Methods and systems for producing neural sequential models

1. A natural language processing method for producing a normalized sequential model using a processor, the method comprising:

providing a sequential energy-based model computed by a parameterized neural network, the sequential energy-based model defining an unnormalized probability distribution over a target sequence of text for a context source of text; and

producing the normalized sequential model by projecting the sequential energy-based model onto a target autoregressive model that approximates a normalized distribution associated with the sequential energy-based model;

wherein the normalized sequential model is adapted to generate a target sequence of text from a context sequence of text; and

wherein the normalized sequential model is configured to perform one of language modeling, dialog, natural language generation, and machine translation.

2. The method of claim 1 , wherein the sequential energy-based model comprises a locally normalized component and a global component.

3. The method of claim 1 , wherein the sequential energy-based model comprises:

an autoregressive model generating a sequence given the context source, the autoregressive model being locally and globally normalized, the autoregressive model being parameterized by a first set of parameters; and

an unnormalized energy-based potential that is combined with the autoregressive model, the energy-based potential modeling expected global features of the target sequence given the context source, the energy-based potential being parameterized by a second set of parameters.

4. The method of claim 3 , wherein the sequential energy-based model comprises a global autoregressive model (GAM).

5. The method of claim 3 , wherein the global features are a priori features that comprise one of a length ratio between source and target, a non-duplication of text, and a consistency with predetermined terminology.

6. The method of claim 3 , wherein the autoregressive model comprises one of a recurrent neural network (RNN), a long short-term memory (LSTM), a ConvSS, and a Transformer.

7. The method of claim 3 , wherein the autoregressive model models incremental aspects of sequential data.

8. The method of claim 3 , wherein the energy-based potential is a log-linear potential.

9. The method of claim 3 , wherein the energy-based potential comprises:

a feature vector representing predefined real features of sequences; and

a weight vector defined by the second set of parameters.

10. The method of claim 9 , wherein the feature vector and the weight vector are of the same dimension and are combined by scalar product.

11. The method of claim 9 , wherein said providing the sequential energy-based model comprises:

training the autoregressive model using a dataset of text sequences to obtain an autoregressive factor r over the first set of parameters; and

using the autoregressive factor r as a proposal distribution, training the sequential energy-based model by fitting the weight vector to the dataset of text sequences using a maximum likelihood approach.

12. The method of claim 11 , wherein said training the sequential energy-based model comprises performing stochastic gradient descent (SGD) over the weight vector.

13. The method of claim 12 , wherein said performing SGD over the weight vector comprises:

computing empirical moments based on an expectation of the features in the feature vector relative to the dataset;

computing model moments based on an expectation of the features in the feature vector predicted by the sequential energy-based model;

computing a gradient based on said computed empirical and model moments; and

updating the weight vector based on said computed gradient.

14. The method of claim 13 , wherein said computing the model moments comprises:

producing random samples of target sequences from the autoregressive model; and

computing a mean of the feature vector over the produced random samples.

15. The method of claim 14 , wherein said producing random samples of target sequences uses a Monte-Carlo sampling technique.

16. The method of claim 15 , wherein said producing random samples of target sequences uses rejection sampling (rs).

17. The method of claim 16 , wherein the rejection sampling comprises determining whether to use respective ones of the produced random samples based on upper bounds of a ratio between an unnormalized probability distribution and the proposal distribution r.

18. The method of claim 13 , wherein said computing the model moments comprises:

producing random samples of target sequences from the autoregressive model; and

directly estimating the expectation of the features using self-normalized importance sampling (snis).

19. The method of claim 1 , wherein the produced normalized sequential model is an autoregressive model.

20. The method of claim 19 , wherein the autoregressive model comprises one of a recurrent neural network (RNN), a long short-term memory (LSTM), a ConvSS, and a Transformer.

21. The method of claim 19 ,

wherein the normalized sequential model provides a policy defined by parameters; and

wherein said producing the normalized sequential model comprises:

training the normalized sequential model to minimize cross-entropy between the policy and the normalized distribution associated with the sequential energy-based model.

22. The method of claim 21 , wherein the normalized distribution associated with the sequential energy-based model is associated with the defined unnormalized probability distribution by a partition function.

23. The method of claim 21 , wherein said training the normalized sequential model uses a distillation approach.

24. The method of claim 21 , wherein said training the normalized sequential model comprises:

producing a set of random samples of target sequences using the provided sequential energy-based model; and

training the normalized sequential model using at least a portion of the produced set of random samples.

25. The method of claim 24 , wherein said training the normalized sequential model comprises performing supervised log-likelihood stochastic gradient descent (SGD) using the portion of the produced set of random samples.

26. The method of claim 24 , wherein said training the normalized sequential model uses a training dataset supplemented by the portion of the produced set of random samples.

27. The method of claim 26 , wherein the training dataset comprises samples from an existing training dataset used to train the provided sequential energy-based model combined with the portion of the produced set of random samples.

28. The method of claim 24 , wherein said producing a set of random samples of target sequences uses a Monte-Carlo sampling technique.

29. The method of claim 28 , wherein said producing a set of random samples of target sequences uses rejection sampling (rs).

30. The method of claim 29 , wherein the sequential energy-based model comprises:

an autoregressive model generating a sequence given the context source, the autoregressive model being locally normalized, the autoregressive model being parameterized by a first set of parameters; and

an unnormalized energy-based potential that is combined with the autoregressive model, the energy-based potential modeling expected global features of the target sequence given the context source, the energy-based potential being parameterized by a second set of parameters;

wherein said providing the sequential energy-based model comprises:

training the autoregressive model using a dataset of text sequences to obtain an autoregressive factor r over the first set of parameters; and

using the autoregressive factor r as a proposal distribution, and training the sequential energy-based model by fitting the weight vector to the dataset of text sequences using a maximum likelihood approach.

31. The method of claim 30 , wherein the rejection sampling comprises determining whether to use respective ones of the produced random samples based on upper bounds of a ratio between an unnormalized target distribution and the proposal distribution r.

32. The method of claim 31 , wherein said training the normalized sequential model further comprises:

determining whether an acceptance rate of the rejection sampling is improving according to predetermined criteria;

if it is determined that the acceptance rate is improving:

retraining the autoregressive model to obtain an updated autoregressive factor r over the first set of parameters;

using the updated autoregressive factor r as a proposal distribution, and retraining the sequential energy-based model by fitting the weight vector to the dataset of text sequences; and

further training the normalized sequential model based on the unnormalized distribution associated with the retrained sequential energy-based model.

33. The method of claim 21 , wherein said training the normalized sequential model uses a distributional policy gradient approach that seeks to maximize an expectation of a reward based on the unnormalized probability distribution and the policy.

34. The method of claim 21 , wherein said training the normalized sequential model comprises:

generating sample target sequences from the policy;

processing the generated sample target sequences using the policy and the unnormalized sequential EBM to generate probability distributions;

computing a gradient based on the generated probability distributions; and

updating the policy using stochastic gradient descent (SGD) based on the computed gradient.

35. The method of claim 34 , wherein the computed gradient is based on a ratio between the generated probability distributions.

36. The method of claim 34 , further comprising:

generating sample target sequences from the updated policy;

processing the generated sample target sequences using the updated policy and the unnormalized probability distribution to generate probability distributions;

computing a gradient based on the generated probability distributions; and

further updating the policy using stochastic gradient descent (SGD) based on the computed gradient.

37. The method of claim 21 , wherein said training the normalized sequential model comprises:

providing an initial policy for generating sample target sequences;

generating sample target sequences from the initial policy;

processing the generated sample target sequences using the policy and the unnormalized probability distribution to generate probability distributions;

computing a gradient based on the generated probability distributions; and

updating the policy using stochastic gradient descent (SGD) based on the computed gradient.

38. The method of claim 37 , further comprising:

periodically determining whether to update the initial policy for generating new sample target sequences.

39. The method of claim 38 , wherein said periodically determining is based on comparing a measured perplexity of the initial policy and the policy on a validation dataset.

40. The method of claim 37 ,

wherein the sequential energy-based model comprises:

an autoregressive model defined by an autoregressive factor r and parameterized by a first set of parameters; and

an unnormalized energy-based potential that is combined with the autoregressive model, the energy-based potential modeling expected global features of the target sequence given the context source, the energy-based potential being parameterized by a second set of parameters; and

wherein the initial policy is based on the autoregressive factor r.

41. A natural language processing method of generating a neural sequential model using a processor, the method comprising:

training an autoregressive model to obtain an autoregressive factor r over a first set of parameters using a dataset of text sequences, the autoregressive model being locally and globally normalized; and

using the trained autoregressive model as a proposal distribution, and training an unnormalized energy-based potential that is combined with the autoregressive model, the energy-based potential modeling expected global features of a target sequence of text given a context source of text, the energy-based potential being parameterized by a second set of parameters;

wherein the neural sequential model is adapted to generate an output target sequence of text given an input context source of text; and

wherein the neural sequential model is configured to perform one of language modeling, dialog, natural language generation, and machine translation.

42. The method of claim 41 , wherein the energy-based potential comprises a feature vector representing predefined real features of sequences and a weight vector defined by the second set of parameters; and

wherein said training an unnormalized energy-based potential comprises fitting the weight vector to the dataset of text sequences using a maximum likelihood approach.

43. The method of claim 42 , wherein said fitting the weight vector to the dataset of text sequences comprises:

computing empirical moments based on an expectation of the features in the feature vector relative to the dataset;

computing model moments based on an expectation of the features in the feature vector predicted by a sequential energy-based model;

computing a gradient based on said computed empirical and model moments; and

updating the weight vector based on said computed gradient.

44. The method of claim 43 , wherein said computing the model moments comprises:

producing random samples of target sequences from the autoregressive model; and

computing a mean of the feature vector over the produced random samples.

45. The method of claim 44 , wherein said producing random samples uses rejection sampling.

46. The method of claim 43 , wherein said computing the model moments comprises:

producing random samples of target sequences from the autoregressive model; and

directly estimating the expectation of the features using self-normalized importance sampling (snis).

47. A natural language processing method for generating a target sequence of text from a context source of text using a processor, the method comprising:

providing a normalized sequential model for generating text trained by a method comprising:

providing a sequential energy-based model computed by a parameterized neural network, the sequential energy-based model defining an unnormalized probability distribution over a target sequence of text for a context source that is used to produce the target sequence; and

producing the normalized sequential model for generating text by projecting the sequential energy-based model onto a target autoregressive model that approximates a normalized distribution associated with the sequential energy-based model;

receiving, by the provided normalized sequential model, the context source of text; and

generating, by the provided normalized sequential model, the target sequence of text;

wherein the normalized sequential model is configured to perform one of language modeling, dialog, natural language generation, and machine translation.

48. The method of claim 47 , further comprising:

one of storing the target sequence of text, printing the target sequence of text, transmitting the target sequence of text to a computing device, and providing the target sequence of text for display on a display.