Objectives and tasks

The general objective of M-FleNS is to develop a new type of Natural Language Generation (NLG) system that combines the best of Rule/Grammar-Based (RGB) and Neural/Language Model-Based (NLMB) methods, called Flexible Neuro-Symbolic (FleNS) NLG.

O1: To develop (1) modular FleNS architectures for flexible combination of RGB and NLMB methods, and (2) new hybridised neuro-symbolic NLG methods.

T1.1: FleNS architecture code design.

T1.2: Sequential RGB and NLMB combination.

T1.3: In-parallel RGB and NLMB combination.

T1.4: Fully hybrid methods.

O2: To develop RGB methods that maximise language independence and coverage, and can flexibly combine with NLMB methods at module and system levels; the RGB system we will start with is FORGe.

T2.1: Maximisation of language-independent rules.

T2.2: Automatic coverage extension .

T2.3: Extension to French.

T2.4: Extension to Irish (Low res.).

O3: To develop NLMB methods that can flexibly combine with RGB methods at the module and system level to form in-parallel and in-sequence FleNS systems.

T3.1: Dataset compilation and creation.

T3.2: Fine-tuning of pre-trained models.

T3.3: Creation of new NLMB sub-modules.

O4: To create a range of evaluation methods for evaluating the quality of generated texts in terms of criteria including Fluency, Grammaticality, and Accuracy.

T4.1: Automatic metrics for output selection.

T4.2: Human assessment for intrinsic text quality.

T4.3: Extrinsic evaluation methods (Wiki articles).

O5: To establish strategic collaborations and build research community interest in neuro-symbolic architectures and methods.

T5.1: Dissemination/Communication (papers, Workshop and Shared Task organisation, etc.).

T5.2: Exploitation.