This paper introduces a deep memory network for aspect-level sentiment classification. Unlike traditional methods such as SVM and LSTM, this approach explicitly captures the importance of each context word when determining the sentiment polarity of an aspect. The model uses multiple computational layers, each of which is a neural attention model over an external memory. These layers calculate the importance of each context word and generate continuous text representations. The final layer's representation is used as the feature for sentiment classification. The model is data-driven, computationally efficient, and does not rely on syntactic parsing or sentiment lexicons. The deep memory network is inspired by memory networks with attention mechanisms. It consists of multiple computational layers, each containing an attention layer and a linear layer. The attention layer assigns weights to context words based on their relevance to the aspect. The model uses these weights to generate a text representation that is then used for sentiment classification. The model is trained end-to-end using gradient descent, with the loss function being the cross-entropy error of sentiment classification. The model is tested on laptop and restaurant datasets from SemEval 2014. Experimental results show that the approach performs comparably to a top feature-based SVM system and outperforms LSTM and attention-based LSTM models in terms of classification accuracy and speed. The model with nine layers is 15 times faster than LSTM with a CPU implementation. The model also demonstrates that using multiple computational layers improves performance. The paper also explores the use of location information in the attention model. Four strategies are tested for integrating location information, with Model 2 showing the best performance. The model with location information outperforms the content-based model in terms of accuracy. The model is able to correctly identify the sentiment polarity of aspects by considering both content and location information. The paper concludes that the deep memory network is a simple and fast approach for aspect-level sentiment classification. It outperforms traditional methods in terms of accuracy and speed, and demonstrates the effectiveness of using multiple computational layers and location information in the attention model. The model is also shown to be effective in capturing both content and location information, leading to better context weight and text representation. Future work includes incorporating sentence structure-like parsing results into the deep memory network.This paper introduces a deep memory network for aspect-level sentiment classification. Unlike traditional methods such as SVM and LSTM, this approach explicitly captures the importance of each context word when determining the sentiment polarity of an aspect. The model uses multiple computational layers, each of which is a neural attention model over an external memory. These layers calculate the importance of each context word and generate continuous text representations. The final layer's representation is used as the feature for sentiment classification. The model is data-driven, computationally efficient, and does not rely on syntactic parsing or sentiment lexicons. The deep memory network is inspired by memory networks with attention mechanisms. It consists of multiple computational layers, each containing an attention layer and a linear layer. The attention layer assigns weights to context words based on their relevance to the aspect. The model uses these weights to generate a text representation that is then used for sentiment classification. The model is trained end-to-end using gradient descent, with the loss function being the cross-entropy error of sentiment classification. The model is tested on laptop and restaurant datasets from SemEval 2014. Experimental results show that the approach performs comparably to a top feature-based SVM system and outperforms LSTM and attention-based LSTM models in terms of classification accuracy and speed. The model with nine layers is 15 times faster than LSTM with a CPU implementation. The model also demonstrates that using multiple computational layers improves performance. The paper also explores the use of location information in the attention model. Four strategies are tested for integrating location information, with Model 2 showing the best performance. The model with location information outperforms the content-based model in terms of accuracy. The model is able to correctly identify the sentiment polarity of aspects by considering both content and location information. The paper concludes that the deep memory network is a simple and fast approach for aspect-level sentiment classification. It outperforms traditional methods in terms of accuracy and speed, and demonstrates the effectiveness of using multiple computational layers and location information in the attention model. The model is also shown to be effective in capturing both content and location information, leading to better context weight and text representation. Future work includes incorporating sentence structure-like parsing results into the deep memory network.