If Boston Power's somersault is helping robots exercise their bones and muscles, then the "drawing" of knowledge map is trying to "create" a working robot brain.
"At present, it is impossible for machines to understand human language." Sun Le, a researcher at the Institute of Software of Chinese Academy of Sciences and vice chairman of China Chinese Information Society, said. Whether it's Siri who can make you happy, Xiao Bing who can write poems, or Watson who can "feel the pulse", they don't really understand what they are doing and why they are doing it.
Let the machine learn to think, relying on "spectrum." This "spectrum" is called knowledge map, which is intended to build the knowledge generated in the human world into the machine world, and then form a knowledge base that can support brain-like reasoning.
In order to build a brand-new Industry-University-Research cooperation model of knowledge mapping in China, a seminar on knowledge mapping was held recently. Researchers from universities and industrial teams have worked together to establish a global knowledge map system and a world-leading artificial intelligence infrastructure.
Technical principle: transforming text into knowledge
"For the sentence" Yao Ming is from Shanghai ",it is just a string of characters stored in the machine. This string of characters is' live' in the human brain. " Sun Le said, for example. For example, when Yao Ming is mentioned, people will think of him as a former American professional basketball player, a "little giant" and a center, while "Shanghai" will make people think of the Oriental Pearl and a bustling city. But for machines, just saying "Yao Ming is from Shanghai" can't understand the meaning behind it like human beings. To understand a passage, a machine needs to know the background knowledge first.
So how to turn the text into knowledge?
"With the help of information extraction technology, people can extract knowledge from texts, which is the core technology of knowledge map construction." Sun Le said that it is popular to use the "triple" storage mode at present. A triple consists of two points and an edge. Points represent an entity or concept, while edges represent various semantic relationships between entities and concepts. A point can extend from many sides, forming various relationships. For example, Yao's point is related to his birthplace in Shanghai, playing for NBA and his height of 2.26 meters.
"If these relationships are perfect enough, the machine will have the foundation to understand the language." Sun Le said. So how to make the machine have such "understanding"?
"In 1960s, Marvin Minsky of Massachusetts Institute of Technology, a pioneer of artificial intelligence, used semantic relations between entities to express the semantics of questions and answers in a question-and-answer system project. Margaret Mastman of Cambridge Department of Linguistics used semantic networks to model world knowledge in 196 1, which can be regarded as the predecessor of knowledge map." Sun Le said.
Subsequently, Wordnet and Hownet in China also built the knowledge base by hand.
"This includes subjective knowledge, such as whether people like or dislike a product on social networking sites; Scene knowledge, such as what to do in a specific scene; Language knowledge, such as grammar of various languages; Common-sense knowledge, such as water, cats and dogs, can be pointed directly when taught, but computers are difficult to understand. " Sun Le explained that from these preliminary classifications, we can feel the vastness of knowledge, not to mention high-level scientific knowledge.
Construction mode: from manual labor to automatic extraction
"After 20 10, Wikipedia began to try crowdsourcing, and everyone can contribute knowledge." Sun Le said that this has greatly accelerated the accumulation of knowledge maps. Baidu Encyclopedia and Interactive Encyclopedia have also adopted similar knowledge collection methods, mobilizing the public to greatly shorten the time of "sand accumulation" and greatly improve the efficiency. Countless knowledge has poured in from all directions and gathered rapidly, just waiting for "building a tower".
Faced with such a large amount of data, or "text", the construction of knowledge map can naturally be no longer manual labor, "let the machine automatically extract structured knowledge and automatically generate' triples'." Sun Le said that academia and industry have developed different frameworks and systems, which can automatically or semi-automatically generate machine-readable knowledge from texts.
There is a vivid picture in Sun Le's demonstration courseware. If you eat a lot of paper, the computer will immediately turn it into "knowledge", but the fact is far from simple. There is no unified scheme for automatic extraction of structured data in different industries. In the introduction of "Baidu knowledge map", it is written as follows: the data submitted to the knowledge map is transformed into entity objects that follow the schema, and unified knowledge calculation such as data cleaning, alignment, fusion and association is carried out to complete the construction of the map. "However, we found that based on Wikipedia, the knowledge maps mined from structured and semi-structured data are still insufficient, so all the work at present focuses on how to extract knowledge from massive texts." Sun Le said, for example, Google's knowledge base and TAC-KBP evaluation sponsored by the National Institute of Standards and Technology are also promoting the technology of extracting knowledge from texts.
In the authoritative international evaluation of automatic construction of knowledge base, extracting knowledge from text is divided into four parts: entity discovery, relationship extraction, event extraction and emotion extraction. In the TAC-KBP Chinese evaluation organized by NIST in the United States, the joint team of Institute of Software of Chinese Academy of Sciences-sogou won the third place in comprehensive performance index, and the event extraction single index was 1 name.
"In this field, China can compete with the international level." Sun Le introduced that the Institute of Software of Chinese Academy of Sciences proposed entity acquisition algorithm based on co-guidance and relationship extraction algorithm based on multi-source knowledge supervision, which greatly reduced the modeling cost and improved the performance of text knowledge extraction tools.
The ultimate goal: to build all human knowledge.
According to the Old Testament, mankind built the Tower of Babel together, hoping to lead to heaven. Now, the human who created AI is building such a Babel to help artificial intelligence reach human intelligence.
Automatic practice makes the amount of knowledge begin to form a scale, reaching an order of magnitude that can support practical application. "But this change is far from reaching the level of human knowledge," Sun Le said. Besides, human knowledge has been constantly increasing, updating and changing dynamically, and understanding should be reflected in the "brain" of the machine with the times.
"Therefore, the knowledge map will not be a static state, but will form a cycle, which is also the concept of endless learning put forward by Carnegie Mellon University and other places in the United States." Sun Le said.
The data shows that at present, Google's knowledge map records more than 3.5 billion facts; Freebase has recorded more than 40 million entities, tens of thousands of attribute relationships and more than 2.4 billion facts. Baidu Encyclopedia contains100000 entries, and Lenovo search function is applied in Baidu search.
"There are also maps of expertise in specific fields such as medical fields and interpersonal relationships." Sun Le introduced that kinship describes the kinship between characters, including 104 entities, 26 relationships and 10800 facts; UMLS uses 135 entities, 49 relationships and 6800 facts to describe the relationship between Chinese medicine concepts in the medical field.
"This is a grand blueprint full of bright prospects." Sun Le said that the ultimate goal of knowledge map is to formalize and structure all human knowledge and use it to build a knowledge-based natural language understanding system.
Although the "system that truly understands language" that satisfies the industry is far from appearing, and the current "Tower of Babel" only stays at the basic level, related applications have shown broad prospects. For example, if you enter "frozen electron microscope" in Baidu Encyclopedia, stone will appear in the vertical bar on the right, and if you enter "coin", Wang Sicong and other related items will appear directly in the search term. It contains the machine's understanding of human intentions.