毕业生知识库

随着信息技术的飞速发展，尤其是智能个人助理、智能客服、智能教育等应用的出现，人们已经不仅仅将互联网作为一个资料查询的工具，而且期望它能够智能地替人们完成信息筛选、归纳总结、模拟演绎等任务，甚至直接解决人们的问题。

如何从大规模网络知识中总结出真实世界的规律和规则，以及如何运用这些规则得到目标问题的答案，已经成为亟待解决的问题，而对知识推理技术的研究正是为了促进这些问题的解决。但是将传统人工智能中的推理方法迁移并运用到基于大规模网络知识库的智能应用时，会遇到两方面的挑战：首先是大规模知识库对推理算法本身提出的新要求，包括需要高效的规则自动挖掘方法以克服专家规则集合不能有效区分噪声的问题，需要降低规则挖掘和推理算法的计算复杂度，需要更好地处理稀疏关系和稀疏规则；

其次是自然语言表示的前端与知识推理所需要的假设形式之间存在沟壑，如何进行有效的转换也是一项重要的任务。

 

本文针对降低规则挖掘和推理算法计算复杂度以及克服自然语言前端向知识推理假设转化的问题展开研究，主要的研究成果包括：

 

1、针对逻辑规则挖掘依赖子图枚举或路径搜索等具有指数型计算复杂度操作的问题，提出基于实例化网络抽样的方法，并针对其中关键步骤——随机游走的效率问题，提出基于目标导向的随机游走算法。

该方法通过随机游走在子图空间中进行均匀采样，

等比例地缩小一条规则正负实例计数以及不同规则之间的实例计数，代替了具有NP难计算复杂度的子图枚举和路径搜索方法，

并利用启发式方法设置状态转移概率，令实例化网络抽样方法的复杂度与知识库规模和逻辑规则长度无关。同时，该方法通过势函数捕捉目标与下一状态之间的联系，令随机游走在每一次跳转时都受到推理目标的指引，

减小无效路径和噪声规则的召回，

从而提高逻辑规则挖掘的效率和准确性。实验结果表明该方法可以在缩短规则挖掘时间的同时保证逻辑规则的质量。

 

2、针对推理过程中由于候选集过大而造成的推理时间长和存在无法区分噪声的问题，提出基于表示学习的候选预选择方法，并针对规则实例化计算复杂度高的问题，提出利用规则向量快速判断规则可用性的方法。

该方法利用了表示学习模型计算速度快和前N准确率高的特点，快速地生成一个小规模高召回的候选子集，并将表示学习方法计算的候选得分作为先验指导后续推理过程中的随机游走以进一步过滤噪声。同时，该方法将逻辑规则表示为分布式语义空间中的向量计算符，去评估两个实体间出现规则实例的可能性，由此代替了复杂的图搜索或抽样的方式，提高了推理效率。

实验结果表明，表示学习是一种高效且准确的候选预选择方法，而将候选得分作为逻辑推理方法的先验也提升了推理准确率，并且逻辑规则向量化的方式进一步地提升了推理的效率。

 

3、针对难以将具有复杂嵌套语义的自然语言表述向适宜的推理假设转化的问题，提出基于虚拟假设解耦的复杂问题推理方法。该方法首先将每一个问题和候选答案对表示成知识库中一个高层三元组，称为虚拟假设，然后在分布式语义空间中捕获虚拟假设与原问题之间，以及虚拟假设与逻辑规则之间的关联，并将这些关联作为证据加入到概率推理模型的目标函数中，从而实现了从自然语言问题向知识库上推理假设的转化。论文还提出了一个专门用于评估推理方法的问题数据集，涉及中英两种语言且包含七个自然与历史学科的真实考试题，实验结果表明基于虚拟假设解耦的推理方法解决了自然语言向推理假设转化的问题，将知识推理有效地应用于深度问答，提升了问答系统的性能。

With the development of information technology, especially with the advent of the intelligent applications, e.g., intelligent personal assistant, intelligent service agent and intelligent education platform, people do not any more treat Internet only as a tool of searching information, but also expect it can filter information, conclude, summarize, simulate and deduce instead of humans, and directly solve people's problems. How to generate laws and rules of the real world from large-scale web knowledge, and how to apply these rules to find answers for target problems have been problems demanding prompt solution. Inference aims to solve these problems. However, it will face two problems when applying traditional inference methods to current intelligent applications with large-scale knowledge bases. First, large-scale knowledge raises new requirements for inference, including requiring high-efficiency methods of automatically rules mining to overcome the problem of that expert rules cannot effectively detect noise, reducing computation complexity of mining rules and performing inference, and handling sparse relation types or sparse rules. Second, there is a gap between the front end in natural language and feasible inference hypotheses, so how to transform is also an important task.

 

In this dissertation, we focus on reducing computation complexity of mining rules and performing inference, and transforming natural language into feasible inference hypotheses. The main achievements are as follows.

 

1. To reduce the computation complexity of enumerating subgraph and searching path during mining formulas, we propose to mine formulas by grounding network sampling. This method performs uniform sampling by random walks on the subgraph space, and sets state transition probabilities by a heuristic method, so the complexity of this method is irrelevant to the scale of knowledge base and the length of formulas. In order to further improve the efficiency of random walk, we propose a goal-directed random walk algorithm, which directs walks by the goal at each transition by introducing a potential function. This method improves the efficiency and effectiveness of mining formulas which may support to infer the goal. The experimental result shows this method can short the running time of mining logic formulas and guarantee the quality of formulas.

 

2. To address the problem of that the candidate set is too large, which leads to a very long inference time and being easy to introduce noise, we propose a method of pre-selection based on representation learning. This method takes advantage of the character of representation learning, i.e., the computation is fast and the accuracy at Top-N is high, and this method can generate a small-scale candidate set with a high coverage of correct answer. Moreover, to further filter noise, we insert the scores calculated by representation learning into the subsequent random walk process. To simplify the process of instancing logic formulas, we propose to employ the distributed representation of a logic formula to fast estimate the usability of the formula. This method represents a logic formula as a vector operator and uses it to estimate whether the formula can be used for two entities. This method takes place of the complex searching or sampling and improves the efficiency of inference.

 

3. To address the problem of that it is difficult to transform natural language descriptions with complex recursive structure into feasible inference hypotheses, we propose to decouple KB-based inference by a virtual hypothesis. This method defines a high level triplet in the knowledge base, which is called a virtual hypothesis. This method also captures associated information between the virtual hypothesis and the original question and correlations between the virtual hypothesis and logic formulas, and directly treats these associations as evidences in the probabilistic inference model. This method achieves that natural language questions can be transformed into inference hypotheses on the knowledge base. After that, we create a question set which contains real-world examination questions in seven subjects with two languages (English and Chinese), where questions that can be answered only by simple retrieval methods have been filtered out. The experimental result shows this method is a promising solution to transform nature language into inference hypotheses.