel RGB and NLMB combination.

• Task 4.2: Human assessment for intrinsic text quality.

• Task 5.1: Dissemination and communication.

03/06/24: Our tutorial 'Human Evaluation of NLP System Quality' will take place at INLG in Tokyo in September, just before the main conference.

• Task 4.2: Human assessment for intrinsic text quality.

• Task 5.1: Dissemination and communication.

14/05/24: Our paper 'New Irish Language Resources and Tools for Responsible Generative AI' got the Best Overall Contribution award at the 2024 ADAPT Scientific Conference!

• Task 5.1: Dissemination and communication.

11/04/24: With Mohammed Sabry (DCU-NLG), we submitted system outputs at the GEM D2T shared task! Our outputs are rule-based generated texts post-edited with a  small language model for improved fluency. Looking forward to the results!

• T1.2: Sequential RGB and NLMB combination.

06/12/23: I presented our paper "Generating Irish Text with a Flexible Plug-and-Play Architecture" (see below) at the Pan-DL workshop at EMNLP in Singapore. 

• Task 5.1: Dissemination and communication.

09/10/23: Our paper "Generating Irish Text with a Flexible Plug-and-Play Architecture" has been accepted at the Pan-DL workshop! The paper is an extension to the technical report submitted to the MM-NLG workshop, and also reports on the creation of  an Irish dataset following the Mod-D2T paper. 

• Task 1.1: FleNS architecture code design.

• Task 2.4:  Extension to Irish.

• Task  3.1: Dataset compilation and creation.

29/09/23:  I attended #ExploreAI event organised by ADAPT, as part of the START European Researcher's night in Trinity college. More info here and here.  Together with Elaine  Uí Dhonnchadha and Lauren Cassidy we presented our live demo for Irish (and English) Wikipedia page generation! The demo is available here, the poster here.

• Task 4.3: Extrinsic evaluation method for NLG.

• Task 5.1: Dissemination and communication.

16/09/23: I attended the INLG conference to present our Irish text generator at the MM-NLG workshop and our ModD2T paper in the main conference. More info here.

• Task 5.1: Dissemination and communication.

12/07/23: Our paper "Mod-D2T: A Multi-layer Dataset for Modular Data-to-Text Generation" has been accepted at INLG! We present a new dataset built using the whole M-FleNS pipeline. 

• Task 1.1: FleNS architecture code design. 

• Task 3.1: Dataset compilation and creation.

10/07/23:  Our paper  from the GEM Human Evaluation group "A Needle in a Haystack: An Analysis of High-Agreement Workers on MTurk for Summarization" was presented today at ACL! This paper and the reproducibility paper led by Anya Belz can now be found on the ACL Anthology  (see Resources).  These contributions will affect the design of M-FleNS's human evaluations.

• Task 4.2: Human assessment for intrinsic text quality.

16/06/23:  We completed a first version of the Irish generator and submitted Irish outputs to the WebNLG shared task! We use the RGB M-FleNS's modules for generation, piped with the Irish NLP tools for morphology processing. A great collaboration with Elaine Ui Dhonnchadha (TCD), Stamatia Dasiopoulou (NTT Data), and the local DCU NLP crew with Brian Davis, Lauren Cassidy and Anya Belz.

• Tasks 2.1 and 2.4: Maximisation of language-independent rules;  Extension to Irish.

08/05/23:  Our (very) collaborative paper "NL-Augmenter: A Framework for Task-Sensitive Natural Language Augmentation" has been published by the  Northern European Journal of Language Technology (NEJLT) and is now available for download! The proposed framework can be used to assess many aspects of the texts produced in M-FleNS.

• Task 4.1: Automatic metrics for output selection.

11/03/23: I presented our work in progress on predicate-argument template extraction at the Georgetown University Round Table, which hosted SyntaxFest'23.  This is the first stone for the automatic extension of the RGB system in M-FleNS!

• Task 2.2: Automatic coverage extension.

• Task 5.1: Dissemination and communication.

02/03/23: I published the code for running WebNLG metrics on Colab (see Resources). We'll use these metrics for the main automatic evaluation of the M-FleNS outputs.

• Task 4.1: Automatic metrics for output selection.

14/12/22: I released the code for running the M-FleNS architecture (see Resources).

• Task 1.1: FleNS architecture code design.

23/11/22: I released the code for converting UD trees into predicate-argument trees (see Resources).

• Task 3.1: Dataset compilation and creation.