The morphological complexity of the Malayalam language poses significant challenges for dependency parsing, demanding accurate syntactic and semantic analysis to advance natural language processing (NLP) for low-resource languages. This study introduces a dependency parsing approach that combines the Cross-Lingual Language Model with Roberta (XLM Roberta) as the shared encoder and a biaffine attention mechanism for parsing with a span-based predictor for SRL. XLM Roberta is a transformer-based multilingual model that produces high-dimensional contextual embeddings for Malayalam sentences to provide a robust syntactic and semantic analysis foundation. The biaffine attention mechanism is employed by the dependency parsing decoder to predict head-dependent relationships and assign syntactic dependency labels. The Span-Based Predictor employed for SRL assigns semantic roles to spans within sentences to effectively handle long-range dependencies common in complex languages like Malayalam. The dataset comprises a manually annotated Malayalam treebank, ensuring complete syntactic and semantic coverage. Parsing performance was evaluated using head detection accuracy, root token identification and the processing of complex sentence structures. Evaluation results indicate that integrating morphological features improves the Unlabeled Attachment Score (UAS) from 93.70% to 95.20% and the Labeled Attachment Score (LAS) from 91.45% to 93.10%. Furthermore, head detection accuracy, root token identification and complex sentence parsing demonstrate significant improvements, with respective scores increasing to 95.40%, 93.80% and 91.60%. By addressing major challenges in Malayalam dependency parsing, this study presents an efficient and scalable solution for language processing tasks. The proposed approach demonstrates significant potential for applications like machine translation, sentiment analysis and knowledge extraction, paving the way for future developments in NLP for low-resource and morphologically rich languages.