SYSTEM FOR SEMANTIC INTERPRETATION

• US 20150331850A1
• Filed: 12/19/2014
• Published: 11/19/2015
• Est. Priority Date: 05/16/2014
• Status: Active Grant

• Current Assignee: Semantix Technologies Corporation
• Original Assignee: Sierra Nevada Corporation
• Inventors: Ramish, John
• Granted Patent: US 9,659,005 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/245 Query processing
• CPC Class Codes: G06F 16/36 Creation of semantic tools, e.g. ontology or thesauri
• CPC Class Codes: G06F 16/367 Ontology
• CPC Class Codes: G06F 40/151 Transformation
• CPC Class Codes: G06F 40/154 Tree transformation for tree-structured or markup documents, e.g. XSLT, XSL-FO or stylesheets
• CPC Class Codes: G06F 40/20 Natural language analysis semantic analysis of natural language G06F40/30
• CPC Class Codes: G06F 40/211 Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 40/40 Processing or translation of natural language natural language analysis G06F40/20; semantic analysis G06F40/30

Abstract

A semantic database is generated to provide answers to questions by users. Text processors can receive text from text sources, and can convert the text into intermediate logical statements. The text processors can then convert these statements into unambiguous semantic representations. A semantic database connected to the text processors can store the semantic representations. Query processors connected to the semantic database can receive a question from a computing device operated by a user, and can convert the question into intermediate logical subqueries. The query processors can then use a disambiguation table to generate unambiguous semantic subqueries from these intermediate logical subqueries. Using the semantic database, the query processors can match each semantic subquery to the stored semantic representation, and join results of the matching as appropriate, to determine one or more answers to the question. The query processors can send the one or more answers to the computing device.

Claims

1. A method comprising:

receiving, by one or more text processors, text from one or more text sources;

converting, by the one or more text processors, the text into a plurality of intermediate logical statements that abstract over the syntactic form of information in the text;

converting, by the one or more text processors, each intermediate logical statement into a semantic representation that equates intermediate logical statements having equivalent meanings, the semantic representation expressing facts that characterize at least one of relations and events; and

sending, by the one or more text processors, the semantic representation to a semantic database configured to store the semantic representation.

2. The method of claim 1, wherein the semantic database is used to determine one or more answers to a question by a user.

3. The method of claim 1, wherein the text comprises a plurality of sentences.

4. The method of claim 1, wherein the one or more text sources comprise at least one of:

web pages and internal documents.

5. The method of claim 1, wherein each intermediate logical statement corresponds to a tree having a root node, one or more other nodes, and one or more edges, the root node representing a verb in the intermediate logical statement, the one or more other nodes representing at least one of a logical subject, a logical object, prepositional objects, and other modifiers in the intermediate logical statement, and the one or more edges identifying relationships among the root and the one or more other nodes.

6. The method of claim 1, wherein the one or more text processors use a semantic model to convert each intermediate logical statement into the semantic representation, such that:

the semantic model comprises a plurality of frames, each frame representing a relation or event;

each frame comprises one or more variables serving roles in the relation or event; and

each frame comprises one or more equivalent patterns, each pattern having a form of an intermediate logical statement including at least one variable and possibly constants.

7. The method of claim 6, wherein each frame comprises constraints on types of the variables.

8. The method of claim 6, wherein the plurality of frames are hierarchically organized as at least one child frame and at least one parent frame, the at least one child frame expressing a meaning that is narrower than a meaning of the at least one parent frame, the at least one child frame logically implying the at least one parent frame.

9. The method of claim 8, wherein the at least one child frame permits specialization that allows facts in the at least one parent frame to be narrowed in meaning so as to belong to the at least one child frame.

10. The method of claim 6, wherein at least one frame is related to at least one other frame with a lateral implication such that facts in the at least one other frame are inferred non-hierarchically from facts in the at least one frame.

11. The method of claim 6, wherein the semantic model is generated using at least one processor, the generation of the semantic model comprising:

receiving one or more intermediate logical statements that abstract over the syntactic form of the information in the text;

substituting variables for nouns in the one or more intermediate logical statements to generate corresponding patterns;

exploiting coreference information about nouns to obtain candidate semantic links between the corresponding patterns;

clustering the linked patterns based on a similarity metric; and

generating, based on a review of the clustered patterns, a frame hierarchy comprising a plurality of frames forming the semantic model.

12. The method of claim 11, wherein the at least one processor is the one or more text processors.

13. The method of claim 1, wherein the converting of the text into the intermediate logical statements of the text by the one or more text processors comprises one or more of:

parsing the text into a parse tree;

extracting and labeling named entities;

resolving entity mentions within and across documents for coreference;

disambiguating word senses, so as to assign sense identifiers to words;

replacing relative pronouns in the parse tree with corresponding referents;

replacing entity mentions including pronouns and definite nominals in the parse tree with canonical mentions;

identifying intermediate logical statements from at least one of verbs, verbals, appositives, event nouns, possessives, and special constructs in the parse tree;

identifying for each intermediate logical statement the corresponding logical subjects, logical objects, prepositional phrases, and other modifiers from the parse tree;

converting possessives from the parse tree into prepositional forms;

converting indirect objects from the parse tree into prepositional forms;

distributing over conjunctions in the parse tree; and

substituting variables for subordinate clauses in the parse tree.

14. The method of claim 6, wherein the conversion of each intermediate logical statement into the semantic representation by the one or more text processors comprises:

receiving the intermediate logical statement;

identifying candidate interpretations associated with the intermediate logical statement, a candidate interpretation being characterized by a frame, bindings for some or all variables of the frame, and possibly supplemental modifiers; and

selecting, from the candidate interpretations, a best interpretation based on at least one of variable type compatibility and likelihood, the best interpretation characterizing the semantic representation.

15. The method of claim 14 wherein the conversion of each intermediate logical statement into the semantic representation by the one or more text processors further comprises:

specializing the selected interpretation to a narrower meaning as applicable based on at least one of variable type compatibility and likelihood;

deducing any applicable lateral implications from the specialized interpretation; and

recursively specializing and deducing lateral implications for each new fact concluded from a lateral implication in order to generate the semantic representation.

16. The method of claim 14, wherein identifying the candidate interpretations by the one or more text processors comprises:

identifying matching pattern forms, a pattern form of the matching pattern forms being a pattern with wildcard characters in place of variables to indicate that the variables are undistinguished;

selecting, from the identified pattern forms, a pattern form that binds to the most elements of the intermediate logical statement and is maximally specific to the intermediate logical statement;

identifying patterns in frames that match the selected pattern form; and

determining interpretations given by each matching pattern, the determined interpretations being the candidate interpretations.

17. The method of claim 1, wherein the one or more text processors and the semantic database are connected via a communication network.

18. A method comprising:

receiving, by one or more query processors, a question from a computing device;

converting, by one or more query processors, the question into one or more intermediate logical subqueries that abstract over the syntactic form of the question;

converting, by the one or more query processors, each intermediate logical subquery into a semantic subquery in a representation that equates intermediate logical subqueries having equivalent meanings;

sending, by the one or more query processors, the semantic query to the database, the semantic query being one or more semantic subqueries joined together;

retrieving, by one or more database processors, facts from the database that match the semantic subqueries;

joining, by the one or more database processors, the facts when necessary to generate one or more answers to the question;

returning, by the one or more database processors, the joined facts to the one or more query processors; and

sending, by the one or more query processors, the one or more answers to the computing device.

19. The method of claim 18, further comprising:

combining, filtering, and reorganizing the one or more answers before sending the one or more answers to the computing device.

20. The method of claim 18, wherein the computing device executes a graphical user interface that displays the one or more answers to the question.

21. The method of claim 18, wherein each intermediate logical subquery corresponds to a tree having a root node, one or more other nodes, and one or more edges, the root node representing a verb in the intermediate logical subquery, the one or more other nodes representing at least one of a logical subject, a logical object, prepositional objects, and other modifiers in the intermediate logical subquery, and the one or more edges identifying relationships among the root and the one or more other nodes.

22. The method of claim 18, wherein the conversion of the question into the intermediate logical subqueries by the one or more query processors comprises one or more of:

parsing the question into a parse tree;

converting the parse tree into a declarative form;

identifying explicit and implicit query variables in the parse tree;

replacing relative pronouns in the parse tree with corresponding referents;

replacing entity mentions including pronouns and definite nominals in the parse tree statements with canonical mentions;

identifying intermediate logical subqueries from at least one of verbs, verbals, appositives, event nouns, possessives, and special constructs in the parse tree;

identifying, for each intermediate logical subquery, corresponding logical subjects, logical objects, prepositional phrases, and other modifiers from the parse tree;

converting possessives from the parse tree into prepositional forms;

converting indirect objects from the parse tree into prepositional forms;

distributing over conjunctions in the parse tree; and

substituting variables for subordinate clauses in the parse tree.

23. The method of claim 18, wherein the conversion of the intermediate logical subquery into the semantic subquery by the one or more query processors comprises:

identifying a surface form that corresponds to the intermediate logical subquery;

retrieving interpretations for the surface form from a disambiguation table in the database; and

selecting, from the retrieved interpretations, a highest-scoring interpretation that is compatible with query variables of the intermediate logical subquery, the highest-scoring interpretation characterizing the semantic subquery.

24. The method of claim 23, wherein the disambiguation table is generated by:

identifying patterns for each semantic frame, aliases for each entity, and words for each noun or adjective sense as surface forms;

generating surface forms for each interpreted semantic fact by combining the surface forms for the semantic frame of the fact with the surface forms of one or more frame elements; and

assigning scores to interpretations for each surface form.

25. The method of claim 24, wherein the scores for interpretations are computed as frequencies of the interpretations.

26. The method of claim 24, wherein:

the scores are proportional to P(surface_form|interpretation)*P(interpretation), the scores being computed as frequencies of the interpretations weighted by the probability of generating the surface form given the interpretation.

27. The method of claim 18, wherein the one or more query processors and the semantic database are connected via a communication network.

28. A system comprising:

one or more text processors receiving text from one or more text sources, the one or more text processors converting the text into intermediate logical statements that abstract over syntactic form, the one or more text processors converting each intermediate logical statement into a semantic representation that equates intermediate logical statements having equivalent meanings;

a semantic database connected to the one or more text processors, the semantic database receiving the semantic representation from the one or more text processors, the semantic database storing the semantic representation;

one or more query processors connected to the semantic database, the one or more query processors receiving a question from a computing device, the one or more query processors converting the question into one or more intermediate logical subqueries that abstract over syntactic form, the one or more query processors converting each intermediate logical subquery into a semantic subquery in a representation that equates intermediate logical subqueries having equivalent meanings, the semantic database matching each semantic subquery to the stored semantic representations and joining results as appropriate to determine one or more answers to the question, the one or more query processors sending the one or more answers to the computing device that outputs the one or more answers.

29. The system of claim 28, wherein:

the one or more text processors use a semantic model to convert the text into the semantic representation; and

the one or more query processors use the semantic model to convert the question to the semantic query.

30. The system of claim 29, wherein:

the one or more text processors and the semantic database are connected via a first communication network; and

the one or more query processors and the semantic database are connected via a second communication network.

31. The system of claim 30, wherein the first communication network is same as the second communication network.

32. The system of claim 30, wherein the first communication network is different from the second communication network.

33. A computer-implemented method comprising:

receiving, by a semantic database, a question from a computing device via a communication network;

identifying, by the semantic database, one or more answers responsive to the question; and

transmitting, by the semantic database, the one or more answers to the computing device,

wherein the semantic database is generated by;

receiving text from one or more text sources;

converting the text into intermediate logical statements that abstract over syntactic form;

converting the intermediate logical statements into a semantic representation that equates intermediate logical statements having equivalent meanings; and

storing the semantic representation, the stored semantic representations forming the semantic database.