Analysis of Alibaba cloud Tianchi competition questions (in-depth learning) -- Reading Notes 1 -- competition question 1

Alicloud Tianchi competition Title Analysis ( Deep learning )– Reading notes 1

[x] It means that you encounter knowledge you don't understand , Will be in 【 Knowledge supplement 】 Give specific explanations .

---

---

Preface

If machine learning mainly solves the application problem of big data , Then the emergence of deep learning , Then begin to solve the problem of machine perception .

---

Tips ： The following is the main body of this article , The following cases can be used for reference

Question one Ruijin Hospital MMC Artificial Intelligence Aided Construction of knowledge map

This competition question will focus on the construction of knowledge map [1] Important technology in — Entity recognition , Explain in detail .

background

There are many patients with diabetes , The cause is complicated , The symptoms are varied , It brings great difficulties to the diagnosis and treatment of diabetes .

Mission

The first stage ： Construction of entity annotation based on diabetes clinical guidelines and research papers ;
The second stage , Construction of inter entity relationship based on diabetes clinical guidelines and research papers .

Introduce the knowledge map （Knowledge Graph）

The development history

The initial purpose is to improve the search quality of search engines , Understand the semantic information behind user queries .（things not strings）
Semantic Web （Semantic Network, the doer of the action in a sentence , event , Beneficiary , object ）–> noumenon （Ontology, Identify commonly recognized terms in the field ）–> The semantic web （Semantic Web, Describe the relationship between resources and data in the World Wide Web ）

The way of expression

How to express knowledge ？ Knowledge map is mainly in the form of graph （ Nodes and edges ） To organize knowledge structurally .
Usually use triples ： Source node , edge , The target node , Namely subject , Predicate , Object of SPO A triple , It is the smallest unit constituting the knowledge map .
Usually use RDF（Resource Description Framework, Resource Description Framework ） To organize triples , It's usually used XML Grammar said , There will be RDFs appear , And there are OWL Language [2] Further expanded RDFs.
It means learning [3] It can greatly improve the related tasks of the knowledge map .

Construction mode

How to construct knowledge map ？ The core of building a knowledge map is to build a description of knowledge map SPO A triple .
Entity recognition ： solve S and O Identification problems of , It mainly determines the boundary and classification of entities in the text .
Relationship extraction ： It is mainly to identify the relationship between given entities , Usually given by the specialty in the field .
Knowledge fusion ; The main purpose is to align triples of several different knowledge maps , To get a larger knowledge map , Mainly the alignment of entities . Unsupervised entity alignment ： By extracting features from entities , Then cluster according to the feature similarity to achieve alignment . The similarity measure means editing distance [4]、Jaccard coefficient [5]、 Cosine similarity [6] etc. .
Anaphora digestion ： It is a key task in information extraction . Such as multiple expressions of an entity , And exist in the form of pronouns , We should accurately understand the semantic information behind it . Common methods are Mention Pair（ All demonstrative pronouns are combined in pairs , Then carry out positive and negative classification ） and Mention Ranking（ Combine a demonstrative pronoun with other words , adopt softmax To determine which candidate word has the greatest probability of this demonstrative pronoun ）.

Knowledge reasoning

How to carry out knowledge reasoning ？ The main task is to infer unknown knowledge from known knowledge （ Complete the knowledge map ） And identify existing errors （ Quality verification ）. Methods include reasoning based on symbolic logic and reasoning based on representation learning .

Knowledge supplement

[1] Knowledge map construction link ： Knowledge modeling （ Agreed mode , Define the entity 、 Relationship 、 event ）、 Knowledge acquisition （ Entity recognition 、 Relationship recognition 、 Event identification ）、 Knowledge fusion （ Merge in the same form ）、 Knowledge storage 、 Knowledge of computing （ Reasoning , Mining correlation ）、 Knowledge application .
[2] OWL Language ： A definition and instantiation web The language of ontology , For computer processing Web Information , Than RDF Have stronger machine interpretation ability .
[3] It means learning ： In the field of deep learning , Representation refers to the parameters through the model , In what form 、 How to represent the input observation sample of the model X. Learning refers to learning about observed samples X A valid representation of , It is a collection of technologies that learn a feature , Transform the raw data into something that can be effectively developed by machine learning .
[4] Edit the distance for two strings （ For example, English words ） Quantitative measurement of the degree of difference , The way is to see how many times it needs to be handled at least （ Insert 、 Delete 、 Rewrite characters ） To change one string into another .
[5]Jaccard Coefficients are used to compare the similarities and differences between finite sample sets .Jaccard The larger the value of the coefficient , The higher the similarity of samples . Defined as ： Given two sets A,B,Jaccard The coefficient is defined as A And B The size of the intersection with A And B The ratio of the size of the Union .
[6] Cosine similarity , Also known as cosine similarity , The similarity between two vectors is evaluated by calculating the cosine value of the angle between them .