Method for learning local syntactic relationships for use in example-based information-extraction-pattern learning

• US 5,841,895 A
• Filed: 10/25/1996
• Issued: 11/24/1998
• Est. Priority Date: 10/25/1996
• Status: Expired due to Fees

• Current Assignee: Price Waterhouse World Firm Services BV
• Original Assignee: PricewaterhouseCoopers LLP (Pricewaterhousecoopers International Limited)
• Inventors: Huffman, Scott B.
• Primary Examiner(s): Boudreau, Leo H.
• Assistant Examiner(s): MEHTA, BHAVESH M
• Application Number:
  US08/736,890
• Time in Patent Office: 760 Days
• Field of Search: 395/21, 395/23, 395/50, 395/54, 395/75, 395/77, 706/15, 706/16, 706/19, 706/20, 706/25, 706/50, 706/59, 706/61, 382/155, 382/156, 382/181, 382/224, 382/229
• US Class Current: 382/155
• CPC Class Codes: G06V 30/246 using linguistic properties, e.g. specific for English or German language

Abstract

A method is provided for learning local syntactic relationships for use in an example-based information-extraction-pattern learning element of an automated information extraction system. The example-based learning element learns information extraction patterns from user-provided examples of texts paired with events the texts contain; these patterns can then be used by the information extraction system to recognize similar events in subsequent texts. The learning element learns patterns by analyzing each example text/event pair to determine paths of local syntactic relationships between constituents in the text that indicate the event. The learning element employs an incomplete dictionary of local syntactic relationships for this analysis. The present invention learns new local syntactic relationships for text/event pairs that cannot be analyzed using the learning element'"'"'s initial, incomplete dictionary of relationships. These new relationships are added to the dictionary, and allow the learning element to learn patterns from example text/event pairs that cannot be analyzed using only the initial dictionary.

Claims

1. In a computerized example-based pattern learning element of an information-extraction system, said element having as input, texts containing user-identified events and as output, information-extraction patterns that can be used to extract similar events from similar texts, a method for learning local syntactic relationships to be used as components within information-extraction patterns learned by said pattern learning element, said method comprising the steps of:

analyzing an example text and event given to the learning element to establish that the learning element contains an incomplete current dictionary of local syntactic relationships needed to form paths of relationships between all constituents of the text that participate in the event, thus making the learning element unable to form a new extraction pattern from said example text and event;

thereupon finding the closest pair of constituents within the example text that cannot be related by local syntactic relationships in the current dictionary;

inferring a new local syntactic relationship between said closest pair, said new local syntactic relationship being expressed as a finite state machine;

adding said finite state machine to said dictionary so that said finite state machine can be embedded in patterns produced by said learning element; and

re-invoking the learning element on said example text and said example event to learn an information extraction pattern from said example text and said example event by making use of the newly-inferred local syntactic relationship finite state machine.

2. In the computer-based system according to claim 1 wherein said finding step produces a closest pair of constituents C1 and C2, and wherein said inferring step comprises:

creating an inbound transition, said inbound transition testing for constituent properties equal to a syntactic category of said C1;

creating an outbound transition, said outbound transition testing for constituent properties equal to the syntactic category of said C2; and

inferring a series of states and intermediate transitions, said series being between said inbound transition and outbound transition, wherein said intermediate transitions in said series testing constituent properties based on the syntactic categories, and one of entity types and of head words, said properties being of the series of constituents appearing in the example text between constituents C1 and C2.

3. The method according to claim 2 wherein a head entity type for C1 exists, said inbound creating step further including in said inbound transition a test for said head entity type.

4. The method according to claim 2 wherein a head entity type for C2 exists, said outbound creating step further including in said outbound transition a test for said head entity type.