1. Natural Language Processing Applied to Archival Description of Textual E-records
William Underwood
Georgia Tech Research Institute
Atlanta, Georgia
WVU/NETL/ERA Workshop
on Digital Preservation of Complex Engineering Data
WVU NRCCE, Morgantown, West Virginia
April 20-21, 2009
We are grateful for the support of this research by the Army Research Laboratory and ERA program of NARA.

2. Overview Archival Description
Method for extracting metadata from textual e-records
Use of the metadata in archival description
Next Steps
Overview

3. Archival Description Archival Description includes:
The titling of records that do not have titles
The summary of the content of records, folders of records and series of records.
When time allows, the creation of other finding aids such as subject indexes to record series.
Archival Description

4. Archival Description: Research Motivation
Archivists cannot describe a series until the record series has been manually read and reviewed.
With increasing volumes of e-records, it may be decades, even centuries, before new acquisitions are described.
In responding to FOIA requests, Archivists need to be able to search collections of e-records with high precision and recall.
However, at the time of responding to FOIA requests, archivists have not read all of the records, so cannot index the records and search on document types, dates of records, author’s and addressee’s names and the topics of records.
The results set of a query is a list of file names, not record titles and summaries of content
Archival Description: Research Motivation

5. Archival Description: Item Scope and Content Note
Descriptions of records include names of author(s) and addressees, topics, actions and sometimes dates.
Example of an item (record) description from NARA’s Archival Research Catalog (ARC)
This letter was typewritten by President George H. W. Bush and addressed to his children: George, Jeb, Neil, Marvin, and Doro. He expresses his happiness at their Christmas celebration held at Camp David, then writes concerning his conflicted feelings as he prepares for the possibility of war with Iraq.
Archival description of textual records is necessary to support access to records and understanding of the records.
Description is of individual items, file units and record series and of the context of the records.
If one reviews guidelines for archival description such as ISAD-G or NARA’s Life Cycle Data Requirements, one will find that the actions of the record should be indicated in the description.
In this example, “expresses his happiness and “expresses his conflicted feelings” are the actions.
Archival Description: Item Scope and Content Note

6. A Method for Extracting Metadata for Archival Description
Input: Textual Document
Information Extraction
Document Type Recognition
Speech Act Transducer
Discourse Analysis for Topic Recognition
Output: [document(e1), author(e1, S), addressee(e1, H), act(e1 F(P)), topic(e1, T), date(e1, D)]
90% performance in identifying person’s names, organization and location names, and dates.
Automatic identification of documentary forms such as memos correspondence, agency, minutes of meetings, press releases
Metadata extraction such as date, author(s), addressee(s) and topics.
A Method for Extracting Metadata for Archival Description

7. Information Extraction: Method
Information extraction (semantic tagging) is a technology used to identify and annotate semantic categories in text (e.g. names of persons, organizations and locations, job titles, dates).
Document Reader
English Tokenizer
Wordlist Lookup + enhanced wordlists
Sentence Splitter
Hepple POS Tagger + lexicon
Semantic Tagger + Named Entity Rules
Information Extraction: Method

8. Information Extraction: Wordlist Lookup
Person_female_first.lst (8263)
Person_female_first_ambig.lst (117)
Person_male_first.lst (3704)
Person_male_first_ambig.lst (1,117)
Person_surname.lst (83,805)
Person_surname_ambig.lst (6,802)
Person_headofstate_90.lst (478)
Location_city_US.lst (33,017)
Location_city_us_ambig.lst (5,478)
Location_foreign_city.lst (3802)
Information Extraction: Wordlist Lookup

9. Java Annotation Pattern Engine (JAPE) Rules

10. Annotated Person Names and Job Titles

11. Information Extraction: Performance

12. Document Types Agenda Bar Chart Biography Briefing Memo Decision Memo
Correspondence
Diary
Executive Order
Information Memo
Job Application
List of Candidates for Federal Office
Mailing List
Memo
Minutes of Meeting
National Security Directive (NSD)
Newsletter
Nomination to Federal Office
Notes
Presidential Statement
Press Pool Report
Press Release
Referral Memo
Resume
Schedule
Signature Memo
Situation Report
Summary
Transcript of Speech
Telephone Call Recommendation
Transcript of News Conference
Document Types

13. Document Type Recognition
Input: Annotated text from Information Extractor
Intellectual Element Annotator + Intellectual Element Rules
SUPPLE Parser/Interpreter + Document Type Grammars augmented with Semantics
Extract Metadata
Output: [document(e1), author(e1, S), addressee(e1, H), topic(e1, T), date(e1, D)]
Document Type Recognition

14. Document Types: Intellectual Element Recognition
The illustration at the left shows a document whose dates, times , person, location and organization names have been annotated by the first six steps of the method.
The illustration on the right shows the same document after the recognition of the intellectual elements of the document./
Document Types: Intellectual Element Recognition

15. Document Types: Grammar for the Structure of a Memorandum

16. Document Types: Grammar for Memorndum with Semantic Rules

17. Parse Tree and Semantics of a Document

18. Extracted Metadata and Item Description
Document_Type = memo
Date = April 27, 1992
Author = SAM SKINNER
Addressee = EDE HOLIDAY
Topic = California Earthquake
A memorandum dated April 27, 1992 from EDE Holiday to Sam Skinner regarding California Earthquake.
Extracted Metadata and Item Description

19. Speech Act Transducer Annotation of Explicit Speech Acts
Annotation of Implicit Speech Acts
Annotation of Speech Acts Indicated by Text Structure
Annotation of Indirect Speech Acts
Annotation of the Primary Speech Acts
Speech Act Transducer

20. Speech Acts I recommend that you attend the conference. recommend
Performative verb - Verb whose action is accomplished merely by saying it or writing it.
I recommend that you attend the conference.
Illocutionary force of a message.
recommend
Propositional content of a message
you attend the conference
An explicit performative sentence is a sentence in which the illocutionary force is made explicit by naming the force.
I promise to be there
An implicit performative sentence is a sentence in which the illocutionary force is not made explicit by naming the force.
I shall be there
John Austin, a philosopher of language, observed that language is not only used to describe acts, but to perform acts.
Verbs such as recommend, request, and promise whose action is performed by meerly saying them are termed performative verbs.
Austin also distinguished the propositional content of a message from the illocutionary force of a message. For instance, (see slide)
An explicit performative sentence is one in which the illocutionary force is made explicit by the use of a performative verb, for example, “I promise to be there.”
An implicit performative sentence is a sentence in which an illocutionary force is not make explicit by a performative verb, for example, “ I shall be there.”. The illocutionary formce is still promise.
Speech Acts

21. Declarative, imperative and interrogative sentences also express speech acts.
Declarative (state)
You completed the report.
Imperative (request)
Please, complete the report.
Interrogative (ask)
Did you complete the report?
These are additional ways in which implicit speech acts are expressed.
Speech Acts: Implicit

22. An indirect speech act is a speech act that is performed indirectly by way of performing another.
Can you pass the salt? (ask)
in the appropriate context means
Please, pass the salt. (request)
Textual structure can also indicate illocutionary force.
Example: a section heading RECOMMENDATIONS can indicate the sentences in a section have the illocutionary force recommend.
Speech Acts

23. Speech Acts in Presidential Records
assert, deny, state, declare(1), tell(1), report, advise(1), remind, inform, certify(1), agree(1), acknowledge, praise(1), commit, pledge, direct, request, ask(1), ask(2), urge, encourage, invite, order(1), prohibit, suggest(2), propose, recommend, declare(2), resign, confirm, nominate, appoint, authorize, pray, terminate, veto, approve(1), disapprove, revoke, mourn, congratulate, thank, apologize, and welcome(2).
concur, salute, amend, counsel, welcome(1), tender(2), call on, block, retire, proclaim, delegate, designate, determine, find, reject(2), endorse, appreciate, regret, trust(1) , believe, want, desire, and intend.
44 of the speech acts were previously identified by Vanderveken
I identified 23 additional speech acts which I defined.
Speech Acts in Presidential Records

24. Uses of Extracted Metadata in Automatic Description
Signature Memorandum from Boyden Gray to the President recommending the nomination of Ronald B. Leighton to be a US District Judge.
Letter from President Bush to President Mikhail Gorbachev suggesting an informal meeting.
Memorandum from President Bush to Boyden Gray requesting an analysis of the War Powers Resolution.
Letter from Susan Black to President Bush expressing appreciation for nomination and commitment to serve.
Referral Memorandum from Sally Kelley to FEMA requesting appropriate action to a letter from Beryl Anthony to the President.
Uses of Extracted Metadata in Automatic Description

25. Next Steps Inducing grammars for documentary form from samples
Create rules for annotating implicit speech acts and speech acts indicated by textual structure.
Evaluate performance of Speech act recognition method
Recognition of the topics of sentences
Discourse Analysis to identify primary topic(s) of records
Generate item, folder and series descriptions and evaluate the method
Implications for Digital Curriculum
To the extent that Digital curators and digital curation curriculum addresses textual records, the curators and curriculum must also address description of textual records.
Many descriptions include an indication of the acts carried out by the records and these acts are explicit or implicit or indirect speech acts.
If we are successful in our research, digital currators will have tools to automatically create descriptions of large volumes of records, thereby alleviating digital currators of the work needed to manually describe those records, and allowing them to focus on other digital curration needs.
Next Steps

26. Additional Information
Website: perpos.gtri.gatech.edu W. Underwood and S. Isbell, Semantic Annotation of Presidential E-Records, Technical Report ITTL/CSITD 08-01, May 2008 W. Underwood and S. Laib. Automatic Recognition of Documentary Forms, Technical Report ITTL/CSITD 08-02, May 2008 W. Underwood. Recognizing Communication Acts in Presidential E-Records. Technical Report ITTL/CSITD 08-03, October 2008
Additional Information