Effective string processing and matching for author entity

Effective String Processing and Matching for Author

Disambiguation

Source:Journal of Machine Learning Research’14Speaker:LIN,CI-JIE

Outline

IntroductionMethodExperimentConclusion

Outline

IntroductionMethodExperimentConclusion

Introduction

Track 2 of KDD Cup 2013 aims at determining duplicated authors in a data set from Microsoft Academic Search

Track 2 in KDD Cup 2013 is a task of name disambiguation

Introduction

1. Author.csv2. Paper.csv3. PaperAuthor.csv4. Conference.csv 5. Journal.csv

Outline

IntroductionMethodExperimentConclusion

Overview of Data sets

1. Alleviation2. Unusual Name3. Inconsistent Information4. Typo5. Incomplete Name6. Empty Entry7. Missing Value8. Nickname9. Wrong matching between authors and papers10.Non-English characters

Main Strategies

The first strategy is that we identify duplicates mainly based on string matching

The second strategy is that if an author in Author.csv has no publication records in PaperAuthor.csv, then we assume that this author has no duplicates

The third strategy is to classify an author as Chinese or non-Chinese before any string matching

Implementation

Two implementations of the framework were finished by two groups within the team

The Framework

1. Chinese-or-not2. Cleaning3. Selection4. Identification5. Splitting6. Linking

The First Implementation1. Chinese-or-not

The First Implementation

2. Cleaning Split two consecutive uppercase characters

replace “CJ" with “C J.” Remove English honorics (e.g., “Mr." and “Dr.") Transform uppercase to lowercase Remove apostrophes and replace punctuation. For

example, “o'relly" becomes “orelly.“

The First Implementation

2. Cleaning Replace European alphabets with similar English

alphabets Replace common English nicknames

replace “bill” with “william."

The First Implementation

3. Selection build a dictionary of (key, value) pairs. Each key is a set

of words, while each value is a set of authors containing the key

The First Implementation

4. Identification Matching Functions

1. Two names share the same set of words “ Chih Jen Lin“ and “ Lin Chih Jen.„

2. A shortened name " Ch. J. Lin" and " Chih Jen Lin.“

3. A partially shortened name "C. J. Lin" and "C. J. Lint.“

4. Alias dry-run procedure

“ C J Lin “ and “ Chih Jen Lion “ are loosely identical, while “ Chih Jen Lin“ and “ Chen Ju Lin “ are not.

The First Implementation

5. Splitting Each author name is a full name with two words. Neither author name is a partially shortened

name of the other. " kazuo kobayashi" and "kunikazu kobayashi"

The Second Implementation

1. Chinese-or-not2. Cleaning3. Selection4. Identification

The Second Implementation

5. Splitting Common extended names(CEN) Longest common extended name(LCEN)

The Second Implementation

6. Linking previous stages group duplicated names rather

than identifiers. However, the competition task is to group

duplicated identifiers

Ensemble

the two implementations detect different sets of duplicates,an ensemble of their results may improve the performance

The filter considers that two authors have a similar background Affiliation、 field of study

Typo Correction

handle typos in author names of Author.csvTwo author names are considered as

duplicates iftheir word sets are the same after treating each

typo the same as after its correction,andtheir affiliations share at least one common word

Outline

IntroductionMethodExperimentConclusion

Experiment

Experiment

Outline

IntroductionMethodExperimentConclusion

Conclusion

try best to keep all information and delay the modification on the data set because the provided data set is noisy and incomplete

an important advantage of using rule-based approaches is that we can easily trace the change of results after a rule is added