METHOD AND APPARATUS FOR TRAINING LANGUAGE MODEL, AND METHOD AND APPARATUS FOR RECOGNIZING LANGUAGE

• US 20160210551A1
• Filed: 01/11/2016
• Published: 07/21/2016
• Est. Priority Date: 01/19/2015
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: LEE, Hodong, MIN, Yunhong, LEE, Hoshik, CHOI, Heeyoul, YOO, Sang Hyun, LEE, Yeha, LEE, Jihyun, CHOI, YoungSang
• Granted Patent: US 10,380,483 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 40/20 Natural language analysis semantic analysis of natural language G06F40/30
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G10L 15/005 Language recognition
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules

Abstract

A method and apparatus for training a language model, include generating a first training feature vector sequence and a second training feature vector sequence from training data. The method is configured to perform forward estimation of a neural network based on the first training feature vector sequence, and perform backward estimation of the neural network based on the second training feature vector sequence. The method is further configured to train a language model based on a result of the forward estimation and a result of the backward estimation.

Claims

1. A method, comprising:

generating a first training feature vector sequence and a second training feature vector sequence from training data;

performing forward estimation of a neural network based on the first training feature vector sequence, and performing backward estimation of the neural network based on the second training feature vector sequence; and

training a language model based on a result of the forward estimation and a result of the backward estimation.

2. The method of claim 1, wherein the neural network comprises first hidden layers for the forward estimation and second hidden layers for the backward estimation, and

the first hidden layers are separate from the second hidden layers.

3. The method of claim 1, wherein the generating comprises:

converting the training data into a word vector sequence; and

generating the first training feature vector sequence starting in a forward direction of the word vector sequence and the second training feature vector sequence starting in a backward direction of the word vector sequence.

4. The method of claim 1, wherein the forward estimation comprises estimating a subsequent word to be connected to a first word included in the training data, and

the backward estimation comprises estimating a previous word connected to a second word included in the training data.

5. The method of claim 1, wherein the training comprises:

calculating an error value between the training data and output data of the neural network based on the result of the forward estimation and the result of the backward estimation; and

updating a connection weight between artificial neurons included in the neural network based on the error value.

6. The method of claim 1, wherein the language model is configured based on a recurrent neural network of a bidirectional long short-term memory (BLSTM) structure.

7. The method of claim 6, wherein the recurrent neural network of the BLSTM structure comprises memory blocks configured to maintain a value output in a previous time period until a subsequent time period.

8. The method of claim 1, wherein the training data comprises sequential data comprising speech data, handwriting data, or a combination thereof.

12. A non-transitory computer-readable storage medium comprising a program comprising instructions to cause a computer to perform the method of claim 1.

9. A method, comprising:

generating a first input feature vector sequence and a second input feature vector sequence from input data; and

performing forward estimation of a neural network based on the first input feature vector sequence and performing backward estimation of the neural network based on the second input feature vector sequence to estimate a result of recognizing the input data.

10. The method of claim 9, wherein the neural network comprises first hidden layers for the forward estimation and second hidden layers for the backward estimation, and

the first hidden layers are separate from the second hidden layers.

11. The method of claim 9, wherein the estimating comprises

estimating the result of recognizing the input data based on a forward estimation result associated with the first input feature vector sequence and a backward estimation result associated with the second input feature vector sequence.

20. A non-transitory computer-readable storage medium comprising a program comprising instructions to cause a computer to perform the method of claim 9.

13. An apparatus, comprising:

a training data preprocessor configured to generate a first training feature vector sequence and a second training feature vector sequence from training data; and

a language model trainer configured to train a neural network based language model based on the first training feature vector sequence and the second training feature vector sequence, perform forward estimation of the neural network with respect to the first training feature vector sequence, and perform backward estimation of the neural network with respect to the second training feature vector sequence.

14. The apparatus of claim 13, wherein the neural network comprises first hidden layers for the forward estimation and second hidden layers for the backward estimation, and the first hidden layers are separate from the second hidden layers.

15. The apparatus of claim 13, wherein the language model trainer is configured to train the language model based on a result of the forward estimation and a result of the backward estimation.

16. The apparatus of claim 13, wherein the language model trainer is configured to calculate an error value between the training data and output data based on a result of the forward estimation and a result of the backward estimation, and update a connection weight between artificial neurons included in the neural network based on the error value.

17. The apparatus of claim 13, wherein the training data preprocessor is configured to convert the training data into a word vector sequence, and generate the first training feature vector sequence starting in a forward direction of the word vector sequence and the second training feature vector sequence starting in a backward direction of the word vector sequence.

18. An apparatus, comprising:

an input data preprocessor configured to generate a first input feature vector sequence and a second input feature vector sequence from input data; and

an input data recognizer configured to perform forward estimation of a neural network based on the first input feature vector sequence, and perform backward estimation of the neural network based on the second input feature vector sequence to estimate a result of recognizing the input data.

19. The apparatus of claim 18, wherein the first hidden layers are separate from the second hidden layers.