Textinator: an Internationalized Tool for Annotation and Human Evaluation in Natural Language Processing and Generation

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 856–866 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 856 Textinator: an Internationalized Tool for Annotation and Human Evaluation in Natural Language Processing and Generation Dmytro Kalpakchi, Johan Boye Division of Speech, Music and Hearing KTH Royal Institute of Technology Stockholm, Sweden {dmytroka,jboye}@kth.se Abstract We release an internationalized annotation and human evaluation bundle, called Textinator, along with documentation and video tutorials. Textinator allows annotating data for a wide variety of NLP tasks, and its user interface is offered in multiple languages, lowering the entry threshold for domain experts. The latter is, in fact, quite a rare feature among the annotation tools, that allows controlling for possible unintended biases introduced due to hiring only English-speaking annotators. We illustrate the rarity of this feature by presenting a thorough systematic comparison of Textinator to previously published annotation tools along 9 different axes (with internationalization being one of them). To encourage researchers to design their human evaluation before starting to annotate data, Textinator offers an easy-to-use tool for human evaluations allowing importing surveys with potentially hundreds of evaluation items in one click. We finish by presenting several use cases of annotation and evaluation projects conducted using pre-release versions of Textinator. The presented use cases do not represent Textinator’s full annotation or evaluation capabilities, and interested readers are referred to the online documentation for more information. Keywords: annotation tool, human evaluation tool, natural language processing, natural language generation 1. Introduction Large-scale pretrained language models, also called foundation models, have brought substantial advances to the Natural Language Processing (NLP) field. In- deed, currently the dominant approach for many NLP tasks is to fine-tune a foundation model on relatively small amounts of task-specific annotated data (Bom- masani et al., 2021, Section 2.1.2). These recent de- velopments introduce two major shifts to the field. The first shift is from focusing on large variable- quality crowdsourced annotated datasets back to rela- tively small high-quality datasets. The second shift is from focusing mostly on methods to focusing mostly on evaluation. In this paper we attempt at addressing both shifts by presenting a tool, called Textinator1, suitable for annotating data for a wide variety of NLP tasks, as well as facilitating human evaluation. Addressing the first shift in more detail, we believe that first and foremost, high-quality data should have correct and consistent annotations. Such annotations are easier to obtain with the help of domain experts, who are, evidently, more expensive to hire than crowd- workers. This is why annotation tools should have a low entry threshold and require no specialized training. Secondly, high-quality data should limit biases only to those necessary for exploring the concept at hand. For instance, factual question answering datasets will nec- essarily be biased towards wh-questions, but will usu- ally not contain why-questions. 1Both the code and the documentation are available at https://github.com/dkalpakchi/Textinator (DOI: 10.5281/zenodo.6497334) One very subtle (and usually unintended) way of intro- ducing bias in the …

Content truncated to 500 words