Inferring informational goals and preferred level of detail of answers based at least on device used for searching

• US 7,430,505 B1
• Filed: 01/31/2005
• Issued: 09/30/2008
• Est. Priority Date: 06/29/2001
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Horvitz, Eric J., Zukerman, Ingrid
• Primary Examiner(s): Smits; Talivaldis Ivars
• Application Number:
  US11/047,491
• Time in Patent Office: 1,338 Days
• Field of Search: 704/2, 704/9, 704/7, 704/270, 704/10, 707/3, 707/10
• US Class Current: 704/9
• CPC Class Codes: G06F 16/35 Clustering; Classification
• CPC Class Codes: G06F 40/35 Discourse or dialogue representation
• CPC Class Codes: Y10S 707/99933 Query processing, i.e. searching
• CPC Class Codes: Y10S 707/99935 Query augmenting and refining, e.g. inexact access

Abstract

A system and method for inferring informational goals and preferred level of details in answers in response to questions posed to computer-based information retrieval or question-answering systems is provided. The system includes a query subsystem that can receive an input query and extrinsic data associated with the query and which can output an answer to the query, and/or rephrased queries or sample queries. The extrinsic data can include at least in part the type of application or device used to perform the query. The query subsystem accesses an inference model to retrieve conditional probabilities that certain informational goals are present. One application of the system includes determining a user'"'"'s likely informational goals and then accessing a knowledge data store to retrieve responsive information. The system includes a natural language processor that parses queries into observable linguistic features and embedded semantic components that can be employed to retrieve the conditional probabilities from the inference model. The inference model is built by employing supervised learning and statistical analysis on a set of queries suitable to be presented to a question-answering system. Such a set of queries can be manipulated to produce different inference models based on demographic and/or localized linguistic data.

Claims

1. A system for learning how to infer information goals from queries, comprising:

a natural language processor that produces a linguistic data concerning one or more linguistic features;

a tagging tool that facilitates manipulating the linguistic data;

one or more taggers that manipulate the linguistic data; and

an inference model that stores information concerning conditional probabilities associated with the likelihood that one or more informational goals exist, where the conditional probabilities of the informational goals are determined, at least in part, from Bayesian statistical analysis performed on the linguistic data, wherein the inference model automatically selects a set of resources to search based upon an inferred age of a user and obtains information from extrinsic data associated with the query based at least in part on a type of device used.

2. The system of claim 1, the linguistic data comprises a parse tree, where the parse tree contains extractable information concerning the nature of and relationships between observable linguistic features.

3. The system of claim 2, the observable linguistic features comprise word-based features, structural features and hybrid linguistic features.

4. The system of claim 3, the word-based features indicate the presence of one or more candidate terms that can be employed in predicting an informational goal.

5. The system of claim 1, the taggers manipulate the linguistic data to conform with one or more schemas associated with reasoning concerning the relevance of a part of a query based on one or more language models.

6. The system of claim 5, the taggers supervise learning associated with computing probabilities associated with the informational goals.

7. The system of claim 1, the inference model comprises one or more decision trees, the decision trees store conditional probabilities associated with one or more informational goals, the decision trees is traversable by the linguistic data.

8. The system of claim 1, the inference model represents a probabilistic dependency model.

9. A method for generating responses comprising:

inputting a question;

employing natural language processing to parse the question;

employing parse data produced by parsing the question to access a decision model, the decision model storing conditional probabilities associated with informational goals;

inferring one or more informational goals where the goals, based least on a type of device used, include information from extrinsic data associated with the question;

selecting a set of resources to search for an answer to the question based upon an inferred age of a user;

producing an output related to the question and the one or more inferred informational goals; and

adapting the decision model based upon automated learning.

10. The method of claim 9, parsing the question produces parse data from which at least one of, the existence of relationships between linguistic components in the question, the nature of relationships between linguistic components in the question, parts of speech in the question, logical forms of words in the question, logical forms of phrases in the question, structural features in the question, the number of distinct parts of speech in the question, whether the main noun in the question is singular and whether the question contains a proper noun, can be extracted.

11. The method of claim 9, accessing the decision model comprises:

identifying one or more decision trees to access, based at least in part on the parse data;

traversing the one or more decision trees based, at least in part on the parse data; and

retrieving one or more conditional probabilities from the one or more decision trees.

12. The method of claim 9, inferring one or more informational goals comprises:

evaluating the one or more conditional probabilities retrieved from the decision model to determine which, if any, informational goals can be inferred from the one or more conditional probabilities.

13. The method of claim 12, inferring the one or more informational goals further comprises inferring one or more levels of detail for a response to the question.

14. The method of claim 9 further comprising:

initializing the decision model;

collecting one or more training sets of questions;

analyzing the one or more training sets of questions;

updating the decision model based on the analysis of the one or more training sets of questions; and

testing the decision model using one or more testing sets of questions.

15. The method of claim 14, analyzing the one or more training sets of questions comprises:

parsing questions in the one or more training sets of questions to produce linguistic data concerning the questions;

employing supervised learning to establish an inference model; and

employing Bayesian statistical analysis to compute one or more conditional probabilities associated with one or more informational goals based, at least in part, on the linguistic data.

16. The method of claim 15, updating the one or more data structures comprises:

adding zero or more nodes to one or more decision trees;

removing zero or more nodes from one or more decision trees;

adding zero or more leafs to one or more decision trees;

removing zero or more leafs from one or more decision trees;

recomputing zero or more split nodes for one or more decision trees;

adding zero or more decision trees to the inference model; and

removing zero or more decision trees from the inference model.

17. The method of claim 14, updating the decision model comprises:

updating one or more data structures employed to store conditional probabilities associated with one or more informational goals.

18. The method of claim 17, testing the decision model comprises:

producing one or more sets of outputs in response to one or more input testing sets of questions; and

analyzing the responsiveness of the one or more sets of outputs to the information sought in the one or more input testing sets of questions.

19. A computer readable medium having a program embodied thereon to perform the method of claim 18.

20. A system for generating a response for a question posed to an automated question answerer comprising:

means for initializing a model representing the likelihood that a certain type of answer is desired, the model is stored in one or more data repositories, based at least in part on a type of device utilized, the model obtains information from extrinsic data associated with a question;

means for decomposing a the question into parts that facilitate accessing the model;

means for selecting a set of resources to search for an answer to the question based upon at least one inferred user attribute;

manual means for adapting the model;

automated means for adapting the model; and

means for constructing one or more responses to the question based on likelihoods retrieved from accessing the model.

21. The system of claim 20, the user attribute, is at least one of age, gender, religion, ethnicity, income-level, intelligence, or marital status.