Contemporary environmental discourse focuses on effectively communicating ecological vulnerability to raise public awareness and encourage positive actions. Hence there is a need for studies to support accurate and adequate discourse production, both by humans and computers. Two main challenges need to be tackled. On the one hand, the language used to communicate about environment issues can be very complex for human and automatic analysis, there being few resources to train and test NLP tools. On the other hand, in the current international scenario, most texts are written in multiple languages or translated from a major to minor language, resulting in different meanings in different languages and cultural contexts. This paper presents a novel parallel corpus comprising the text of World Wide Fund (WWF) 2024 Annual Report in English and its translations into Italian and Brazilian Portuguese, and analyses their linguistic features.

We seek to explain the causes of the misclassification of the most challenging documents, namely those that no classifier using state-of-the-art, very semantically-separable contextual embedding representations managed to predict accurately. To do so, we propose a taxonomy of incorrect predictions, which we used to perform qualitative human evaluation. We posed two (research) questions, considering three sentiment datasets in two different domains – movie and product reviews. Evaluators with two different backgrounds evaluated documents by comparing the predominant sentiment assigned by the model to the label in the gold dataset in order to decide on a likely misclassification reason. Based on a high inter-evaluator agreement (81.7%), we observed significant differences between the product and movie review domains, such as the prevalence of ambivalence in product reviews and sarcasm in movie reviews. Our analysis also revealed an unexpectedly high rate of incorrect labeling in the gold dataset (up to 33%) and a significant amount of incorrect prediction by the model due to a series of linguistic phenomena (including amplified words, contrastive markers, comparative sentences, and references to world knowledge). Overall, our taxonomy and methodology allow us to explain between 80%-85% of the errors with high confidence (agreement) – enabling us to point out where future efforts to improve models should be concentrated.