Entity Augmentation Service from Latent Relational Data

• US 20130238621A1
• Filed: 03/06/2012
• Published: 09/12/2013
• Est. Priority Date: 03/06/2012
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Chakrabarti, Kaushik, Chaudhuri, Surajit, Ganjam, Kris K., Yakout, Mohamed A.
• Granted Patent: US 9,171,081 B2
• Time in Patent Office: Days
• US Class Current: 707/737
• CPC Class Codes: G06F 16/2465 Query processing support for facilitating data mining operations in structured databases
• CPC Class Codes: G06F 16/951 Indexing; Web crawling techniques
• CPC Class Codes: G06F 2216/03 Data mining

Abstract

The subject disclosure is directed towards providing data for augmenting an entity-attribute-related task. Pre-processing is preformed on entity-attribute tables extracted from the web, e.g., to provide indexes that are accessible to find data that completes augmentation tasks. The indexes are based on both direct mappings and indirect mappings between tables. Example augmentation tasks include queries for augmented data based on an attribute name or examples, or finding synonyms for augmentation. An online query is efficiently processed by accessing the indexes to return augmented data related to the task.

Claims

1. In a computing environment, a method performed at least in part on at least one processor, comprising, processing an augmentation task, including accessing relationship-based data corresponding to relationships, including at least one indirect relationship, between entities and attributes mined from at least one corpus, and using the relationship-based data to find data that completes the augmentation task.

2. The method of claim 1 further comprising, receiving the augmentation task as a query for augmented data, in which the query identifies an entity set and an attribute name, and where using the relationship-based data to complete the augmentation task comprises looking up at least one attribute name for each entity of the entity set.

3. The method of claim 1 further comprising, receiving the augmentation task as a query for augmented data, in which the query identifies an entity set and one or more attribute examples, and wherein using the relationship-based data to complete the augmentation task comprises looking up at least one attribute name for at least one entity of the entity set based on the one or more attribute examples.

4. The method of claim 1 further comprising, receiving the augmentation task as a query for synonym data, in which the query identifies a name, and Wherein using the relationship-based data to complete the augmentation task comprises looking up at least one synonym for the name.

5. The method of claim 1 wherein accessing the relationship-based data comprises identifying seed tables and computing topic sensitive pagerank scores via a preference vector and stored vectors for tables.

6. The method of claim 5 wherein identifying the seed tables comprises accessing an index.

7. The method of claim 1 wherein using the relationship-based data to find data that completes the augmentation task comprises aggregating and selecting values from final predictions.

8. The method of claim 1 further comprising, preprocessing corpus data into one or more indexes, including extracting relational tables from the at least one corpus, processing the tables into a graph, and processing the tables into vectors.

9. The method of claim 1 further comprising, extracting relational tables from the at least one corpus, processing the tables into a graph using holistic matching based on topic sensitive pagerank over the graph to provide the relationship-based data.

10. The method of claim 1 further comprising, extracting relational tables from the corpus, and processing the tables into a graph using a model, including using a feature set comprising one or more features corresponding to context associated with each table.

11. The method of claim 10 wherein using the feature set comprises, for at least one of the features, providing a representative document associated with a table, and computing a document similarity matrix from a plurality of documents.

12. The method of claim 10 further comprising, automatically generating labeled training data, and training the model based at least in part on the automatically generated labeled training data.

13. The method of claim 1 further comprising, extracting relational tables from the corpus, and splitting at least one non-binary relational table into a plurality of binary tables based upon subject.

14. The method of claim 1 further comprising, extracting relational tables from the corpus, including classifying to separate the relational tables from other structured data.

16. The system of claim 14 wherein the service is further configured to process a request for synonyms by accessing a synonym-related index built from the relational tables, including based upon indirect mappings.

17. The system of claim 14 wherein the vectors of each table correspond to personal page rank vectors.

18. The system of claim 14 wherein the indirect mappings are based upon topic sensitive pagerank over a graph built from the tables.

20. The one or more computer-readable media of claim 18 wherein performing the holistic matching includes using context associated with at least two of the tables.

21. The one or more computer-readable media of claim 18 having further computer-executable instructions comprising, receiving the augmentation task as a query for augmented data in which the query identifies an entity set and an attribute name, receiving the augmentation task as a query for augmented data, in which the query identifies an entity set and one or more attribute examples, or receiving the augmentation task as a query for synonym data, in which the query identifies a name.

15. A system comprising, a service configured to process domain-independent entity augmentation tasks, including to preprocess relational tables obtained from at least one corpus into a plurality of indexes, in which the indexes include data based upon indirect mappings between at least some of the tables, and to respond to a query corresponding to an entity augmentation task with data obtained via the indexes, including to identify seed tables via the index, compute scores for each seed table, compute a preference vector based upon vectors of the seed tables, compute prediction scores based upon the preference vectors and vectors associated with at least some of the relational tables, aggregate the scores, and return a final prediction based upon aggregation of the scores to complete an entity augmentation task to respond to the query.

19. One or more computer-readable media having computer-executable instructions, which when executed perform steps, comprising:

pre-processing entity attribute relation tables extracted from a corpus into indexes used for entity augmentation, including performing holistic matching between tables that includes computing values for direct relationships and indirect relationships between at least some of the tables; and

accessing the indexes to process an entity augmentation task.