WKU_NLP at SemEval-2023 Task 9: Translation Augmented Multilingual Tweet Intimacy Analysis

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1525–1530 July 13-14, 2023 ©2023 Association for Computational Linguistics WKU_NLP at SemEval-2023 Task 9: Translation Augmented Multilingual Tweet Intimacy Analysis Qinyuan Zheng Wenzhou-Kean University zhengqin@kean.edu Abstract This paper presents a system designed for the SemEval 2023 Task 9: Multilingual Tweet In- timacy Analysis. This system is composed of a pretrained multilingual masked language model as a text encoder and a neural network as a regression model. To enhance model per- formance in low-resource scenarios, the system employs data augmentation techniques lever- aging neural machine translation models. Fur- thermore, I demonstrate the system can be fur- ther improved through the ensemble of top- performing models in each language. This system ranks 4th in languages unseen in the training data and 16th in languages seen in the training data. The code and data are accessible vias the following link: https://github. com/Cloudy0219/Multilingual. 1 Introduction Intimacy is a fundamental aspect of interpersonal relationships within various social settings. Lan- guages inherently encode rich social information pertaining to intimacy, offering opportunities to ex- plore and analyze intimacy through computational linguistics perspectives. Assessing intimacy in a language not only provides insights into the com- plexities of human communication but also serves as a test for computational models’ ability to com- prehend human social interactions and the under- lying social norms related to intimacy. Assessing intimacy in a language not only provides insights into the complexities of human communication but also serves as a test for computational models’ abil- ity to comprehend human social interactions and the underlying social norms related to intimacy. Social media platforms, such as Twitter, present a treasure trove of diverse interactions with varying degrees of intimacy, making them an ideal source for investigating linguistic intimacy patterns. Pei and Jurgens (2020) introduced a benchmark to ex- amine intimacy interactions on Twitter, comprising Twitter Train Translation Models Translated Data Tweet Regression Model Intimacy Score Text Encoder Figure 1: This figure illustrates the architecture of the system. The left part demonstrates the workflow to data augmentation based on the translation model. The right side demonstrates the workflow for intimacy analysis based on the text encoder and regression model. English tweets with human-labeled intimacy scores. Pei et al. (2022) further expanded this benchmark to a multilingual setting, covering ten languages, fur- ther increasing the scope and relevance of intimacy analysis across diverse linguistic communities. Despite its significance, intimacy analysis re- mains a challenging task due to the limited data available and the inherent difficulties for models to interpret and understand human intimacy. This challenge is further intensified in multilingual con- texts, as the expression of intimacy varies across languages and cultures. In response, I propose a novel regression model that leverages recent ad- vancements in language modeling and neural net- works to capture the semantics of intimacy in mul- tiple languages effectively. To address the issues of limited resources and language differences, I also incorporate data augmentation and model en- 1525 sembling techniques to enhance …

Content truncated to 500 words