Data processing: speech signal processing – linguistics – language – Linguistics
Reexamination Certificate
1999-09-03
2001-10-30
Edouard, Patrick N. (Department: 2747)
Data processing: speech signal processing, linguistics, language
Linguistics
C704S257000
Reexamination Certificate
active
06311150
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to natural language understanding systems, and more particularly to a method and system for building a natural language understanding system with hierarchical translation of commands to improve speed and accuracy of translations.
2. Description of the Related Art
State-of-the-art natural language understanding systems typically perform a direct translation of the user input to a formal command in a single step. However, when the number of formal commands is large, a large number of features may be required to perform a direct one-step translation. Increasing the number of features introduces additional statistical noise that results in a degradation of accuracy of the system. Additional computational difficulties include added computation time and memory, and possible numerical instability due to the large number of features. Consider for example translating commands such as “forward this message” (action), “how do I forward a message” (help), or “did I forward a message” (query). All three of these sentences are related to forwarding a message, and as such the key word “forward” may appear to be important, but a feature set including the word “forward” would only introduce noise and lead to inaccuracies in the translation.
Therefore, a need exists for a system and method for reducing a number of features needed to translate a command in a natural language understanding system. A further need exists for a hierarchical translation method and system for translating command in a faster and more accurate manner.
SUMMARY OF THE INVENTION
A method for hierarchical translation of input to a formal command in a natural language understanding system, which may by implemented by a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform the method includes presenting an input command to be translated to a natural language understanding engine. At least two translator levels are provided in the natural language understanding engine. The input command is translated into at least one category by a first translator level of the at least two translator levels by associating the input command with the at least one category for a next level of translators. A formal language command is output for the input command from a last of the at least two translator levels based on the input command and the at least one category.
In other methods which may be implemented with a program storage device, the step of providing at least two translator levels may include providing N translator levels where N is defined as a system parameter. The step of translating the command into at least one category may include the step of providing a plurality of models. Each model is preferably associated with a category, and each category may include a subset of formal language commands to narrow a search space for determining the formal language command corresponding to the input command. The step of applying the formal language command to at least one application may be included. The at least one category may include formal language command entries relevant to at least one application and may further include the step of applying the formal language command entries to the at least one application in accordance with the input command. The at least one category may include all entries for the input command and may further include the step of associating a first portion of the entries with the at least one category by providing a formal command for each of the entries relevant for the at least one category. The step of mapping a second portion of the entries to an error command may also be included.
In still other methods, the at least one category may include a do nothing category, and a formal language command may be outputted by the step of outputting a do nothing command associated with incomplete and improper input commands. The at least one category may include a general category for addressing ambiguous commands, commands relevant to multiple categories and system related commands. The at least two translator levels may search for a category selection for a next translator level or may search for the formal language command for the last level, the category selection being limited to a given category from a previous level translator thereby narrowing a search space for each translator level to improve speed and accuracy.
Another method for hierarchical translation of input to a formal command in a natural language understanding system includes the steps of presenting an input command to be translated to a natural language understanding engine, providing a plurality of translator levels in the natural language understanding engine, providing a plurality of categories for each translator level, each category including a group of formal language commands stored in at least one model, associating the input command with the categories for a given translator level, scoring a probability of correct translation for each of the at least two categories, selecting a category of the at least two categories yielding a highest score, associating the input command with the selected category for a next level of translators and outputting a formal language command for the input command from a last of the translator levels based on the input command and the selected category.
In other methods, the step of providing a plurality of translator levels may include providing N translator levels where N is defined as a system parameter. Each category may include a subset of formal language commands to narrow a search space for determining the formal language command corresponding to the input command. The step of applying the formal language command to at least one application may be included. The categories may include formal language command entries relevant to at least one application and may further include the step of applying the formal language command entries to the at least one application in accordance with the input command. The at least two categories may include all entries for the input command and may further include the step of associating a first portion of the entries with a given category by providing a formal command for each of the entries relevant for the given category. The step of mapping a second portion of the entries to an error command may also be included. The step of selecting a category of the at least two categories yielding a highest score may include the step of selecting a category with a lower score if a translation error is encountered for a category or categories yielding a higher score. If a translation error is encountered for all scored categories, a do nothing category may be selected.
In still other methods, at least one category includes a do nothing category and the step of outputting a formal language command may include the step of outputting a do nothing command associated with incomplete and improper input commands. The method may further invlude the step of assigning formal language commands to multiple categories to protect against translation errors in intermediate levels of the plurality of translator levels. The above methods may be performed by a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform the method steps.
A method for building hierarchical models for translating an input command to a formal command in a natural language understanding system includes the steps of collecting training data from a domain of inputs, separating the training data into translation levels, the translation levels arranged to provide a given output at each level for associating the inputs to a category, the training data including the inputs and an associated formal language command for each input, for each translation level, indicating categories for searching for a formal language command by dividing up a formal language command search space and associating the categories with e
Kleindienst Jan
Ramaswamy Ganesh N.
Edouard Patrick N.
F. Chau & Associates LLP
International Business Machines - Corporation
LandOfFree
Method and system for hierarchical natural language... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and system for hierarchical natural language..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and system for hierarchical natural language... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2589965