Generation of matched corpus for language model training

1. A computer-implemented method for generating a text, the method comprising:

obtaining a first text collection matched with a target domain and a second text collection including a plurality of samples each describing rewriting between a first text and a second text having a style different from the first text;

training a text generation model of a learning machine with the first text collection and the second text collection, the text generation model having, in a vocabulary, one or more operation tokens indicating rewriting;

creating a seed model used for obtaining the plurality of texts in conjunction with the text generation model, the seed model created by a nonlinear interpolation of a first seed model base trained with the first text collection and a second seed model base trained with the second text collection, the nonlinear interpolation including a plurality of conditions for giving weights for mixing the first and second seed model bases; and

outputting a plurality of texts obtained from the text generation model, the plurality of texts being configured for training a language model of a speech recognition engine.

2. The method of claim 1 , wherein the one or more operation tokens includes one or more insertion tokens and/or one or more deletion tokens, each insertion token indicating insertion of a text part, each deletion token indicating deletion of a text part.

3. The method of claim 1 , the method further comprising:

generating a plurality of tagged texts having at least one operation token using at least the text generation model; and

applying a rewriting rule to each tagged text to obtain a plurality of untagged texts as the plurality of the texts used for training a language model.

4. The method of claim 1 , wherein the first seed model base is a topic seed base (T) and the second seed model base is a rewriting seed base (R).

5. The method of claim 4 , wherein the topic seed base (T) and the rewriting seed base (R) are unigram language models that provide a probability for each token.

6. The method of claim 5 , wherein the nonlinear interpolation includes a first condition giving priority to the second seed model base for filler and operation tokens and a second condition giving priority to the first seed model base or the second seed model base depending on probabilities for word tokens.

7. The method of claim 1 , wherein the text generation model has more complexity than a language model to be trained with the plurality of texts.

8. The method of claim 1 , wherein the plurality of texts is used for training a language model in conjunction with at least one of a base language model trained with a corpus other than the first text collection and second text collection, a first language model trained with the first text collection and a second language model trained with the second text collection.

9. The method of claim 1 , wherein, the plurality of texts is used for training a language model and an amount of the plurality of texts is larger than a total amount of the first text collection and the second text collection.

10. The method of claim 1 , wherein the first text has a spontaneous style and the style of the second text is a more formatted style than the first text.

11. A computer system for generating a text, by executing program instructions, the computer system comprising:

a memory storing the program instructions;

a processor device in communications with the memory for executing the program instructions, wherein the processing unit is configured to:

obtain a first text collection matched with a target domain and a second text collection including a plurality of samples each describing rewriting between a first text and a second text having a style different from the first text;

train a text generation model with the first text collection and the second text collection, wherein the text generation model has, in a vocabulary, one or more operation tokens indicating rewriting;

create a seed model used for obtaining the plurality of texts in conjunction with the text generation model, the seed model created by a nonlinear interpolation of a first seed model base trained with the first text collection and a second seed model base trained with the second text collection, the nonlinear interpolation including a plurality of conditions for giving weights for mixing the first and second seed model bases; and

output a plurality of texts obtained from the text generation model the plurality of texts being configured for training a language model of a speech recognition engine.

12. The computer system of claim 11 , wherein the one or more operation tokens includes one or more insertion tokens and/or one or more deletion tokens, each insertion token indicating insertion of a text part, each deletion token indicating deletion of a text part.

13. The computer system of claim 11 , wherein the processor device is further configured to:

generate a plurality of tagged texts having at least one operation token using at least the text generation model; and

apply a rewriting rule to each tagged text to obtain a plurality of untagged texts as the plurality of texts used for training a language model.

14. The computer system of claim 11 , wherein the first seed model base is a topic seed base (T) and the second seed model base is a rewriting seed base (R).

15. The computer system of claim 14 , wherein the topic seed base (T) and the rewriting seed base (R) are unigram language models that provide a probability for each token.

16. A computer program product for generating a text, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a computer-implemented method comprising:

obtaining a first text collection matched with a target domain and a second text collection including a plurality of samples each describing rewriting between a first text and a second text having a style different from the first text;

training a text generation model of a learning machine with the first text collection and the second text collection, the text generation model having, in a vocabulary, one or more operation tokens indicating rewriting;

creating a seed model used for obtaining the plurality of texts in conjunction with the text generation model, the seed model created by a nonlinear interpolation of a first seed model base trained with the first text collection and a second seed model base trained with the second text collection, the nonlinear interpolation including a plurality of conditions for giving weights for mixing the first and second seed model bases; and

outputting a plurality of texts obtained from the text generation model the plurality of texts being configured for training a language model of a speech recognition engine.

17. The computer program product of claim 16 , wherein the one or more operation tokens includes one or more insertion tokens and/or one or more deletion tokens, each insertion token indicating insertion of a text part, each deletion token indicating deletion of a text part.

18. The computer program product of claim 16 , the method further comprises:

generating a plurality of tagged texts having at least one operation token using at least the text generation model; and

applying a rewriting rule to each tagged text to obtain a plurality of untagged texts as the plurality of texts used for training a language model.

19. The computer program product of claim 16 , wherein the first seed model base is a topic seed base (T) and the second seed model base is a rewriting seed base (R).

20. The computer program product of claim 19 , wherein the topic seed base (T) and the rewriting seed base (R) are unigram language models that provide a probability for each token.