In the era of online shopping, the volume of product reviews for user products on e-commerce platforms is massively increasing on a daily basis. For any given user product, it consists of a flood of reviews and manually analysing each of these reviews to understand the important aspects or opinions associated with the products is difficult and time-consuming task. Furthermore, it becomes nearly impossible for the customer to make decision of buying the product or not. Thus, it becomes necessary to have an aspect-based summary generated from these user reviews, which can act as a guide for the interested buyer in decision-making. Recently, the use of Large Language Models (LLMs) has shown great potential for solving diverse Natural Language Processing (NLP) tasks, including the task of summarization. Our paper explores the use of various LLMs such as Llama3, GPT-4o, Gemma2, Mistral, Mixtral and Qwen2 on the publicly available domain-specific Amazon reviews dataset as a part of our experimentation work. Our study postulates an algorithm to accurately identify product aspects and the model’s ability to extract relevant information and generate concise summaries. Further, we analyzed the experimental results of each of these LLMs with summary evaluation metrics such as Rouge, Meteor, BERTScore F1 and GPT-4o to evaluate the quality of the generated aspect-based summary. Our study highlights the strengths and limitations of each of these LLMs, thereby giving valuable insights for guiding researchers in harnessing LLMs for generating aspect-based summaries of user products present on these online shopping platforms.

Sentiment Analysis plays a crucial role in understanding user opinions in various languages. The paper presents an experiment with a sentiment analysis model fine-tuned on Marathi sentences to classify sentiments into positive, negative, and neutral categories. The fine-tuned model shows high accuracy when tested on Konkani sentences, despite not being explicitly trained on Konkani data; since Marathi is a language very close to Konkani. This outcome highlights the effectiveness of Zero-shot learning, where the model generalizes well across linguistically similar languages. Evaluation metrics such as accuracy, balanced accuracy, negative accuracy, neutral accuracy, positive accuracy and confusion matrix scores were used to assess the performance, with Konkani sentences demonstrating superior results. These findings indicate that zero-shot sentiment analysis can be a powerful tool for sentiment classification in resource poor languages like Konkani, where labeled data is limited. The method can be used to generate datasets for resource-poor languages. Furthermore, this suggests that leveraging linguistically similar languages can help generate datasets for low-resource languages, enhancing sentiment analysis capabilities where labeled data is scarce. By utilizing related languages, zero-shot models can achieve meaningful performance without the need for extensive labeled data for the target language.