Abstract Aspect sentiment classification has a dependency over the aspect term extraction. The majority of the existing studies tackle these two problems independently, i.e., while performing aspect sentiment classification, it is assumed that the aspect terms are pre-identified. However, such assumptions are neither practical nor appropriate. In this paper, we address these impractical limitations and propose a multi-task learning framework for the identification and classification of aspect terms in a unified model. At first, the proposed approach employs a BiLSTM followed by a self-attention mechanism to identify the aspect terms in a given sentence. Subsequently, the architecture utilizes a CNN framework to predict the sentiments of the identified aspect terms. We evaluate our proposed approach for the three benchmark datasets across two languages, i.e., English and Hindi. Experimental results suggest that the proposed multi-task model achieves competitive performance with reduced complexity (i.e., a single model for the two tasks compared to two separate models for each task) for both the languages.