Aspect-based sentiment analysis (ABSA) is vital for text comprehension which benefits applications across various domains. This field involves the two main sub-tasks including aspect extraction and sentiment classification. Existing methods to tackle this problem normally address only one sub-task or utilize topic models that may result in overlapping concepts. Moreover, such algorithms often rely on extensive labeled data and external language resources, making their application costly and time-consuming in new domains and especially for resource-poor languages like Urdu. The lack of aspect mining studies in Urdu literature further exacerbates the inapplicability of existing methods for Urdu language. The primary challenge lies in the preprocessing of data to ensure its suitability for language comprehension by the model, as well as the availability of appropriate pre-trained models, domain embeddings, and tools. This paper implements an ABSA model (CITATION) for unlabeled Urdu tweets with minimal user guidance, utilizing a small set of seed words for each aspect and sentiment class. The model first learns sentiment and aspect joint topic embeddings in the word embedding space with regularization to encourage topic distinctiveness. Afterwards, it employs deep neural models for pre-training with embedding-based predictions and self-training on unlabeled data. Furthermore, we optimize the model for improved performance by substituting the CNN with the BiLSTM classifier for sentence-level sentiment and aspect classification. Our optimized model achieves significant improvements over baselines in aspect and sentiment classification for Urdu tweets with accuracy of 64.8% and 72.8% respectively, demonstrating its effectiveness in generating joint topics and addressing existing limitations in Urdu ABSA.