This paper proposes AdaRNN, an adaptive recursive neural network for target-dependent Twitter sentiment classification. AdaRNN adaptively propagates the sentiments of words to the target based on context and syntactic relationships. It uses multiple composition functions and models adaptive sentiment propagation as distributions over these functions. The model automatically learns composition functions and how to select them from supervision, rather than using heuristics or hand-crafted rules. The paper also introduces a manually annotated dataset for target-dependent sentiment analysis. The paper discusses the challenges of target-dependent sentiment classification, where traditional methods fail to account for multiple targets in a single tweet. AdaRNN integrates target information with RNN to enhance deep learning models. It uses distributed representation to automatically learn features for sentiment classification. RNNs use recursive structures to achieve state-of-the-art results in sentiment analysis. Recursive neural models use semantic composition functions to handle complex compositionalities in sentiment analysis. The paper presents a framework that learns to propagate sentiments of words towards the target based on context and syntactic structure. It introduces a novel adaptive multi-compositionality layer in RNN, named AdaRNN. The model uses a softmax classifier to predict sentiment labels. The paper also presents experiments showing that AdaRNN outperforms baseline methods. It introduces a manually annotated dataset for target-dependent sentiment analysis and conducts extensive experiments on it. The results show that AdaRNN achieves better performance than baseline methods. The paper also discusses the effects of the hyper-parameter β in AdaRNN. Different values of β lead to different composition selection schemes. The paper concludes that AdaRNN improves baseline methods for target-dependent sentiment classification by using adaptive composition functions and syntactic tags to guide sentiment propagation. The model is more robust to parsing imprecision than hand-crafted rules.This paper proposes AdaRNN, an adaptive recursive neural network for target-dependent Twitter sentiment classification. AdaRNN adaptively propagates the sentiments of words to the target based on context and syntactic relationships. It uses multiple composition functions and models adaptive sentiment propagation as distributions over these functions. The model automatically learns composition functions and how to select them from supervision, rather than using heuristics or hand-crafted rules. The paper also introduces a manually annotated dataset for target-dependent sentiment analysis. The paper discusses the challenges of target-dependent sentiment classification, where traditional methods fail to account for multiple targets in a single tweet. AdaRNN integrates target information with RNN to enhance deep learning models. It uses distributed representation to automatically learn features for sentiment classification. RNNs use recursive structures to achieve state-of-the-art results in sentiment analysis. Recursive neural models use semantic composition functions to handle complex compositionalities in sentiment analysis. The paper presents a framework that learns to propagate sentiments of words towards the target based on context and syntactic structure. It introduces a novel adaptive multi-compositionality layer in RNN, named AdaRNN. The model uses a softmax classifier to predict sentiment labels. The paper also presents experiments showing that AdaRNN outperforms baseline methods. It introduces a manually annotated dataset for target-dependent sentiment analysis and conducts extensive experiments on it. The results show that AdaRNN achieves better performance than baseline methods. The paper also discusses the effects of the hyper-parameter β in AdaRNN. Different values of β lead to different composition selection schemes. The paper concludes that AdaRNN improves baseline methods for target-dependent sentiment classification by using adaptive composition functions and syntactic tags to guide sentiment propagation. The model is more robust to parsing imprecision than hand-crafted rules.