The escalating volume and frequency of social media complaints necessitate robust automated complaint analysis techniques. Much of the existing body of research in this area has been devoted to two primary aspects: identifying complaint-specific content amidst other noncomplaint communications, and predicting the severity of a complaint, which involves classifying complaints into different severity levels based on the anticipated resolution from the complainant's perspective. These automated analysis tools equip companies with the means to effectively manage complaints and generate suitable responses. In our study, we present a unified generative approach for complaint detection, transforming the multitask learning problem into a text-to-text generation task. As part of our training strategy, we adopt the Seq2Path training paradigm that conceptualizes the outcome as a tree structure as opposed to a traditional sequence. This innovative approach tackles the drawbacks of conventional sequences, such as the lack of order among the outputs, yielding a more coherent and structured output. Our model's effectiveness is assessed against the benchmark Complaints dataset, highlighting its superior performance across diverse evaluation metrics when compared with state-of-the-art models and other baselines 1 .