Unsupervised learning of deep patterns for semantic parsing

• US 9,292,490 B2
• Filed: 08/16/2013
• Issued: 03/22/2016
• Est. Priority Date: 08/16/2013
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Kimelfeld, Benny, Vaithyanathan, Shivakumar
• Primary Examiner(s): ROBERTS, SHAUN A
• Application Number:
  US13/968,462
• Publication Number:
  US 20150051900A1
• Time in Patent Office: 949 Days
• Field of Search: 704/9
• US Class Current: 1/1
• CPC Class Codes: G06F 40/205 Parsing
• CPC Class Codes: G06F 40/211 Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
• CPC Class Codes: G06F 40/289 Phrasal analysis, e.g. finite state techniques or chunking
• CPC Class Codes: G06F 40/30 Semantic analysis

Abstract

Using exemplary sentences, usage patterns and thematic roles ascribed in VerbNet to generate “deep pattern trees” for the exemplary sentences. Then, when an arbitrary natural language subject sentence is input, these deep pattern trees can be matched to the natural language subject sentence in order to assign thematic roles to at least some of the “grammatical portions” of the natural language subject sentence.

Claims

1. A method of unsupervised learning of deep patterns for semantic parsing, using an algorithm that operates over a VerbNet corpus, the method comprising:

receiving a set of exemplary sentences, including a first exemplary sentence, from a language net;

extracting, from VerbNet, a plurality of markup language files corresponding to a usage pattern P;

parsing each of the exemplary sentences to yield a set of training constituency trees respectively corresponding to the exemplary sentences;

performing maximal frequent subtree analysis on the set of training constituency trees to yield an unfiltered set of deep pattern trees; and

filtering at least one irrelevant tree(s) from the set of deep pattern trees to obtain a filtered set of deep pattern trees, with the at least one irrelevant tree(s) do not contain a part of speech that is included in the usage pattern P;

deleting redundant leaves from at least one of the following;

the filtered set of deep pattern, and/or the unfiltered set of deep pattern trees; and

matching nodes of each tree pattern, of the filtered set of deep tree patterns, with items of the usage pattern P using machine logic based rules to facilitate translation of the filtered set of deep pattern trees into annotation query language rules;

At least the parsing step is performed by computer software running on computer hardware.

2. The method of claim 1 further comprising:

responsive to the matching of the nodes, translating the filtered set of deep pattern trees into annotation query language (AQL) rules.

3. The method of claim 1 wherein:

the language net organizes the exemplary sentences into a set of usage patterns; and

the performing maximal subtree analysis step is performed separately for each sub-set of training constituency trees respectively corresponding to each usage pattern in the language net.

4. A computer program product, for unsupervised learning of deep patterns for semantic parsing, using an algorithm that operates over a VerbNet corpus, the computer program product comprising software stored on a software storage device, the software comprising:

first program instructions programmed to extract, from VerbNet, a plurality of markup language files corresponding to a usage pattern P;

second program instructions programmed to parse each of the exemplary sentences to yield a set of training constituency trees respectively corresponding to the exemplary sentences;

third program instructions programmed to perform maximal frequent subtree analysis on the set of training constituency trees to yield an unfiltered set of deep pattern trees; and

fourth program instructions programmed to filter at least one irrelevant tree(s) from the set of deep pattern trees to obtain a filtered set of deep pattern trees, with the at least one irrelevant tree(s) do not contain a part of speech that is included in the usage pattern P;

fifth program instructions programmed to delete redundant leaves from at least one of the following;

the filtered set of deep pattern, and/or the unfiltered set of deep pattern trees; and

sixth program instructions programmed to match nodes of each tree pattern, of the filtered set of deep tree patterns, with items of the usage pattern P using machine logic based rules to facilitate translation of the filtered set of deep pattern trees into annotation query language (AQL) rules.

5. The product of claim 4 wherein the software further comprises:

seventh program instructions programmed to, responsive to the matching of the nodes, translate the filtered set of deep pattern trees into annotation query language (AQL) rules.

6. The product of claim 4 wherein:

the language net organizes the exemplary sentences into a set of usage patterns; and

the third program instructions are further programmed to perform maximal subtree analysis separately for each sub-set of training constituency trees respectively corresponding to each usage pattern in the language net.

7. A computer system for unsupervised learning of deep patterns for semantic parsing, using an algorithm that operates over a VerbNet corpus, the computer system comprising:

a processor(s) set; and

a software storage device;

wherein;

the processor set is structured, located, connected and/or programmed to run software stored on the software storage device; and

the software comprises;

first program instructions programmed to extract, from VerbNet, a plurality of markup language files corresponding to a usage pattern P;

second program instructions programmed to parse each of the exemplary sentences to yield a set of training constituency trees respectively corresponding to the exemplary sentences;

third program instructions programmed to perform maximal frequent subtree analysis on the set of training constituency trees to yield an unfiltered set of deep pattern trees; and

fourth program instructions programmed to filter at least one irrelevant tree(s) from the set of deep pattern trees to obtain a filtered set of deep pattern trees, with the at least one irrelevant tree(s) do not contain a part of speech that is included in the usage pattern P;

fifth program instructions programmed to delete redundant leaves from at least one of the following;

the filtered set of deep pattern, and/or the unfiltered set of deep pattern trees; and

sixth program instructions programmed to match the nodes of each tree pattern, of the filtered set of deep tree patterns, with items of the usage pattern P using machine logic based rules to facilitate translation of the filtered set of deep pattern trees into annotation query language (AQL) rules.

8. The system of claim 7 wherein the software further comprises:

Seventh program instructions, responsive to the matching of the nodes, translate the filtered set of deep pattern trees into annotation query language (AQL) rules.

9. The system of claim 7 wherein:

the language net organizes the exemplary sentences into a set of usage patterns; and

the third program instructions are further programmed to perform maximal subtree analysis separately for each sub-set of training constituency trees respectively corresponding to each usage pattern in the language net.