Distant supervision for relation extraction without labeled data

Distant supervision for relation extraction without labeled data remote supervision: Uses unlabeled corpus for relation extraction 1. Background:

Link extraction (whether an individual belongs to an organization or not)

Used in link Extraction3Methods:

A) Supervised Learning

Advantage: High Accuracy

Disadvantages: 1. manual tagging of gold corpus is expensive, which requires a lot of time and money, and the number of materials is limited, so a large amount of training data is not available; 2. fields are restricted, and tags are all in a specific corpus. The training system is restricted by that field.

B) unsupervised learning

Advantage: large-scale data can be used to extract a large number of relationships.

Disadvantage: It is often difficult to map extracted results to a specific knowledge base.

C) bootstrap Learning

There are often low accuracy issues.

D) remote supervision

Use freebase to obtain weekly labeled training data.

Features: Compared with supervised learning, the knowledge base is used to provide training data instead of manual tagging to obtain training data, which has no issue of fitting and Field Dependence. Compared with unsupervised learning, you do not need to resolve the problem of ing clustering results to relationships, and use large-scale training data to obtain rich features.

2. Method Introduction

Basic assumptions:If two entities are participants of a relationship, any sentence containing the two entities may express this relationship.

Training Phase

1. Use Net (Named Entity tagger) to mark persons organizations and locations;
2. Extract features from entity pairs in freebase and construct training data;
3. Train the multi-class logistic regression model.

Test phase:

1. Use Net (Named Entity tagger) to label persons organizations and locations
2. Each pair of entities appearing in a sentence is considered as a potential link instance and used as test data.
3. Use the trained model to classify object pairs.

 

3. Feature Selection 3.1. Vocabulary features:

A) word sequence between two entities;

B) part-of-speech tagging of these words;

C) The flag indicates the entity that appears first;

D) Left window in K size;

E) Right window in K size.

3.2. Syntax features:

A) The shortest dependency path between two entities;

B) The left and right windows of two entities.

3.3. Features of the named OBJECT Tag:

Name, place name, organization name, and others

4. Other notes

Link features to be thrown into multiple logistic regression models.

 

Negative Sample Construction: randomly selects entity pairs that are not in freebase (possibly incorrect)

 

Training and test data structure: the Relationship instance in freebase is used for training and the other half for testing. Data is allocated using Wikipedia data, with training and test data. During the test, only instance pairs that do not appear during training (not in freebase during training) are classified.

 

Test result Selection: the confidence level of all object pairs, and each object pair is assigned a classification result. Sort their confidence level and select Top N.

Distant supervision for relation extraction without labeled data