Improving the performance of personal name disambiguation using web directories

Intelligent Database Systems Lab

國立雲林科技大學National Yunlin University of Science and Technology

1

Improving the performance of personal name disambiguation using web directories

Quang Minh Vu, Atsuhiro Takasu, Jun AdachiIPM, 2008

Presented by Hung-Yi Cai2010/09/01


N.Y.U.S.T.

I. M.

2

Outlines

Motivation Objectives Methodology Experiments Conclusions Comments


N.Y.U.S.T.

I. M.Motivation

Searching for information about a person on the internet is an increasing requirement in information retrieval.

Search results returned from search engines for a personal name query often contain documents relevant to several people because a name is usually shared by several people.

Due to this name ambiguity problem, users have to manually investigate the result documents to filter out people in whom they have no interest.

3


N.Y.U.S.T.

I. M.Previous Studies

4

Name Method of Detail disadvantage

Co-reference->VSM(In news articles)

First, Extracting text relevant to a person in a document. Then, using the VSM to measure similarities between articles.

People in the web may have several appearances related to different events.

Second order context vector

Applying the log likelihood method together with singular value decomposition for co-occurrence information to calculate context vectorsof terms.

This method may not work well when dealing with people who are not famous.The building of context vectors difficult.

c-value/nc-value

Extracting key phrases related to people. Then, senting key phrases as queries tosearch engines and built key phrases’ contexts using snippets of the resulting documents.

It requires many query transactions to build contexts for key phrases.

Link information->ACDC

Using link information in web pages and the other uses the A/CDC algorithm to group together web pages having the same topic.

When we search for a person on the web, We may not know his or her social network in advance.

Pattern-MatchingDatabase(DBLP&Amzon)

Extracting personal profiles, or using databases, such as DBLP and Amazon, to extract authors’ names and research keywords.

The method of extracting personal profiles may not work well with web pages other than profile pages, while the method that uses a dictionary-like database cannot extract terms not listed in the database.

Natural language-processing

Extracting named entities in documents. Because web documents contain much noisy information, the extraction of named entities may not work well.


N.Y.U.S.T.

I. M.Objectives

Propose Similarity via Knowledge Base (SKB) that uses web directories to improve the disambiguating performance in Name Disambiguation System (NDS).

SKB can be divided into two components:─ Using web directories as a knowledge base to find common

contexts by TF-IDF in documents.─ Then, using the common contexts measure to determine

document similarities.

5


N.Y.U.S.T.

I. M.TF-IDF

6

Term weights are calculated using the terms’ occurrences in the document concerned and in a set of documents.─ Tf (t, doc) is the number of times term t appears in the

document doc.


N.Y.U.S.T.

I. M.Methodology

In SKB, using web directories to measures features of terms in a document.1) Measurement of term weights using a knowledge base

A knowledge base Modification of term weight in documents Modification of term weight in directories

2) Measurement of document similarities Find directories close in topic with the document Measure document similarities

7


N.Y.U.S.T.

I. M.Name Disambiguation System

The operational details are as follows:1) Preprocessing documents

2) Calculation of document similarities

3) Discrimination by reranking documents

8


N.Y.U.S.T.

I. M.Experiments

Step 1. Data Sets─ Documents of people─ Creation of pseudo namesake document sets and real

namesake document sets

9


N.Y.U.S.T.

I. M.Experiments

Step 2. Web directory structures

10


N.Y.U.S.T.

I. M.Experiments

Step 3. Baseline methods─ Comparing SKB with two conventional methods:

VSM： Calculating the weight of these terms by TF-IDF Building the feature vectors of documents

NER： Extracting the entity names in the documents by LingPipe

software Using these names to construct feature vectors of the documents

(the constituents of vectors were binary values)

11


N.Y.U.S.T.

I. M.Experiments

Step 4. Evaluation metrics─ We recorded the precision values at 11 recall points:

0%, 10%, ... ,90%, and 100% and denoted these as P(doci, 0%), P(doci, 10%), ... , P(doci, 90%) and P(doci, 100%), respectively.

12


N.Y.U.S.T.

I. M.

Step 5. Experimental results─ The overall performance for each method

In this experiment, we set the window size n = 50 and the number of representative directories k = 20. We set the frequency document ratio threshold for SKB2 r = 5.

Experiments

13


N.Y.U.S.T.

I. M.Experiments

Step 5. Experimental results─ Performance of SKB2 when varying the frequency ratio

threshold

14


N.Y.U.S.T.

I. M.Experiments

Step 5. Experimental results─ Performance of SKB systems when varying the window size

15


N.Y.U.S.T.

I. M.Experiments

Step 5. Experimental results─ Performance of SKBs when varying the number of

representative directories

16


N.Y.U.S.T.

I. M.Experiments

Step 5. Experimental results─ Performance for each method on real namesake document sets

17


N.Y.U.S.T.

I. M.

18

Conclusions

Disambiguation of people will be a trend in web search, and we propose a new method that uses web directories as a knowledge base to improve the disambiguation performance.

The experimental results showed a significant improvement with our system over the other methods, and we also verified the robustness of our methods experimentally with di erent ffweb directory structures and with di erent ffparameter values.


N.Y.U.S.T.

I. M.

19

Comments

Advantages─ Just requiring little preparation─ Broad range of people

Shortages─ Cost of computation is proportional─ Some mistake

Applications─ Information retrieval