Automatic generation of assert statements for unit test cases

1. A system comprising:

one or more processors; and

a memory that stores one or more programs that are configured to be executed by the one or more processors, the one or more programs including instructions to perform actions that:

pre-train a neural transformer model with attention with a first unsupervised training dataset to learn to predict natural language text, the first unsupervised training dataset including a plurality of sequences of natural language text;

pre-train the neural transformer model with attention with a second unsupervised training dataset to learn to predict a sequence of source code, the second unsupervised training dataset including a plurality of sequences of source code from source code methods of a programming language;

fine-tune the neural transformer model with attention with a supervised training dataset to learn to predict an assert statement for a given test method, the supervised dataset including a plurality of test-assert triplets, wherein a test-assert triplet includes a test method annotated with an asset placeholder, a focal method, and an assert statement, wherein the fine-tuning is a translation task with a training objective that learns to translate a source code snippet, s i , to a corresponding assert statement, a i , as a conditional probability P(a i |tm i ′+fm i ), where tm i ′ is a corresponding test method annotated with an assert placeholder and fm i is a corresponding focal method; and

deploy the neural transformer model with attention in a software development environment to generate a target assert statement given a specified test method and a specified focal method.

2. The system of claim 1 , wherein the one or more programs include further instructions to perform actions that:

mine a source code repository for a plurality of test methods having a single assert statement; and

generate the plurality of test-assert triplets from the plurality of test methods.

3. The system of claim 1 , wherein the one or more programs include further instructions to perform actions that:

apply a span mask to sequences of the second unsupervised training dataset; and

wherein the pre-trained neural transformer model with attention is trained to learn syntax of the programming language of the source code methods by replacing the span mask of a sequence with appropriate source code.

4. The system of claim 1 , wherein the one or more programs include further instructions to perform actions that:

apply a span mask to sequences of the first unsupervised training dataset; and

wherein the pre-trained neural transformer model with attention is trained to learn semantics of the natural language by replacing the span mask in a sequence with appropriate words of the natural language text.

5. The system of claim 1 , wherein the natural language text is English language and the source code of the second unsupervised training dataset and the source code of the supervised training dataset is written in a same programming language.

6. A method performed by one or more computing devices, comprising:

pre-training a neural transformer model with attention with a first unsupervised training dataset, the first unsupervised training dataset including a plurality of sequences of natural language text;

pre-training the neural transformer model with attention with a second unsupervised training dataset, the second unsupervised training dataset including a plurality of sequences of source code;

fine-tuning the neural transformer model with attention with a supervised training dataset, wherein the fine-tuning is a translation task with a training objective where the neural transformer model with attention learns to translate a source code snippet, s i , to an assert statement, a i , as a conditional probability P(a i |tm i ′+fm i ), where tm i ′ is a test method annotated with an assert placeholder, fm i is a corresponding focal method, and a i is a corresponding assert statement; and

deploying the neural transformer model with attention in a software development environment to automatically generate a target assert statement for a test method of a focal method.

7. The method of claim 6 , further comprising:

mining a source code repository for a plurality of test methods having a single assert statement.

8. The method of claim 7 , further comprising:

identifying the focal method subject to a select test method by identifying a latest method call in the select test method.

9. The method of claim 6 , further comprising:

applying a span mask to sequences of the second unsupervised training dataset; and

wherein the pre-trained neural transformer model with attention is trained to learn syntax of a programming language of the source code methods by replacing the span mask of a sequence with appropriate source code.

10. The method of claim 6 , further comprising:

applying a span mask to sequences of the first unsupervised training dataset; and

wherein the pre-trained neural transformer model with attention is trained to learn semantics of the natural language by replacing the span mask in a sequence with appropriate words of the natural language.

11. The method of claim 6 , wherein the neural transformer model with attention includes a plurality of encoder blocks and a plurality of decoder blocks, an encoder block having a multi-head attention layer, a decoder block having a masked multi-head attention layer.

12. The method of claim 11 , wherein at least one of the plurality of encoder blocks includes a gaussian error linear unit activation function.

13. The method of claim 6 , wherein the natural language text is English, and

wherein the source code of the second unsupervised training dataset and the supervised training dataset are written in a same programming language.