Introduction to Information Extraction Chia-Hui Chang Dept. of Computer Science and Information Engineering, National Central University, Taiwan chia@csie.ncu.edu.tw
Problem Definition The output template of the IE task Information Extraction (IE) is to identify relevant information from documents, pulling information from a variety of sources and aggregates it into a homogeneous form. The output template of the IE task Several fields (slots) Several instances of a field
Difficulties of IE tasks depends on … Text type From Wall Street Journal articles, or email message, to HTML documents. Domain From financial news, or tourist information, to various language. Scenario
Various IE Tasks Free-text IE: Semi-structured IE: For MUC (Message Understanding Conference) E.g. terrorist activities, corporate joint ventures Semi-structured IE: E.g.: meta-search engines, shopping agents, Bio-integration system
Types of IE from MUC Named Entity recognition (NE) Finds and classifies names, places, etc. Coreference Resolution (CO) Identifies identity relations between entities in texts. Template Element construction (TE) Adds descriptive information to NE results. Scenario Template production (ST) Fits TE results into specified event scenarios.
Name Entity Recognition http://www.cs.nyu.edu/cs/faculty/grishman/NEtask20.book_3.html
NE Recognition (Cont.) Spanish: 93% Japanese: 92% Chinese: 84.51%
Coreference Resolution Coreference resolution (CO) involves identifying identity relations between entities in texts. For example, in Alas, poor Yorick, I knew him well. Tie “Yorick" with “him“. The Sheffield system scored 51% recall and 71% precision. http://www.cs.nyu.edu/cs/faculty/grishman/COtask21.book_4.html
Template Element Production Adds description with named entities Sheffield system scores 71%
Scenario Template Extraction STs are the prototypical outputs of IE systems They tie together TE entities into event and relation descriptions. Performance for Sheffield: 49% http://www.cs.nyu.edu/cs/ faculty/grishman/ IEtask15.book_2.html
Example The operational domains that user interests are centered around are drug enforcement, money laundering, organized crime, terrorism, …. 1. Input: texts dealing with drug enforcement, money laundering, organized crime, terrorism, and legislation; 2. NE: recognizes entities in those texts and assigns them to one of a number of categories drawn from the set of entities of interest (person, company, . . . ); 3. TE: associates certain types of descriptive information with these entities, e.g. the location of companies; 4. ST: identifies a set (relatively small to begin with) of events of interest by tying entities together into event relations.
Example Text
Output Example (NE, TE)
Output (STs)
Another IE Example Corporate Management Changes Purpose which positions in which organizations are changing hands? who is leaving a position and where the person is going to? who is appointed to a position and where the person is coming from? the locations and types of the organizations involved in the succession events; the names and titles of the persons involved in the succession events http://www.cs.umanitoba.ca/~lindek/ie-ex.htm
Input Text President Clinton nominated John Rollwagen, the chairman and CEO of Cray Research Inc., as the No. 2 Commerce Department official. Mr. Rollwagen said he wants to push the Clinton administration to aggressively confront U.S. trading partners such as Japan to open their markets, particularly for high-tech industries. In a letter sent throughout the Eagan, Minn.-based company on Friday, Mr. Rollwagen warned: "Whether we like it or not, our country is in an economic war; and we are at a key turning point in that war." ...... Cray said it has appointed John F. Carlson, its president and chief operating officer, to succeed him. ......
Extraction Result Corporate Management Database Person Organization Position Transition John Rollwagen Cray Research Inc. chairman out CEO John F. Carlson in Organization Database Name Location Alias Type Cray Research Inc. Eagan, Minn. Cray COMPANY Commerce Department GOVERNMENT
MUC Data Set for MET2 http://www.itl.nist.gov/iaui/894.02/related_projects/muc/met2/met2package.tar.gz MUC3&4 http://www.itl.nist.gov/iaui/894.02/related_projects/muc/muc_data/muc34.tar.gz MUC6&7 from LDC http://www.ldc.upenn.edu/ MUC-6: http://www.cs.nyu.edu/cs/faculty/grishman/muc6.html MUC-7 http://www.itl.nist.gov/iaui/894.02/related_projects/muc/ proceedings/muc_7_toc.html
Summary Evaluation Design Methodology Precision= Recall= Natural Language Processing Machine Learning # of correctly extracted fields # of extracted fields # of correctly extracted fields # of fields to be extracted
IE from Semi-structured Documents Output Template: k-tuple Multiple instances of a field Missing data
Various IE Tasks for Semi-structured Documents Multiple-record page extraction One-record (singular) page extraction
Multiple-record page extraction
One-record (singular) page extraction
Summary Evaluation Design Methodology Precision= Recall= Machine Learning Pattern Mining # of correctly extracted records # of extracted records # of correctly extracted records # of records to be extracted
News Group IE Example: Computer-Related Jobs
Output Template Between free-text IE and semi-structured IE [CaliffRapier 99]
Annotated Training Examples Most systems require annotated training examples (answer keys) AutoSlog, Rapier, SRV, WIEN, Softmealy, Stalker Very few systems require unannotated training examples AutoSlog-TS, IEPAD, OLERA
The Type of Extraction Rule Delimiter-based Rule WIEN, Stalker Content-based Rule Context-based Rule Rapier, AutoSlog, SRV, IEPAD
Background Knowledge For Rule Generalization Example Implicit or Explicit Example Specified format for date, email, etc. Special feature for color, location, etc.
Conclusion Define the IE problem Specify the input: training example with annotation, or without annotation Depict the extraction rule Use necessary background knowledge
References *H. Cunningham, Information Extraction – a User Guide, http://www.dcs.shef.ac.uk *MUC-6, http://www.cs.nyu.edu/cs/faculty/ grishman/muc6.html *I. Muslea, Extraction Patterns for Information Extraction Tasks: A Survey, The AAAI-99 Workshop on Machine Learning for Information Extraction. Califf, Relational Learning of Pattern-Matching Rule for Information Extraction, AAAI-99.