人工智能近年来发展迅速，以自然语言理解为目标的认知智能技术研究成为大家广泛关注的课题。为了让机器具备对自然语言的深入理解及推理能力，关键在于我们需要研究如何实现机器对语言的建模学习。本文从自然语言理解领域一个基础的句子自动答题任务出发，探索研究具备高精准和灵活的可扩展性的自动答题技术。自动答题旨在从给定的词语或短语集合中选择最佳的候选答案填在目标句子的合适位置，使目标句子具有正确的语法和完整的语义。现阶段的自动答题技术，仍然以传统的词语特征和语言模型等方法为代表，缺乏对各类信息的有效利用，答题系统效果较低。与此同时，近年来以深度学习为代表的技术在自然语言理解领域广泛应用，但真正将其高效应用于句子自动答题任务的工作较少，且无法在统一的深度学习框架下实现对词法、句法等各类信息的合理融合学习，方法的扩展性很差。 本文正是从句子自动答题领域的技术研究现状入手，提出开展英语考试自动答题技术研究，并围绕多维度语义分析、深度语义建模、基于句法的深度语义建模以及深度语义信息融合等多个方面开展研究工作，具体包括： 首先，研究基于多维度语义分析自动答题方法。在自动答题领域被广泛使用的传统语言模型和隐语义分析方法，只能建模短距离的单词依赖，无法对长距离依赖以及语法问题进行精准建模。针对上述问题，本文相应提出了自动固定搭配挖掘和动词时态预测的方法，在此基础上，设计多维度语义分析方法，该方法相比传统方法效果提升明显，有效提升了答题系统的问题解决能力。 其次，研究基于深度学习语义建模答题方法。针对传统机器学习中存在的数据表达稀疏问题，本文将词嵌入表达引入到答题任务上。在此基础上，采用最新的深度学习框架，提出了基于排序学习的自动答题框架。进一步地，为了缓解词嵌入存在的语义不精确问题，本文创新性提出了对立语义的词嵌入增强模型，除了在公开的GRE对立语义数据集上取得最优效果，还在统一的深度学习技术框架中，成功应用于句子答题任务中，有效提升了答题效果。 再次，研究结合句法结构信息的深度语义答题方法。针对传统的循环神经网络结构的句子语义建模方法无法刻画句子语法结构等信息，在实际系统构建中往往存在信息利用单一，语义建模不够精准的问题。近年来虽有面向如何对句法结构进行建模使用的研究工作，但都表现出效率低、扩展性差等问题。本文提出了基于序列化句法结构的句子语义建模方法，有效解决了句法信息的使用问题，并在答题任务大幅度提升了系统性能。该方法为后续的融合模型提供了重要支撑。 最后，研究基于深度语义特征融合方法。该方法针对英语试题统一句子层面建模存在细节丢失的问题，提出了基于位置隐含状态信息的深度排序模型，同时结合本文前述提出的多种方法，构建形成一个统一的多源信息融合模型，并在实际自动答题任务中得到了最优的系统效果。 关键词：自然语言处理 深度学习 循环神经网络 卷积神经网络 长短时记忆网络 自动答题 词嵌入 句子语义表达 句法分析 语言模型

 Artificial intelligence has developed rapidly in recent years, and the research on cognitive intelligence technology with the goal of natural language understanding has attracted extensive attention. In order to give the machine a deep understanding and reasoning ability of natural language, the key is to study how to model the language of different level, such as word, sentence, document. This thesis starts from a basic automatic sentence answering task in the field of natural language understanding, and explores the automatic sentence answering technology with high precision and flexible scalability. The task of sentence automatic answering is to select the best candidate answer from a given set of words or phrases to fill in the appropriate position of the target sentence, so that the target sentence has the correct grammar and complete semantics. At present, the representative automatic sentence answering technology includes traditional word features and language model based methods. It lacks effective use of various types of information, and the sentence answering system has a poor accuracy. At the same time, in recent years, the technology represented by deep learning has been widely used in the field of natural language understanding, but the work of effectively applying it to the sentence automatic answering task is less, and it is impossible to fuse various features, such as lexical features, syntactic features, etc., and make reasonable integration learning under the unified deep learning framework. The scalability of existing methods is very poor. This thesis starts with the analysis of the technical research status in the field of automatic sentence answering, and proposes the research of automatic answering technology for English test. The thesis conducts research work in four aspects: multi-dimensional semantic analysis, deep semantic modeling, deep sentence embedding based on syntax, and deep semantic feature fusion. The specific description of each work is as follows: Firstly, the sentence automatic answering method based on multi-dimensional semantic analysis is studied. Traditional language models and latent semantic analysis methods, which are widely used in this field, can only model short-term word contexts, and cannot accurately model long-distance dependences and grammatical problems. Aiming at the above problems, this paper proposes a method of automatic regular collocation mining and verb tense prediction. We proposed a model that fuse multi-dimensional semantic features. Compared with the traditional method, the method significantly improves the performance of sentence automatic answering. Secondly, the method of sentence answering questions based on deep semantic representation is studied. Aiming at the problem of high-dimensional and sparse representation in traditional machine learning, this paper introduces word embedding which represent semantics in a low dimensional continuous vector space into the task of sentence automatic answering. An automatic sentence answering framework based on deep learning to rank is proposed. Furthermore, in order to alleviate the semantic inaccuracy of word embedding, this paper innovatively proposes a word embedding enhancement model for contrast meaning representation. This model gets the state of art performance on the public GRE "most opposite question" dataset. When we integrate the enhanced word embedding into a deep sentence embedding framework, it makes significant improvement in automatic sentence answering task. Thirdly, this thesis studies the deep semantic answering method based on syntactic structure information. The traditional sentence embedding methods based the existing deep learning models, such as recurrent neural network and convolutional neural network, cannot describe the sentence grammatic structure and other information. In recent years, although there are research work on how to model the syntax structure, they all show problems such as low efficiency and poor scalability. This paper proposes a sentence embedding method based on serialized syntax structure to effectively solve the problem of using syntactic information. And in the automatic sentence answering task, this approach greatly improves system performance and provides important support for subsequent fusion models. Finally, a method based on deep semantic feature fusion is studied. This method is motivated by the incapability of some semantic details in unified sentence embedding. A deep ranking model based on hidden state information around the position of answer is proposed. Combined with the various methods proposed above, a unified multi-source information fusion model is constructed. The optimal performance is obtained in the automatic sentence answering task of English test questions. Key Words: Natural Language Processing, Deep learning, Recurrent Neural Network, Convolution Neural Network, Long Short-Term Memory network, Automatic Sentence Answering, Word Embedding, Sentence Embedding, Syntactic Analysis, Language Model