Why Knowledge Graph can be applied to Q&A systems

Knowledge Graphs and Q&A Systems

Abstract: Domain Knowledge Graphs (KM) are the infrastructure for the next generation of search engines, Q&A systems and other intelligent applications. In this article, we will introduce the basics of knowledge graphs and their applications in natural language processing (mainly Q&A systems).

1. Introduction

Knowledge graph is the infrastructure for next-generation search engines, Q&A systems, and other intelligent applications. "This article will introduce the foundation of knowledge graph. This article will introduce the basics of knowledge graph, and its application in natural language processing (mainly Q&A systems).

2. Knowledge Graph Concepts

Knowledge Graph, a data structure based on directed graph, consists of nodes (points) and directed edges (directed edges), each node in the graph is called an Entity, and the edges represent the logical relationships between entities (Relationships). The edges represent the logical relationships between entities.

As an example, here is a knowledge graph that simply describes a tourist attraction:

Now, let's explain why "Knowledge graphs are the infrastructure for next-generation search engines, question-and-answer systems, and other intelligent applications. If we look at the intelligent system as a brain, then the knowledge graph is a knowledge base in the brain, which enables the machine to analyze and think about the problem from the perspective of "relationship". The above figure is an example of simple knowledge that can be obtained from the knowledge graph, such as "the altitude of Mount Tai is 1545 meters" and "Hengshan and Hengshan have the same pronunciation".

3. Representation of Knowledge Graph

Knowledge Graph can be represented by ternary (entity-1, relation, entity-2), each record describes a fact, for example: (Wuyue, one of the five mountains, Taishan) represents "Taishan is one of the five mountains". A fact. It should be noted that, if the relation? is determined, then the position of entity-1 and entity-2? can not be reversed, because a ternary group describes a directed edge (fact); entity does not have to be a real-life specific things, but can also be the value of an attribute of the thing, when the relationship is the attribute.

We use triples to store knowledge graphs, and there is one more issue to consider, which is Entity Recognition and Entity Disambiguation. For example, the entity "apple" may refer to a fruit apple, or it may refer to an iphone, and then we need to do some processing of the knowledge graph to modify the search strategy.

In the process of building the knowledge graph, if ambiguity is found, then the corresponding parent node is added, and the parent node is used to eliminate the ambiguity when filtering the word "apple". Here again, it is emphasized that the knowledge graph is only the infrastructure. Later, we will illustrate the application value of the knowledge graph through a few examples; and then introduce the Knowledge Reasoning (Knowledge Reasoning) technology, that is, how to interact with the user to teach the intelligent system to improve the knowledge graph.

4. Application of Knowledge Graph

Traditional search engines simply filter the target web pages according to the keywords entered by the user, and then give a bunch of web links. Knowledge graph applications, in addition to giving links to the corresponding web pages, try to provide some more intelligent answers. For example, if a user enters "taj mahal" in Bing Search, the following results will be provided:

There are synonyms for Taj Mahal, travel information, geographic location, and the Seven Wonders of the Ancient World, etc., which can better explore the user's intent, rather than the rigidity of traditional search engines that require users to filter the information on their own, one by one. The knowledge graph technology has a great commercial value.

For example, I searched Baidu and typed in "table tennis" and got the following results (I just wanted to search for "Zhang Jike" but forgot his name):

At the same time, the use of knowledge graphs enables search engines to gain a certain degree of reasoning power. The search engine has the ability to do some reasoning. For example, if you type "Liang Qichao's son's wife" into Baidu search, the traditional search engine simply matches web pages, making it difficult to truly understand the user's intent, let alone answer the question. However, knowledge mapping can make the problem easier, we first get Liang Qichao's son is Liang Sicheng from the knowledge base, and then get Liang Sicheng's wife is Lin Weiyin.

This enhances the interaction between the search engine and the user, gradually turning it into an intelligent Q&A system.

5. Knowledge Reasoning Technology

Here again, we emphasize that the knowledge graph is only the infrastructure, because it is really simple, and there is no high technology, we just want to store the knowledge in such a form in order to learn from simple knowledge to a higher level of knowledge. For example, the knowledge graph only contains the following information:

Then the intelligent quiz system cannot answer questions such as "What is the relationship between Kang Xi and Qianlong?". Knowledge reasoning questions, this can be used to simply add artificial rules: "father + father → grandfather" to update the knowledge graph, or directly from user interaction to learn the rules, of course, user interaction will need to be on the statistical knowledge, can not say that there are users to answer the "nationality of the same relation "The first thing you need to do is to make sure that all the grandfathers and grandchildren are of the same nationality.

Next we look at a more complex example, in the Baidu search engine enter "pregnant women can eat lychee?" Get the following results:

Then the intelligent Q&A system can return "59% of the netizens think you can eat, 28% of the netizens think you can not eat, 13% of the netizens think you can not eat". If the user continues to ask, "What is the reason for not being able to eat it? , then return to the "can not eat" users to answer. Here's an example of how statistical machine learning techniques can be used for knowledge inference.

We analyze, these "netizen answer" some from "baby tree", some from "have questions and answers", some from "parenting". "Parenting", we can use the crawler to crawl these sites to the question and its answer, and then do clustering of questions, first build the following knowledge graph:

Specific questions and answers between the logical relationship between the edge of the beginning of the blank, we then classify the semantic edge on the classification technology, to fill in the blank edge of the complete, but of course, may also be misclassified! For example, the first answer in the Baidu example, "eat carefully" is categorized as "can eat". Finally, these edges can be counted to answer similar questions, and the corpus can be used to do knowledge inference, learn knowledge that is not in the knowledge base, and improve the knowledge graph.

Reference

Knowledge graph - the knowledge base in the machine's brain

Applications of knowledge graph

Knowledge graph for mobile search