Relation Recognition

(1)

lecture of Internet-based IE

Relation Recognition

Fang Li

Dept. of Computer Science &

Engineering

(2)

lecture of Internet-based IE Technology

• Relation Representation

• Relation Identification

1.Knowledge Engineering Approach 2.Machine Learning Approach

--Supervised learning

--Semi-supervised learning -- Distant supervised learning -- Deep Learning

3. Knowledge graph

(3)

Relation Example:

(4)

Relation with/without Time

 Relations may be timeless or bound to time intervals. For example,

father(x,y) vs. boss(x,y)

 Time type is divided into temporal points and intervals

(5)

Relation and Event

 Events: a particular kind of simple or complex relation among entities

involving a change in relation state at the end of a time interval.

 Eg: Company-founding

Company: IBM

Location: New York Date: June 16,1911

Original-Name: Coputer- Tabulating-Recording Co.

Founding-year (IBM, 1911) Founding-location (IBM, New York)

(6)

Relation Examples

 Physical--Located PER--‐GPE

He was in Tennessee.

 Part--Whole-Subsidiary ORG--‐ORG

XYZ, the parent company of ABC.

 Person--Social--Family PER--‐PER

John’s wife Yoko!

 Org--AFF-Founder PER--‐ORG

Steve Jobs, co-founder of Apple.

(7)

Explicit and Implicit Relations

Explicit relations are expressed by certain surface linguistic forms:

 Prepositional phrase: The CEO of Microsoft…

 Prenominal modification: the American envoy…

 Possessive: Microsoft’s chief scientist…

 Nominalizations: Anan’s visit to Baghdad..

 Apposition: Tony Blair, Britain’s prime minister….

(8)

Explicit and Implicit Relations (cont.)

 A relation is implicit in the text if the text provides convincing evidence that the relation actually holds.

 Example:

Prime Minister Tony Blair attempted to convince the British Parliament of the necessity of intervening in Iraq.

Question: Was Tony Blair a British Prime Minister?

(9)

A conceptual view of entities and relations

 E’s are the

entities found in a sentence.

 R’s are the relations

between any two entities.

 mutually dependent

(10)

Example:

Mo Yan is the first Chinese writer to win the Nobel Prize in Literature, who was born in Shan Dong province.

(11)

Three Cases of Binary

Relation Extractions R(E

¹

,E

²

)

 For a given fixed pair of entities (E¹,E²), to find out the type of

relationship (R) that exists between the pair.

 Given relationship R and an entity name E¹, to extract the entities (E² ) with which E¹ has relationship R.

 Given a fixed relationship type R, to find all the entity pairs (E¹,E²).

(12)

Relation Extraction

 A harder task than entity extraction

 Relation extraction requires a

skillful combination of local and nonlocal noisy clues from

diverse syntactic and semantic structures in a sentence.

(13)

Preprocessing Steps for relation extraction

E.g. Haifa located 53 miles from Tel Aviv will host ICML in 2010.  located

1) Named entity identification:

<LOC>Haifa</LOC> located 53 miles from

<LOC>Tel Aviv</LOC> will host ICML in 2010.

2) POS tagging:

Haifa/NNP located/VBN 53/CD miles/NNS from/IN Tel/NNP Aviv/NNP will/MD host/VB ICML/NNP

in/IN 2010/CD

(14)

Preprocessing Steps of Relation Extraction (cont.)

3)Syntactic Parse Tree

4) dependency Graph

Haifa located 53 miles from Tel Aviv will host ICML in 2010

(15)

Methods of Relation Recognition

1. Pattern-based methods:

 hand made patterns

 learning based on seeded pattern.

2. Supervised method

3. Semi-supervised method

4. Distant-supervised method 5. Deep learning method

6. …

(16)

Hearst’s Patterns

Some patterns extracted from the sentence or between the two entities: X is an instance of Y

(17)

Learning Patterns

-- based on seeds

 <Mark Twain, Elmira> Seed tuple

“MarkTwain is buried in Elmira, NY.”

X is buried in Y --pattern1 induced

“The grave of Mark Twain is in Elmira”

The grave of X is in Y –pattern2 induced

“Elmira is Mark Twain’s final resting place”

Y is X’s final resting place. –pattern3 induced

 Use those patterns to grep for new facts.

(18)

Problems with patterns

 Hand-built

Plus: High-precision, tailored to specific domains

Minus: low-recall, huge labor

 Learning based on seeds

Plus: high-recall, less human labor Minus: noise, low-precision

(19)

Supervised machine learning methods (overview)

1. Choose a set of relations to extract 2. Find and label data

 Label the named entities and the relations between these entities.

 Break into training, development and test sets

3. Train a classifier on the training set

4. Find all pairs of named entities (usually in the same sentence)

5. Use the classifier to identify the relation

(20)

For example: to identify the

employee relation (Org, Per)

Input:

Acme Inc. hired Mr Smith as their new CEO, replacing Mr Bloggs.

Org: Acme Inc.

Per: Smith and Bloggs

(Acme Inc., Smith) true (Acme Inc., Bloggs) false Model

(21)

Train the Model

 Extract features:

1. Features similar to those used in the entity recognition: capitalized, suffix and so on.

2. Conjunctions of the features from the two entities: spouse_of needs person type of both entities.

 Choose models: many models.

lecture of Internet-based IE lecture of Internet-based IE

(22)

Word Features

Acme Inc (mention 1). hired Mr Smith (mention 2) as their new CEO, replacing Mr Bloggs.

 Headwords of M1 and M2, and combination Inc. Smith Inc.Smith

 Bag of words and bigrams in M1 and M2

{Acme, Inc, hired, Mr., Smith, Acme Inc, Mr. Smith,…}

 Words or bigrams in particular positions left and right of M1 and M2

M1: +1 hired M2: +1 as, -1 hired

 Bag of words or bigrams between M1 and M2 {hired}

(23)

Named entity type and mention level

Features

 Named-entity types (ORG, PER, etc) M1: ORG M2: person

 Concatenation of the two entity type ORG-PERSON

 Entity level of M1 and M2 {name, nominal, pronoun}

M1: name M2: name

(24)

Parse Features

 Base syntactic chunk sequence from one to the other

VP

 Constituent path through the tree from one to the other

NP VP NP

 Dependency path

Acme Inc. hired Mr Smith

(25)

Other Features : Gazeteers and trigger words

 Personal relative trigger list from Wordnet: parent, wife, husband,…

 Country name list

 Wikipedial

(26)

Acme Inc (mention 1). hired Mr Smith

(mention 2) as their new CEO, replacing Mr Bloggs.

 Entity-based features M1 type: ORG

M1 head: Inc M2 type: PERS M2 head: Smith

 Word-based features

Between entity bag of words: {hired}

Words before M1: none Words after M2: as

 Syntactic features

Constituent path: NP VP NP

Basic syntactic chunk path: VP Typed-dependency path:

Acme Inc.<- subj hired obj Mr.Smith

Feature summary

(27)

From Dan Jurafsky 为什么需要两

个实体的 Entity head 特征呢？

(28)

Features  Feature

representation of the model

Acme Inc (mention 1). hired Mr Smith (mention 2) as their new CEO, replacing Mr Bloggs.

 Feature between Mention1 and

mention2: word sequences or number of words between them.

 Morphologic feature of mention 1（形态特征）:

ACME INC, A.C.M.E Inc, Acme Inc, acme inc

 Combination feature （顺序关系）:

(company name, person name), (person name, company name)

(29)

Classifiers for supervised methods

(ref. chapter 5 of textbook)

 Choose models:

1. MaxEnt(maximum entropy model) 2. NB(Naïve Bayes)

3. SVM(support vector machines) 4. …

 Train it on the training set, turn on the development set, test on the test set.

(30)

Relationship Extraction using Support Vector Machine (SVM)

•

Support vector machine (SVM) is recognized as one of the best classification algorithm over various applications and domains.

•

SVM is a method that finds a function that discriminates between two classes.

Positive example

Negative example

(31)

Support Vector Machine (SVM)

 When classifying natural language data, it is not always possible to linearly separate the data  map them into a feature space where they are linearly separable.

Kernel function

Positive example

Negative example

(32)

SVMLight: an open software

 Install an SVM package such as SVMlight (http://svmlight.joachims.org/)

 Transfer your training data format in order to be matched.

 Use training command for SVMlight.

SVM Ref:

http://nlp.stanford.edu/IR-book/html/htmledition/support-

vector-machines-the-linearly-separable-case-1.html#svm-sv- classifier

(33)

A Guide to SVM

 Transform data to the format of an SVM package.

 Conduct simple scaling on the data.

 Choose a kernel for SVM.

 Use cross-validation to the best parameter.

 Train the whole training set.

 Test

(34)

Data Preprocessing

 SVM requires that each data instance is represented as a vector of real

numbers.

 Use m numbers to represent a m- category attribute. For example a

three-category attribute such as (red, green, blue) can be represented as

(0,0,1), (0,1,0), and (1,0,0).

(35)

Scaling

 Some attribute may be a value, such as the length of a sentence.

 Scaling before using SVM  [0,1] or [- 1,1], for example, [-10,10] to [-1,1]

 How ?

X= (x-min)/(max-min)

Using the same scaling factors for training and test sets, obtain better result.

(36)

Choose a kernel

 Linear kernel when the number of features is very large.

 RBF kernel can handle nonlinear problem.

(37)

Cross-validation & grid-search

 In v-fold cross-validation, first divide the training set into v subsets of equal size.

Sequentially one subset is tested using the classier trained on the remaining v-1

subsets.

 Each instance of the whole training set is predicted once so the cross-validation

accuracy is the percentage of data which are correctly classified.

 Grid-search parameter using cross- validation. lecture of Internet-based IE

(38)

Problems of Supervised methods

 High precision with enough hand-labeled training data.

 Labeling is expensive.

 Supervised models can not generalize well to different genres.

(39)

Comparison of Classification Models

 Test corpus:

Reuters-21578 Data Set

 21578 documents

 118 categories

 An article can be in more than one category

 Learn 118 binary category distinctions

 Common categories (#train, #test)

• Earn (2877, 1087)

• Acquisitions (1650, 179)

• Money-fx (538, 179)

• Grain (433, 149)

• Crude (389, 189)

• Trade (369,119)

• Interest (347, 131)

• Ship (197, 89)

• Wheat (212, 71)

• Corn (182, 56)

(40)

Technology 40

(41)

Semi-supervised method

Relation Bootstrapping

Gather a set of seeds

Iterate:

1.Find sentence with these seeds 2.Look at the context between or

around the seeds to define a pattern 3.Use the pattern for more examples

(42)

Bootstrapping from seed entity

pairs to learn relations

(43)

Confidence Value for Bootstrpping

 Given a document collection D, a current set of tuples T, and a proposed pattern P, two factors need to be considered:

 Hits: the set of tuples in T that p matches while looking in D.

 Finds: The total set of tuples that p finds in D

A corpus D:

ABCDEF BDCDFE CDEFG HHECDE

Tuple set T:

BCDE ECDE AUD Conf(CD)=2/4 X log (4)=30%

(44)

(45)

 Patterns:

[person][was assigned│was selected │was appointed as][position]

 New examples:

Example: Extract Person name and position title

Search Engine Keywords: Wang Ning + vice Mayor

(46)

Summarization

 What is relations recognition? Three cases

 How to identify relations?

 Pattern-based methods

 Supervised methods

 Semi-supervised methods

(47)

References



Text book chapter 5 Supervised Classification



Sunita Sarawagi. Information Extraction Foundations and Trends in Databases vol.1,No.3 2007 261-377.

 Jun Zhu,et al. StatSnowball:a statistical apprpach to extracting entity relationships In Proceedings of

WWW 2009, Madrid.

回家作业课程网站下载，两篇文章选一篇阅读，完成截止日10月28日24点以前

 Mintz,Bills,Snow,Jurafsky.Distant supervision for relation extraction without labeled data. ACL 2009

 TransE

(48)

Distant supervision method

“Distant supervision for relation extraction without labeled data”

 What means “distant supervision”?

 What are the advantages of the method?

 What are the disadvantages of the method?

Relation Recognition