Annotation Lab v2.4.0 adds relation creation features for Visual NER projects and redesigns the Spark NLP Pipeline Config on the Project Setup Page. Several bug fixes and stabilizations are also included. Here’s are the highlights:
Relations on Visual NER Projects
Annotators can create relations between annotated tokens/regions in Visual NER projects. This functionality is similar to what we already had in text-based projects. It is also possible to assign relation labels using the contextual widget (the “+” sign displayed next to the relation arrow).
Spark NLP Pipeline Config redesigned
The Spark NLP Pipeline Config
Setup Page was redesigned to ease filtering, collapsing, and expanding models and labels. For a more intuitive use, the
Add Label button was moved to the top right side of the tab and no longer scrolls with the config list.
This version adds a lot of improvements to the new Setup Page. The
Training & Active Learning Tabs are only available to projects for which Annotation Lab supports training. When unsaved changes are present in the configuration, the user cannot move to the
Training & Active Learning Tab and/or
Training cannot be started. When the OCR server was not deployed, imports in the Visual NER project were failing. With this release, when a valid Spark OCR license is present, the OCR server is deployed and the import of pdf and image files is executed.
When a login session is timed out and cookies are expired, users had to refresh the page to get the login screen. This known issue has been fixed and the user will be redirected to the login page.
When a task was assigned to an annotator who does not have completions for it, the task status was shown incorrectly. This was fixed in this version.
While preannotating tasks with some specific types of models, only the first few lines were annotated. We have fixed the Spark NLP pipeline for such models and now the entire document gets preannotations. When Spark NLP for Healthcare license was expired, the deployment was allowed but it used to fail. Now a proper message is shown to Project Owners/Managers and the deployment is not started in such cases.
Inconsistencies were present in training logs and some embeddings were not successfully used for models training.
Along with these fixes, several UI bugs are also fixed in this release.