Mining Wiki Resources for Multilingual Named Entity Recognition

1
Mining Wiki Resources for Multilingual Named
Entity Recognition

• Alexander E. Richman Patrick Schone

Department of Defense ACL 2008
Reporter Chia-Ying Lee Advisor Prof.
Hsin-Hsi Chen
2
Introduction

• Using the multilingual Wikipedia to automatically
  create an annotated corpus of text in any given
  language.
• Languages French, Ukrainian, Spanish, Polish,
  Russian, and Portuguese.
• Do not use of any non-English linguistic
  resources outside of the Wikimedia domain and any
  semantic resources such as WordNet or POS tagger.
• Use an internally modified variant of BBN's
  IdentiFinder (Bikel et al., 1999), specifically
  modified to emphasize fast text processing,
  called PhoenixIDF.

2
3
Related Work

• Toral and Muñoz (2006) used Wikipedia to create
  lists of named entities.
• Rely on WordNet, and need a manual supervision
  step
• Kazama and Torisawa (2007) used Wikipedia to
  building entity dictionaries.
• Rely on POS tagger
• Cucerzan (2007) used Wikipedia primarily for
  Named Entity Disambiguation, following the path
  of Bunescu and Pasca (2006)
• Using Category, but specific to English

4
Wikipedia

• Multilingual, collaborative encyclopedia on the
  Web which is freely available
• As of October 2007, there were over 2 million
  articles in English, and 30 languages with at
  least 50,000 articles and another 40 with at
  least 10,000 articles.

4
5
Wikipedia - feature

• Article links, links from one article to another
  of the same language.
• Category links, links from an article to special
  Category pages.
• Interwiki links, links from an article to a
  presumably equivalent, article in another
  language.
• Redirect pages, short pages which often provide
  equivalent names for an entity
• Disambiguation pages, a page with little content
  that links to multiple similarly named articles.
• Example http//en.wikipedia.org/wiki/FBI

5
6
Training Data Generation

1. Initial Set-up
2. English Language Categorization
3. Multilingual Categorization
4. The Full System

6
7
Initial Set-up

• ACE Named Entity types
• PERSON, GPE (Geo-Political Entities),
  ORGANIZATION, VEHICLE, WEAPON, LOCATION,
  FACILITY, DATE, TIME, MONEY, and PERCENT.
• MUC tags like ltENAMEX TYPEGPEgtPlace
  Namelt/ENAMEXgt
• Process
• Identifies words and phrases that might represent
  entities.
• Uses category links and/or interwiki links to
  associate that phrase with an English language
  phrase or set of Categories.
• Determines the appropriate type of the English
  language data and assumes that the original
  phrase is of the same type.

8
English Language Categorization(1)

• Wiki Useful Category gt Key Category Phrase
• gt Disambiguation Pages? gt Wiktionary
• Useful Category
• CategoryLiving People PERSON
• CategoryCities in NorwayGPE
• Useless Category
• Category1912 Establishments which includes
  articles on Fenway Park (a facility), the
  Republic of China (a GPE), and the Better
  Business Bureau (an organization).

9
English Language Categorization(2)
10
Multilingual Categorization

• Not all articles have English equivalent, but
  many of the most useful categories have English
  equivalents.
• French CatégorieCommune des Côtes-d'Armor,
  CatégorieVille portuaire de France,
  CatégoriePort de plaisance, and
  CatégorieStation balnéaire.
• English Category Communes of Côtes-d'Armor,
  UNKNOWN, CategoryMarinas, and
  CategorySeaside resorts

11
The Full System

• The first pass uses the explicit article links
  within the text.
• We then search an associated English language
  article, if available, for additional
  information.
• A second pass checks for multi-word phrases that
  exist as titles of Wikipedia articles.
• We look for certain types of person and
  organization instances.
• We perform additional processing for alphabetic
  or space-separated languages, including a third
  pass looking for single word Wikipedia titles.
• We use regular expressions to locate additional
  entities such as numeric dates.

12
Evaluation All

• Wiki test set
• Three human annotated newswire test sets
  Spanish, French and Ukrainian.

F-score Spanish French Ukrainian Polish Portuguese Russian
ALL .846 .844 .807 .859 .804 .802
DATE .925 .910 .848 .891 .861 .822
GPE .877 .868 .887 .916 .826 .867
ORG .701 .718 .657 .785 .706 .712
PERSON .821 .823 .690 .836 .802 .751
12
13
Evaluation Spanish (1)

• Spanish is a substantial, well-developed
  Wikipedia, consisting of more than 290,000
  articles at October 2007.
• Newswire 25,000 words from the ACE 2007 test
  set, manually modified extended MUC-style
  standards.
• Wiki test set 335,000 words.

14
Evaluation Spanish (2)

• Either Wikipedia is relatively poor in
  Organizations or that PhoenixIDF underperforms
  when identifying Organizations relative to other
  categories or a combination.
• Traditional Training trained PhoenixIDF on ACE
  2007 data converted to MUC-style tag.

15
Evaluation French

• French is one of the largest Wikipedias,
  containing more than 570,000 articles at October
  2007.
• Newswire 25,000 words from Agence France Presse
• Wiki test set 920,000 words.
• Similar to Spanish.

15
16
Evaluation Ukrainian (1)

• Ukrainian is a medium-sized Wikipedia with 74,000
  articles at October 2007.
• The typical article is shorter and less
  well-linked to other articles than in the French
  or Spanish versions.
• Newswire approximately 25,000 words from various
  online news sites covering primarily political
  topics.
• Wiki test set 395,000 words.
• Traditional Training trained PhoenixIDF Newswire
  data

16
17
Evaluation Ukrainian (2)

• The Ukrainian newswire contained a much higher
  proportion of organizations than the French or
  Spanish versions.
• The Ukrainian language Wikipedia contains very
  few articles on organizations relative to other
  types

17
18
Conclusion

• Wikipedia can create a NER system with
  performance comparable to one developed
  human-annotated Newswire, while not requiring any
  linguistic expertise.
• This level of performance can likely be obtained
  currently in 20-40 languages.
• Wikipedia-derived system could be used as a
  supplement to other systems for many more
  languages.
• An automatically generated entity dictionary
  embedded in our system .

18
19
Future Work

• Automatically generate the list of key words and
  phrases for useful English language categories.
• The authors also believe performance could be
  improved by using higher order non-English
  categories and better disambiguation.
• Lists of organizations might be particularly
  useful, and List of pages are common in many
  languages.

19
20
Thank you!
20