Mapping Explicit and Implicit Discourse Relations between the RST-DT and the PDTB 3.0

Proceedings of Recent Advances in Natural Language Processing, pages 22–31 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_003 22 Cross-lingual Classification of Crisis-related Tweets Using Machine Translation Shareefa Al Amer1,2, Mark Lee1, Phillip Smith1 1School of Computer Science, University of Birmingham, United Kingdom 2College of Computer Science & Information Technology, King Faisal University, Saudi Arabia alamersharifah@gmail.com, {m.g.lee,p.smith.7}@bham.ac.uk Abstract Utilisation of multilingual language models such as mBERT and XLM-RoBERTa has in- creasingly gained attention in recent work by exploiting the multilingualism of such mod- els in different downstream tasks across differ- ent languages. However, performance degra- dation is expected in transfer learning across languages compared to monolingual perfor- mance although it is an acceptable trade-off considering the sparsity of resources and lack of available training data in low-resource lan- guages. In this work, we study the effect of machine translation on the cross-lingual trans- fer learning in a crisis event classification task. Our experiments include measuring the effect of machine-translating the test data into the source language and vice versa 1. We evalu- ated and compared the performance in terms of accuracy and F1-Score. The results show that translating the source data into the target language improves the prediction accuracy by 14.8% and the Weighted Average F1-Score by 19.2% when compared to zero-shot transfer to an unseen language. 1 Introduction We are interested in discovering methods to en- hance the detection of emerging crises in social media and the transferability of a classification model fine-tuned on a specific language to other languages. Crisis event detection typically relies on two aspects: (1) the detection of burstiness of cer- tain keywords or trends in the timeline, and (2) the classification of the detected bursts and whether they indicate the occurrence of a disaster or not. The former aspect can be obtained using unsuper- vised learning to cluster posts that have certain commonalities such as common keywords, time, 1We refer to Arabic as the target language (i.e., language of testing data) and to English as the source language (i.e., language of training data). and location. However, the latter usually depends on the use of supervised learning algorithms to fil- ter out non-relevant posts since bursts can occur for non-crisis related events such as concerts and media events, political events, and other trends that can interfere with the task of responding to emer- gencies. It is especially important to develop meth- ods to increase the accuracy of classifying relevant posts to support multilingual settings and there- fore help provide better response to emergencies. Although the use of machine translation for cross- lingual transfer learning has shown promising re- sults, there are several drawbacks to the existing work including quality of machine translation, lim- ited parallel data, and structural differences which affect the overall performance of the final model. In this work, we conducted several experiments to assess the effect of machine translation in bridg- ing language gaps for zero-shot cross-lingual clas- sification of crisis-related tweets. Additionally, we investigated potential limitations on the final pre- dictions. Our study focused …

Content truncated to 500 words