According to an embodiment, a document analysis device includes one or more hardware processors configured to function as a sentence extraction unit, an analysis unit, a neural network unit. The analysis unit generates, for each of sentences, initial element information representing an initial value of relevance to each of predetermined attribute items. The neural network unit receives sentence information and outputs execution result of a main task on the target document, for each of the sentences. The neural network unit includes an attention unit and a main task execution. The attention receives the sentence information and the initial element information, calculates an attention weight and outputs attention information according to the attention weight, for each of the sentences. The main task execution unit executes the main task based on the attention information for each of the sentences.