| In recent years,although the development of the military command information systems has gained remarkable achievements,in the future war there will be more complex situational understanding,rapid decision-making and other issues.These are the bottlenecks for promoting intelligent systems.Faced with this situation,and taking into account the "human-computer" characteristics of information-based military command and decision-making,human-computer interaction is used in the military command information system to assist commanders to improve the ability to obtain key information and improve efficiencies of decision-making.It is an important research direction to break the bottleneck of perception intelligence of a system.The form of human-computer interaction,mainly based on voice interaction,can greatly improve the commander’s abilities about situational awareness and decisionmaking.This entire interaction process involves many modules in natural language processing such as speech recognition,speech synthesis,and dialogue systems.At present,the existing speech recognition and speech synthesis theories and technologies are relatively mature,and related industrial products have been put on the market.The core component of the dialogue system,question answering,is one of main challenges of the frontier research in the field of natural language processing,and plays a vital role in assisting decision-making and improving efficiencies of interactions.Question Answering refers to a series of operations such as analyzing the question and retrieving passages based on the user’s question in the form of natural language,and returning an accurate answer to the user.There are several shortcomings in traditional grammatical feature-based methods to solve the question answering problem: 1)relying heavily on syntactic analysis,which requires a lot of labor and material resources;2)paying attention to keyword matching and ignoring similarities at the semantic level;3)poor flexibility and there is some information loss when extracting answers.Deep learning methods can automatically generate a vector representation with semantic information from the original text,avoiding the above-mentioned shortcomings of traditional methods,and can effectively deal with the problem of question answering.Based on the frontier research on deep learning,this paper mainly researches on algorithms about constructing word vector models and evaluating candidate answers in the military field’s question answering tasks based on unstructured documents.since current mainstream static word embedding models cannot effectively deal with the problem of "polysemy",this paper proposed an question answering method based on improved dynamic word embedding,QA-IDWE.The algorithm uses an improved dynamic word embedding model to dynamically construct word vectors according to the context,which effectively solves the "polysemy" problem and makes the constructed word vectors more semantic,but the answers predicted by this algorithm are still misplaced,missing parts,or redundancy.To solve this problem,this paper proposed an question answering method based on reranking candidate answers,QA-RCA.The algorithm is divided into two parts: the candidate answer extraction module and the candidate answer set reranking module.The former is used to extract answers from the original text,which can generate a candidate answer set;the latter combines the background knowledge in the given original text to calculate the matching relationship between the candidate answer and the question,so as to improve accuracies of predicting answer.Experimental results on a typical data set show that the question answering methods based on deep learning proposed in this paper can improve the overall performance in the task of question answering by constructing a more accurate word vector model and a more reasonable answer generation model. |
