Skip to content
This repository has been archived by the owner on Nov 18, 2021. It is now read-only.

Latest commit

 

History

History
112 lines (82 loc) · 5.43 KB

deepfrog.md

File metadata and controls

112 lines (82 loc) · 5.43 KB
Raw

Deep-learning for Dutch text mining

Metadata

  • Status: Proposed
  • Type: Generic
  • Work Package: WP3
  • Research Coordinators: Maarten van Gompel
  • Coordinators for CLARIAH: Maarten van Gompel
  • Participating Institutes: KNAW HuC, Radboud University Nijmegen
  • End-users: Researchers/developers, specific research is being conducted already by Tess Dejaeghere and Vincent vandeGhinste in the scope of CLARIAH-VL
  • Developers: Maarten van Gompel
  • Interest Groups: Text
  • Task IDs: T139b (Deepfrog)

Description

This generic use-cases proposes builds upon the Automatic linguistic enrichment for Dutch texts use-case. Do note that there is currently no official time or budget allocated for this initiative yet. Nevertheless, we already got started on it and have a first prototype.

What is the research about?

The various NLP modules in Frog were built on k-NN classifiers, whereas the current state-of-the art tends to converge to using a variety of neural transformer-based model.

What problem is hindering the research?

The current Frog may or may not leap behind the current state-of-the-art, whether that is the case is still to be determined. To research this we propose a next generation Frog which we call DeepFrog that offers ground for comparison.

What is needed to do the research?

Frog currently consists of multiple components addressing different kinds of linguistic enrichment, this will be no different in the proposed DeepFrog. Initial emphasis will be on:

  • Part-of-Speech Annotation
  • Lemmatisation
  • Named-entity Recognition

This leaves morphological analysis, dependency parsing and shallow parsing for a later stages>

The aim is to use the same training data for the modules of DeepFrog as has been used for Frog, in this way, a fair comparison between the two systems can be drawn. This comparison constitutes the main research objective of this research; the hypothesis is that by applying state-of-the-art deep learning techniques we can improve the accuracy for each of our very specific Dutch linguistic enrichment tasks.

Using the same training data also implies that the same tagsets will be used (i.e. the CGN tagset for Part of Speech tagging). This provides a necessary continuity for the possible future adoption of the new tool if it proves successful, as it can then be a stand-in replacement for the current Frog.

Unlike earlier work on Frog, this project will build an entirely new code base rather than an update the existing Frog codebase, as such they can live alongside and independently of eachother. The reason for a new codebase is the fact that we are effectively replacing the very memory-based algorithms that have powered Frog for years with different ones based on neural networks. A new Python codebase would be more appropriate than the current C++ one in the light of the deep learning libraries that are available and the need to do rapid prototyping to develop and test the new models. The performance penalty that comes with the interpreted overhead (Python) can be largely disregarded as the underlying deep learning libraries that perform the bulk of the work are native implementations anyway.

We will adhere to the same input and output formats as the current Frog: FoLiA XML input/output, plain text input, and columned output. The idea is that DeepFrog can act as a more-or-less drop-in replacement.

Data

We will leverage existing pre-trained models for Dutch such as BERTje, Roberta, and fine-tune those on specific NLP tasks.

Tools

We will build on existing third-party libraries such as transformers library offered by huggingface and its port to Rust.

DeepFrog will be a command-line tool that integrates the necessary data layers, but most of the models we train can also be used independently of the tooling we develop and have merit on their own.

What software and services are involved?

How to evaluate this?

A study needs to be done to assess the efficacy of the DeepFrog with Frog. An initiative to this end, using the first protype we delivered, has already started in the Flemish CLARIAH with Vincent Vandeghinste and intern Tess Dejaeghere. Further interest from Flanders has also been expressed by Walter Daelemans.

References

Related use-cases:

Links: