This is the official GitHub repository for the following paper:

DAInfer: Inferring API Aliasing Specifications from Library Documentation via Neurosymbolic Optimization (FSE 2024)

Step 1: Install the packages in the requirements.txt file.

pip install -r requirements.txt --use-deprecated=legacy-resolver

Step 2: Download the necessary resources for the tagging model. Execute the following code in Python environment:

> import nltk
> nltk.download('brown')
> nltk.download('universal_tagset')
> nltk.download('averaged_perceptron_tagger')

Step 3: Add your openai key in src/config.py

global_openai_key = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

If you want to run the DAInfer with the default setting, you can simply execute the following command:

cd src
python run.py 1 1 0.7 0.7 --lazy

If you want to apply the self-consistency to the two-staged prompting and adjust the temperatures, you can change the four arguments passed to main.py. For example, if you want to set K = 5 for both the two stages and the temperatures are both set to 1.0, you can execute the following command:

cd src
python run.py 5 5 1.0 1.0 --lazy

If you want to disable the lazy strategy in the neural-symbolic optimization, you can replace the -lazy flag with -eager.

cd src
python run.py 1 1 0.7 0.7 --eager

ATTENTION: The eager mode would invoke OpenAI API for a large number of methods in the library documentation, which may cost a lot of money. We recommend you to use the lazy mode to avoid the high cost.

We parse the documentations of Java classes used for the evaluation and provide the documentation model in the directory data/javadoc/benchmark. All the analyzed Java classes are listed in the json file data/javadoc/evalSubject.json. All the methods and their semantic descriptions are summarized in the json file benchmark_fullMethodDoc.json.

We provide docParser.py in the directory src. You can utilize it to extract the documentation of the Java library you focus on and run DAInfer to infer the API aliasing specifications for the library.

The directory oracle contains the three sources of oracles, including the ones specified in FlowDroid and USpec and the manually specified ones. The generated models of atlas is stored in baseline/atlas/models. We referred to all the other oracles when we manually specified ours, which is stored in the directory oracle/ManualOracle.

The output of DAInfer is stored in the directory data/output. Specifically, the directory autoPrompt_FourTypes_m_n_temp1_temp2 stores the detailed output of DAInfer under the following setting of the two-staged prompting.

• m and n: The numbers of sampling in self-consistency prompting for the typical verb retrieval and the memory operation abstraction, respectively.

• temp1 and temp2: The temperatures of the prompting for the typical verb retrieval and the memory operation abstraction, respectively.

In each directory autoPrompt_FourTypes_m_n_temp1_temp2, the three sub-directories contain the following (intermediate) results:

• The directory methoInfo: contains the json files that record the method type signatures and the results of memory operation abstraction.

• The directory prompt contains the prompting results of the first stage, i.e., the typical verbs indicating the four kinds of memory operations.

• The directory inferResult: The results of the API aliasing specification inference. Specifically, the json file benchmark_inferredSpecs.json contains all the inferred specifications. The json file benchmark_CHADic.json maintains the class hierachy relation. The json file benchmark_NNSet.json stores the identified semantic units, which are named entities in the method names. The json file benchmark_retArgSpecCandidate.json stores all the API aliasing specifications that satisfy the degree constraint and validity constraint but not necessarily statify the semantic unit constraint and memory operation constraint.

• The directory compareResult: The comparison results of the inferred specifications and the oracle in the directory oracle. Specifically, the json files inferredCorrectSpecs.json and inferredWrongSpecs.json list correct specifications, and wrong specifications related to the sampled 60 Java classes, respectively. The files missedFlowDroidSpecs.json and missedUSpecArgSpec.json store the specifications that are not inferred by DAInfer but are labeled by FlowDroid's developers or inferred by USpec, respectively.

Apart from atlas and USpec, we propose an LLM-based API aliasing specification inference approach as a baseline. To run the baseline, you also need to fill the openai key in the following statement:

openai_key_baseline = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

Then you can simply runn the baseline by executing the following command:

cd baseline/LLM-Alias
python ChatGPTBaseline.py

The results of the baseline is dumped to output.json file in the directory baseline/LLM-Alias.

If you use DAInfer in your research, please cite the following paper:

@inproceedings{DAInfer,
  author={Chengpeng Wang 
  and Jipeng Zhang 
  and Rongxin Wu 
  and Charles Zhang},
  title={DAInfer: Inferring API Aliasing Specifications from Library Documentation via Neurosymbolic Optimization},
  booktitle={The Proceedings of the ACM on Software Engineering},
  volume={1},
  number={FSE},
  year={2024},
  doi = {10.1145/3660816},