GMNLP at SemEval-2023 Task 12: Sentiment Analysis with Phylogeny-Based Adapters

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1172–1182 July 13-14, 2023 ©2023 Association for Computational Linguistics GMNLP at SemEval-2023 Task 12: Sentiment Analysis with Phylogeny-Based Adapters Md Mahfuz Ibn Alam∗ Ruoyu Xie∗ Fahim Faisal∗ Antonios Anastasopoulos Department of Computer Science, George Mason University {malam21, rxie, ffaisal, antonis}@gmu.edu Abstract This report describes GMU’s sentiment analy- sis system for the SemEval-2023 shared task AfriSenti-SemEval. We participated in all three sub-tasks: Monolingual, Multilingual, and Zero-Shot. Our approach uses models ini- tialized with AfroXLMR-large, a pre-trained multilingual language model trained on African languages and fine-tuned correspondingly. We also introduce augmented training data along with original training data. Alongside fine- tuning, we perform phylogeny-based adapter- tuning to create several models and ensemble the best models for the final submission. Our system achieves the best F1-score on track 5: Amharic, with 6.2 points higher F1-score than the second-best performing system on this track. Overall, our system ranks 5th among the 10 systems participating in all 15 tracks. 1 Introduction With the increasing use of the internet and social media, the digital availability of various languages is rapidly expanding. This expansion opens av- enues for Natural Language Processing (NLP) ap- plications such as Sentiment Analysis (SA) and Machine Translation (MT). Nevertheless, despite African languages comprising 30% of around 6,000 living languages (Skutnabb-Kangas et al., 2003), most of them are not supported by modern lan- guage technologies, leading to an ever-widening gap in language technology access (Joshi et al., 2020; Blasi et al., 2022). Recently, SA has gained increasing attention, with its applications in various domains, such as public health, literature, and social sciences (Mo- hammad, 2022). Despite the growth in this area, most previous works do not include African lan- guages. This shared task focuses on SA on a *Joint contributions: MA performed model training, RX worked on data processing and paper writing, FF constructed the model. Ranking Team Name Avg. F1 1 BCAI-AIR3 69.77 2 UM6P 68.08 3 mitchelldehaven 67.85 4 tmn 65.84 5 GMNLP 65.76 6 UCAS 65.62 7 Masakhane-Afrisenti 63.58 9 PA14 63.09 9 DN 64.10 10 FUOYENLP 59.46 Table 1: We calculate the macro-average F1-score for the 10 systems (out of 47) participating in all 15 tracks in this Shared Task. Overall, our system ranks 5th. Twitter dataset in 14 African languages, includ- ing Hausa (ha), Yoruba (yo), Igbo (ig), Nigerian Pidgin (pcm), Amharic (am), Tigrinya (tg), Oromo (or), Swahili (sw), Xitsonga (ts), Algerian Arabic (dz), Kinyarwanda (kr), Twi (twi), Mozambican Portuguese (pt), and Moroccan Darija (ma). This paper presents a novel SA system that ef- fectively addresses the challenge of low-resource and multilingual sentiment classification for mul- tiple African languages. We leverage multilingual language models and propose data augmentation methods to increase the training data size. In addi- tion, we perform phylogeny-based adapter-tuning to create several models. These models are then ensembled to create the final model. 2 Related Work While SA is a popular research area in NLP, its application in African languages is relatively rare. …

Content truncated to 500 words