We propose a new method which can improve the information retrieval quality and efficiency in maintainability domain. Our method incorporates the state-of-the-art natural language generation models. The main process of our retrieval system is as follows: when a user needs to find a certain information in a maintenance manual, the manual is first loaded into our document parsing module to be decomposed into text paragraphs. Than the text paragraphs are processed by the information extraction module which uses a deep learning model named SQUASH to generate hierarchical question-answer pairs. At the same time, the user needs to input the retrieval question, which will be reformulated to several similar questions. These questions will be matched with the generated question-answer pairs. The final output is ranked matching results linked with their corresponding paragraph. Two specific cases are presented to better illustrate our method.
Our method outperforms the traditional retrieval method in several scenarios. For example, when there is too much similar expression, our system can target the desired information accurately. Our system also helps significantly when the query sender is not familiar with the term used in the manual, or the document is either complicated or not standardized (i.e. the information that should be available is not written in the commonly used form, or not in place where it should appear).