Leveraging corporal data for data parsing and predicting

• US 10,311,092 B2
• Filed: 06/28/2016
• Issued: 06/04/2019
• Est. Priority Date: 06/28/2016
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Inventors: Ganjam, Kris K., Chakrabarti, Kaushik
• Primary Examiner(s): Chau, Dung K
• Application Number:
  US15/195,981
• Publication Number:
  US 20170371958A1
• Time in Patent Office: 1,071 Days
• Field of Search: 707739, 707748, 707600, 706 62, 706 47
• CPC Class Codes: G06F 16/353 into predefined classes
• CPC Class Codes: G06F 40/216 using statistical methods
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

The techniques discussed herein leverage structure within data of a corpus to parse unstructured data to obtain structured data and/or to predict latent data that is related to the unstructured and/or structured data. In some examples, parsing and/or predicting can be conducted at varying levels of granularity. In some examples, parsing and/or predicting can be iteratively conducted to improve accuracy and/or to expose more hidden data.

Claims

1. A method comprising:

obtaining an unstructured token, the unstructured token being unassociated with a class identifier, attribute identifier, or attribute value;

parsing the unstructured token, the parsing including associating, based at least in part on a probabilistic database derived from a corpus, a class identifier with the unstructured token to obtain a classified token;

predicting that an attribute label or an attribute value is associated with the classified token based at least in part on another probabilistic database derived from the corpus and the class identifier; and

associating the attribute label or the attribute value with the classified token.

2. A method as claim 1 recites, further comprising iteratively predicting an additional attribute label or an additional attribute value based at least in part on a previously predicted attribute label, a previously predicted attribute value, or a previously parsed classification.

3. A method as claim 1 recites, the parsing the unstructured token further including identifying one or more of a semantic class, syntactic class, or a syntactic function of the unstructured token.

4. A method as claim 3 recites, the associating the class identifier with the undefined token to obtain the classified token being further based at least in part on the syntactic class or the syntactic function of the unstructured token.

5. A method as claim 1 recites, wherein the unstructured token is one token of a set of tokens and the associating the class identifier with the unstructured token to obtain the classified token is further based at least in part on an arrangement of individual tokens of the set of tokens relative to other individual tokens of the set of tokens.

6. A method as claim 1 recites, further comprising calculating a second probabilistic database from the corpus for the unstructured token or the classified token (“

the token-of-interest (“

TOI”

)”

), the calculating including;

identifying class identifiers, class values, attribute labels, or attribute values associated with the TOI, the identified class identifiers, class values, attribute labels, or attribute values being “

candidate classifications”

; and

calculating a frequency with which the TOI is associated with the candidate classifications, the frequency being associated with a probability that the candidate classifications is associated with token.

7. A method as claim 1 recites, wherein the associating the class identifier with the unstructured token or associating the attribute label or the attribute value with the classified token includes:

ranking the candidate classifications according to highest probability of being associated with the token; and

assigning a top-k number of the candidate classifications to the token, where k is an integer value greater than or equal to 1.

8. A method as claim 1 recites, the parsing including:

calculating, from the corpus, the another probabilistic database based at least in part on the unstructured token, the another probabilistic database including;

candidate classifications, the candidate classifications including the class identifier; and

probabilities that the candidate classifications are associated with the unstructured token; and

associating the class identifier with the unstructured token based at least in part on the probabilities.

9. A method as claim 1 recites, the predicting that an attribute label is associated with the classified token including:

calculating, from the corpus, the another probabilistic database, the another probabilistic database including;

candidate attribute labels;

probabilities that the candidate attribute labels are associated with one or more of the class identifier, the unstructured token, or the unstructured token;

associating one of the candidate attribute labels with the token based at least in part on the probabilities, the one of the candidate attribute labels being a latent attribute label.

10. A method as claim 9 recites, the predicting that an attribute value is associated with the classified token including:

calculating, from the corpus, a third probabilistic database, the third probabilistic database including;

candidate attribute values;

second probabilities that the candidate attribute values are associated with one or more of the latent attribute label, the class identifier, the unstructured token, or the classified token; and

associating one of the candidate attribute values with the latent attribute label or the defined token based at least in part on the second probabilities.

11. A method as claim 1 recites, wherein one or more of the class identifier, the attribute label, or the attribute value include a token of the corpus.

12. A system comprising:

one or more processors;

computer-readable media having stored thereon computer-executable instructions that, when executed by the one or more processors, configure the system to perform operations comprising;

obtaining a token-of-interest (“

TOI”

);

generating a schema for the TOI based at least in part on relational data of a corpus, the schema including;

a first relation between the TOI and a parsed label classifying the TOI,

a second relation between the parsed label and a predicted label, and

a third relation between the predicted label and a predicted token, the predicted label classifying the predicted token and the predicted token including latent data associated with the TOI.

13. A system as claim 12 recites, the generating including:

identifying the corpus, the corpus containing one or more of the TOI, the parsed label, the predicted label, or the predicted token; and

selecting the parsed label, the predicted label, or the predicted token from among tokens of the corpus, the selecting based at least in part on the relational data of the corpus and the relational data being based at least in part on;

a frequency with which the parsed label is associated with the TOI in the corpus,

a frequency with which the predicted label is associated with the TOI or the parsed label in the corpus, or

a frequency with which the predicted token is associated with the predicted label or the TOI in the corpus.

14. A system as claim 13 recites, the parsed label, the predicted label, or the predicted token being in a top-k number of tokens ranked by frequency, wherein k is an integer value greater than or equal to 1.

15. A system as claim 13 recites, the relational data further including:

distribution data of frequencies with which tokens of the corpus are associated with the parsed label, the predicted label, or the predicted token;

contextual data, the contextual data including one or more of;

tokens associated with the TOI,

labels associated with the tokens associated with the TOI, or

an arrangement of the TOI relative to the tokens associated with the TOI.

16. A system as claim 12 recites, the generating including:

selecting, from tokens of a corpus of documents containing the TOI, the parsed label based at least in part on a frequency with which the parsed label is related to the TOI in the corpus of documents containing the TOI;

selecting, from tokens of one or more of the corpus of documents containing the TOI or from a corpus of documents containing the parsed label, the predicted label based at least in part on a frequency with which the predicted label is associated with the parsed label in one or more of the corpus of documents containing the TOI or the corpus of documents containing the parsed label; and

selecting, from tokens of a corpus of documents containing the predicted label, the predicted token based at least in part on a frequency with which the predicted token is related to the predicted label in the corpus of documents containing the second label.

17. A method comprising:

tokenizing unstructured data to obtain tokens, the unstructured data lacking relational structure the unstructured data and identifiers of what the data is or what the data is like;

parsing the tokens, based at least in part on probabilities calculated from a corpus, to form structured data;

predicting, based at least in part on one or more of the tokens and probabilities calculated from a corpus, additional tokens or additional structural information, the additional structural information including an attribute identifier and the additional tokens including an attribute value; and

adding the additional tokens or the additional structural information to the structured data.

18. A method as claim 17 recites, the parsing being further based, at least in part, on one or more of:

an arrangement of the tokens relative to each other;

a parsing of one of the tokens;

a predicted additional token or predicted structural information;

or

tail data of probabilities calculated from a corpus.

19. A method as claim 17 recites, further comprising:

receiving a target attribute identifier;

performing one or more predictions of attribute identifiers until the target attribute identifier is predicted, the one or more predictions being based at least in part on one or more of the tokens, the structured data, one or more of previously predicted attribute identifiers, or probabilities calculated from a corpus; and

predicting an attribute value for the target attribute identifier based at least in part on one or more of the tokens or the one or more predicted attribute identifiers.

20. A method as claim 17 recites, the corpus including structured information.