Text style transfer using reinforcement learning
A computer-implemented method is provided for transferring a target text style using Reinforcement Learning (RL). The method includes pre-determining, by a Long Short-Term Memory (LSTM) Neural Network (NN), the target text style of a target-style natural language sentence. The method further includes transforming, by a hardware processor using the LSTM NN, a source-style natural language sentence into the target-style natural language sentence that maintains the target text style of the target-style natural language sentence. The method also includes calculating an accuracy rating of a transformation of the source-style natural language sentence into the target-style natural language sentence based upon rewards relating to at least the target text style of the source-style natural language sentence.
1. A computer-implemented method for transferring a target text style using Reinforcement Learning (RL), comprising:
pre-determining, by a Long Short-Term Memory (LSTM) Neural Network (NN), the target text style of a target-style natural language sentence;
transforming, by a hardware processor using the LSTM NN, a source-style natural language sentence into the target-style natural language sentence that maintains the target text style of the target-style natural language sentence; and
calculating an accuracy rating of a transformation of the source-style natural language sentence into the target-style natural language sentence based upon rewards relating to at least the target text style of the source-style natural language sentence,
wherein the rewards comprise style rewards which are determined using a style classifier built upon a bidirectional recurrent neural network with an attention mechanism.
2. The computer-implemented method of claim 1 , wherein the rewards further comprise, semantic rewards and fluency rewards, and wherein said calculating step comprises evaluating the target-style natural language sentence with respect to content preservation, target text style, and fluency using the style rewards, the semantic rewards, and the fluency rewards, respectively.
3. The computer-implemented method of claim 2 , wherein the semantic rewards are determining using a semantic module configured to determine a Word Mover's Distance (WMD) as an embedding-based similarity metric calculated as a sum of distances between co-occurring words in the source-style natural language sentence relative to the target-style natural language sentence.
4. The computer-implemented method of claim 2 , wherein the fluency rewards are determined using a Recurrent Neural Network (RNN)-based language model.
5. The computer-implemented method of claim 1 , wherein the style classifier is pre-trained on a source and target corpus in style classification.
6. The computer-implemented method of claim 1 , wherein the style classifier is adversarially trained on target-style natural language sentences generated by said transforming step.
7. The computer-implemented method of claim 1 , further comprising guaranteeing, by a LSTM-based style discriminator, a style transfer strength of the transformation above a threshold amount.
8. The computer-implemented method of claim 1 , further comprising guaranteeing, by a sentence discriminator, a content preservation between the source-style natural language sentence and the target-style natural language sentence.
9. The computer-implemented method of claim 1 , further comprising guaranteeing, by a Recurrent Neural Network (RNN)-based language model, a fluency of the target-style natural language sentence.
10. The computer-implemented method of claim 1 , guiding, based on the rewards, the generator in performing a subsequent transformation that achieves a higher accuracy rating, responsive to the accuracy rating being below a threshold amount.
11. The computer-implemented method of claim 1 , repeating said guiding step until the accuracy rating is equal to or greater than the threshold amount.
12. A computer program product for transferring a target text style using Reinforcement Learning (RL), 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 method comprising:
pre-determining, by a Long Short-Term Memory (LSTM) Neural Network (NN), the target text style of a target-style natural language sentence;
transforming, by the LSTM NN, a source-style natural language sentence into the target-style natural language sentence that maintains the target text style of the target-style natural language sentence; and
calculating an accuracy rating of a transformation of the source-style natural language sentence into the target-style natural language sentence based upon rewards relating to at least the target text style of the source-style natural language sentence,
wherein the rewards comprise style rewards which are determined using a style classifier built upon a bidirectional recurrent neural network with an attention mechanism.
13. The computer program product of claim 12 , wherein the rewards further comprise semantic rewards and fluency rewards, and wherein said calculating step comprises evaluating the target-style natural language sentence with respect to content preservation, target style, and fluency using the style rewards, the semantic rewards, and the fluency rewards, respectively.
14. The computer program product of claim 12 , further comprising guaranteeing, by a LSTM-based style discriminator, a style transfer strength of the transformation above a threshold amount.
15. The computer program product of claim 12 , further comprising guaranteeing, by a sentence discriminator, a content preservation between the source-style natural language sentence and the target-style natural language sentence.
16. The computer program product of claim 12 , further comprising guaranteeing, by a Recurrent Neural Network (RNN)-based language model, a fluency of the target-style natural language sentence.
17. The computer program product of claim 12 , further comprising guiding, based on the rewards, the generator in performing a subsequent transformation that achieves a higher accuracy rating, responsive to the accuracy rating being below a threshold amount.
18. The computer program product of claim 17 , further comprising repeating said guiding step until the accuracy rating is equal to or greater than the threshold amount.
19. A computer processing system for transferring a target text style using Reinforcement Learning (RL), comprising:
a memory device including program code stored thereon;
a hardware processor, operatively coupled to the memory device, and configured to run the program code stored on the memory device to
pre-determine, using a Long Short-Term Memory (LSTM) Neural Network (NN), the target text style of a target-style natural language sentence;
transform, using the LSTM NN, a source-style natural language sentence into the target-style natural language sentence that maintains the target text style of the target-style natural language sentence; and
calculate an accuracy rating of a transformation of the source-style natural language sentence into the target-style natural language sentence based upon rewards relating to at least the target text style of the source-style natural language sentence,
wherein the rewards comprise style rewards which are determined using a style classifier built upon a bidirectional recurrent neural network with an attention mechanism.