Automatic Extractive Summarization using GAN Boosted by DistilBERT Word Embedding and Transductive Learning

Abstract

Text summarization is crucial in diverse fields such as engineering and healthcare, greatly enhancing time and cost efficiency. This study introduces an innovative extractive text summarization approach utilizing a Generative Adversarial Network (GAN), Transductive Long Short-Term Memory (TLSTM), and DistilBERT word embedding. DistilBERT, a streamlined BERT variant, offers significant size reduction (approximately 40%), while maintaining 97% of language comprehension capabilities and achieving a 60% speed increase. These benefits are realized through knowledge distillation during pre-training. Our methodology uses GANs, consisting of the generator and discriminator networks, built primarily using TLSTM -an expert at decoding temporal nuances in timeseries prediction. For more effective model fitting, transductive learning is employed, assigning higher weights to samples nearer to the test point. The generator evaluates the probability of each sentence for inclusion in the summary, and the discriminator critically examines the generated summary. This reciprocal relationship fosters a dynamic iterative process, generating toptier summaries. To train the discriminator efficiently, a unique loss function is proposed, incorporating multiple factors such as the generator's output, actual document summaries, and artificially created summaries. This strategy motivates the generator to experiment with diverse sentence combinations, generating summaries that meet high-quality and coherence standards. Our model's effectiveness was tested on the widely accepted CNN/Daily Mail dataset, a benchmark for summarization tasks. According to the ROUGE metric, our experiments demonstrate that our model outperforms existing models in terms of summarization quality and efficiency.

A Predictive Sales System Based on Deep Learning

“…A great asset of LSTM [15], is its capacity to unravel intricate temporal structures and seize the fluid dynamics of systems that undergo time-driven fluctuations. By dissecting the inherent motifs and inclinations embedded within the data, LSTM architectures can reveal interconnections that may not be easily distinguished using established statistical or machine learning methodologies.…”