RESUMO
Data annotation in NLP is a costly and time-consuming task, traditionally handled by human experts who require extensive training to enhance the task-related background knowledge. Besides, labeling social media texts is particularly challenging due to their brevity, informality, creativity, and varying human perceptions regarding the sociocultural context of the world. With the emergence of GPT models and their proficiency in various NLP tasks, this study aims to establish a performance baseline for GPT-4 as a social media text annotator. To achieve this, we employ our own dataset of tweets, expertly labeled for stance detection with full inter-rater agreement among three annotators. We experiment with three techniques: Zero-shot, Few-shot, and Zero-shot with Chain-of-Thoughts to create prompts for the labeling task. We utilize four training sets constructed with different label sets, including human labels, to fine-tune transformer-based large language models and various combinations of traditional machine learning models with embeddings for stance classification. Finally, all fine-tuned models undergo evaluation using a common testing set with human-generated labels. We use the results from models trained on human labels as the benchmark to assess GPT-4's potential as an annotator across the three prompting techniques. Based on the experimental findings, GPT-4 achieves comparable results through the Few-shot and Zero-shot Chain-of-Thoughts prompting methods. However, none of these labeling techniques surpass the top three models fine-tuned on human labels. Moreover, we introduce the Zero-shot Chain-of-Thoughts as an effective strategy for aspect-based social media text labeling, which performs better than the standard Zero-shot and yields results similar to the high-performing yet expensive Few-shot approach.
Assuntos
Mídias Sociais , Humanos , Processamento de Linguagem Natural , Aprendizado de MáquinaRESUMO
BACKGROUND: Over the past years, homelessness has become a substantial issue around the globe. The largest social services organization in Thunder Bay, Ontario, Canada, has observed that a majority of the people experiencing homelessness in the city were from outside of the city or province. Thus, to improve programming and resource allocation for people experiencing homelessness in the city, including shelter use, it was important to investigate the trends associated with homelessness and migration. OBJECTIVE: This study aimed to address 3 research questions related to homelessness and migration in Thunder Bay: What factors predict whether a person who migrated to the city and is experiencing homelessness stays or leaves shelters? If an individual stays, how long are they likely to stay? What factors predict stay duration? METHODS: We collected the required data from 2 sources: a survey conducted with people experiencing homelessness at 3 homeless shelters in Thunder Bay and the database of a homeless information management system. The records of 110 migrants were used for the analysis. Two feature selection techniques were used to address the first and third research questions, and 8 machine learning models were used to address the second research question. In addition, data augmentation was performed to improve the size of the data set and to resolve the class imbalance problem. The area under the receiver operating characteristic curve value and cross-validation accuracy were used to measure the models' performances while avoiding possible model overfitting. RESULTS: Factors predicting an individual's stay duration included home or previous district, highest educational qualification, recent receipt of mental health support, migrating to visit family or friends, and finding employment upon arrival. For research question 2, among the classification models developed for predicting the stay duration of migrants, the random forest and gradient boosting tree models presented better results with area under the receiver operating characteristic curve values of 0.91 and 0.93, respectively. Finally, home district, band membership, status card, previous district, and recent support for drug and/or alcohol use were recognized as the factors predicting stay duration. CONCLUSIONS: Applying machine learning enables researchers to make predictions related to migrants' homelessness and investigate how various factors become determinants of the predictions. We hope that the findings of this study will aid future policy making and resource allocation to better serve people experiencing homelessness. However, further improvements in the data set size and interpretation of the identified factors in decision-making are required.
RESUMO
Amid ongoing health crisis, there is a growing necessity to discern possible signs of Wellness Dimensions (WD) manifested in self-narrated text. As the distribution of WD on social media data is intrinsically imbalanced, we experiment the generative NLP models for data augmentation to enable further improvement in the pre-screening task of classifying WD. To this end, we propose a simple yet effective data augmentation approach through prompt-based Generative NLP models, and evaluate the ROUGE scores and syntactic/semantic similarity among existing interpretations and augmented data. Our approach with ChatGPT model surpasses all the other methods and achieves improvement over baselines such as Easy-Data Augmentation and Backtranslation. Introducing data augmentation to generate more training samples and balanced dataset, results in the improved F-score and the Matthew's Correlation Coefficient for upto 13.11% and 15.95%, respectively.