Test Environment for a TagTool_WiZArD Named Entity Recognition Plugin

Introductory remarks

In the long run TagTool_WiZArD application (ttw) (see https://github.com/pBxr/TagTool_WiZArd) shall be enabled to integrate Natural Language Processing features. In this first step an environment is beeing created in which Named Entity Recognition processes can be tested using the Hugging Face Transformers pipeline (see https://huggingface.co) and a variant of the BERT language model family (see below).

The final aim is to create a plugin for the productive version of ttw application.

Approach

In this version the focus lies on the extraction of location names. To be able to vary and to handle the test environments in a simple way the Anaconda plattform (see https://www.anaconda.com/) is used in combination with Jupyter Notebooks where the Python code is run.

• It is simulated that the plugin receives path, name of the text file and further settings from ttw.
• The plugin gets the plain text from the selected text file and saves it in a folder ("NER_results").
• It extracts a tokenized list of result entries that are beeing re-merged to a list of place names.
• The place names are run through the iDAI.gazetteer webservice to identify the locations and extract the gazetteer-IDs, that ttw needs as .csv list to enrich the article file with geographic authority data (see also https://github.com/pBxr/ID_Extractor for ttw).
• A log file with the tokenized results, a draft for the final .csv list and the complete gazetteer query result as .json file will be saved in the NER_results folder.

Technical details

• dslim/bert-base-NER is very suitable for this testing purposes concerning its performance and accuracy. Here are the main environments parameters this version is tested with:
• Anaconda Navigator 2.6.0 (Windows)
• Python 3.7.0
• PyTorch 1.4.0
• TensorFlow 2.3.0
• Transformers 4.11.3
• For all other questions see especially https://huggingface.co/docs/transformers/installation.
• In case of problems check whether to use pip or conda-forge channel to install components.

Other Aspects

• The NER results show quality differences that seem to depend on the state of the plain text (with or without title, keywords and abstracts, extracted with bs4 from structured text files or from plain text etc.). So the pre-processing of the texts needs more focus.
• The insufficient quality of the iDAI.gazetteer query results was ignored for this first test version (as well as the webservice´s default query limit). To work on filter mechanisms to improve the quality of the result will be a task for forthcoming commits.