WVOQ at SemEval-2021 Task 6: BART for Span Detection and Classification

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 270–274 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 270 WVOQ at SemEval-2021 Task 6: BART for Span Detection and Classification Cees Roele cees.roele@gmail.com Abstract Simultaneous span detection and classification is a task not currently addressed in standard NLP frameworks. The present paper describes why and how an EncoderDecoder model was used to combine span detection and classifi- cation to address subtask 2 of SemEval-2021 Task 6. 1 Introduction Task 6 of SemEval-2021 studies the detection of persuasion techniques (Dimitrov et al., 2021). The task considers English language memes, which in subtasks are to be classified, divided into classified fragments having a begin and end, and classified when text is combined with images. Of the three subtasks described in the paper, the present paper primarily addresses resolving subtask 2: Given only the ”textual content” of a meme, identify which of the 20 tech- niques are used in it together with the span(s) of text covered by each technique. This is a multilabel sequence tagging task. The figure below illustrates span detection and classification for three technique classes for a meme. Figure 1: Span detection and classification: overlap- ping spans and spans extending over multiple sentences In the above figure the ellipsis at the end of the sentence denotes the continuation of the second sentence. Note that loaded language and name calling both apply to the same span, that is, the word ”CRACKHEAD”. Note furthermore, that the span for smears overlaps with both of these spans and ranges over more than one sentence. The present paper describes a novel approach to resolving these requirements by generating XML- like start and end tokens to delineate spans. The following illustrates this for the message in figure 1. <SMEARS> WHY DO THEY KEEP SAYING RELEASE THE <LOADED−LANGUAGE> <NAME−CALLING> CRACKHEAD </NAME−CALLING> </LOADED−LANGUAGE> ? “n HUNTER ... </SMEARS> It attained an F1 score on the test set that is about in the middle of the baseline and the highest ranking score. The choice of this approach of generating markup to identify spans was made on the basis that it was technically possible, easily understand- able at a behavioral level of input and output, and using a model that is pre-trained for dealing with spans. The aim was not so much to attain the high- est score as to explore how effective this approach is in a proof-of-concept and what problems need to be overcome to bring it to good performance.1 2 Background Propaganda messages are constructed using spe- cific rhetorical techniques. The current task is to identify within a message in what fragment a par- ticular technique is invoked. 1The code for the described system is avail- able at: https://github.com/ceesroele/ SemEval-2021-Task-6. 271 A simpler task is to identify fragments of a mes- sage in which any propaganda technique is used. Effectively, this comes down to classifying any part of a message as being either propaganda or not. It is a sequence labeling …

Content truncated to 500 words