Performance Analysis of Serial and Parallel Computing Techniques for Natural Language Processing Tasks

This paper explores the efficiency and scalability of serial and parallel computing techniques applied to common Natural Language Processing (NLP) tasks, including keyword extraction, sentiment analysis, named entity recognition, and text summarization. Using a combination of Python libraries such as spaCy, transformers, Dask, and multiprocessing, we evaluate the execution time and memory usage for these tasks on varying dataset sizes and numbers of processing threads. The experiments are conducted on a system with an integrated AMD Radeon GPU, where all computations run on the CPU due to the lack of CUDA support. The challenges and limitations of utilizing GPU acceleration for NLP tasks on such hardware are also discussed. Our findings reveal that while serial execution is straightforward, parallel computing can significantly reduce execution time for large datasets when appropriately configured. However, increasing the number of threads beyond the system's core count can lead to performance degradation due to overhead from context switching, memory contention, and the Global Interpreter Lock (GIL). Additionally, the limitations of integrated GPUs in supporting popular NLP frameworks are analyzed. Recommendations for optimal thread usage, multiprocessing strategies, and profiling techniques are provided to improve the performance of NLP workloads. The study highlights the importance of understanding system architecture and task characteristics when designing efficient NLP pipelines. These insights can benefit researchers and practitioners aiming to optimize NLP workflows in resource-constrained environments.