The study reported in this paper aims to contribute to the development of an annotation scheme for evaluative language, based on Appraisal Theory, that addresses key sources of classification problems. In particular. it aims to develops a unified annotation scheme that proposes (1) a three-component annotation model comprising Appraiser, Appraised and Appraisal Element, (2) the operationalised distinction between Affect and Appreciation governed by a criterion of experiencer salience and a criterion distinguishing personal emotions from evaluations of conduct, and (3) a decision framework for the Judgement-Appreciation distinction structured on the target and lexis types operating through override conditions and substitution tests. The revised framework is illustrated with examples selected from a corpus of news discourse in English and is designed to be replicable across future Appraisal-based studies of evaluative language.

As precursor work in preparation for an international standard ISO/PWI 24617-16 Language resource management – Semantic annotation – Part 16: Evaluative language, we aim to test and enhance the reliability of the annotation of subjective evaluation based on Appraisal Theory. We describe a comprehensive three-phase workflow tested on COVID-19 media reports to achieve reliable agreement through progressive training and quality control. Our methodology addresses some of the key challenges through the refinement of targeted guideline refinements and the development of interactive clarification tools, alongside a custom platform that enables the pre-classification of six evaluative categories, systematic annotation review, and organized documentation. We report empirical results that demonstrate substantial improvements from the initial moderate agreement to a strong final consensus. Our research offers both theoretical refinements addressing persistent classification challenges in evaluation and practical solutions for the implementation of the annotation workflow, proposing a replicable methodology for the achievement of reliable annotation consistency in the annotation of evaluative language.

In recent years, distantly-supervised relation extraction has achieved a certain success by using deep neural networks. Distant Supervision (DS) can automatically generate large-scale annotated data by aligning entity pairs from Knowledge Bases (KB) to sentences. However, these DS-generated datasets inevitably have wrong labels that result in incorrect evaluation scores during testing, which may mislead the researchers. To solve this problem, we build a new dataset NYTH, where we use the DS-generated data as training data and hire annotators to label test data. Compared with the previous datasets, NYT-H has a much larger test set and then we can perform more accurate and consistent evaluation. Finally, we present the experimental results of several widely used systems on NYT-H. The experimental results show that the ranking lists of the comparison systems on the DS-labelled test data and human-annotated test data are different. This indicates that our human-annotated data is necessary for evaluation of distantly-supervised relation extraction.