1. Introduction
Seven Code search tool 分类
- text-based code search.
- I/O example code search.
- API-based code search.
- ADECK [147]
- code clone search.
- binary code search.
Source code 会编译成不同二进制代码，对二进制代码检索. - UI search.
使用 UI 草图进行搜索pix2code [14] - programming video search.
在视频中搜索相关的代码。[10]
67个code search tool 仅仅只有12个开源。
Evaluation
主要是 ranking metrics. MRR (Mean Reciprocal Rank), Precision.
最近几篇文章还报告了 R@1, R@5 这些。
Challenges
- Standard Benchmark.
- Inprove Machine Learning Models. 训练数据，跨模态表示，loss function.
- Model Fusion. DL model, traditional IR model…
- Cross-Language Searches.
- Search Tasks. UI codes, code used in programming videos. 这些新任务。
2. Background
一个通常的 working flow, 有7个模块。
- Query. 主要是自然语言。 [56, 85, 103] 支持结构化的 code-based query.
- Codebase. 不同语言，不同来源。
- Code Analysis Technique. 如何从 code 中挖掘更多programming knowledge? AST (抽象语法树) [136, 145]. CFG (控制流程图) [21, 130]. Call Graph (变量方法的调用关系) [74, 75].
- Modeling Technique.
- Traditional IR.
- 启发式、手工设计特征、matching score.
- ML model
- Auxiliary Technique (辅助技术，可以调研一下).
- Query reformulation.
- Code Cluster.
- Feedback learning.
- Evaluation Method.
- Performance Measures.
3. Methodology
Empirical study: Analyze the search logs of existing code search tools [6, 7, 26, 37, 38, 89, 104, 105, 138, 142]. 这个挺有意思的，分析真实场景中的 search logs.
5 现有代码搜索工具中的关键组件
5.1 Code Analysis Technique
Semantics Analysis 解析程序组成和依存关系
AST (抽象语法树) [136, 145]. CFG (控制流程图) [21, 130]. Call Graph (变量方法的调用关系) [74, 75].
Relevancy Analysis: code 与 query 的关系。
- Code difference. 识别 query 和 code 相似和不同的部分，缩小搜索范围
- static code slice. 过滤不相关的代码。
- symbolic execution.
5.2 Modeling Techniques
IR Models
- TF-IDF, BM25
- Boolean models 支持使用 “AND” 操作符。[86, 143]
Heuristic Models (启发式)
ML Models
5.3 Auxiliary techniques (辅助技术)
Inverted Index
Query Reformulation: Expand & replace
Code Clustering
Feedback Learning
7 Code Search Evaluation
8. CHALLENGES AND OPPORTUNITIES
Challenge 1: Diversity of the Codebase.
Challenge 2: Limited Queries.
Challenge 3: Model Construction Issues. DL model 参数过多，对训练数据的质量有要求。
Challenge 4: Evaluation Issues.
Challenge 5: Limited Performance Measures.
Challenge 6: Replication Issues.
Opportunity 1: Better Benchmarks.
Opportunity 2: DL-Based Model with Big Data.
Opportunity 3: Fusion of Different Types of Models.
Opportunity 4: Multi-Language Tool.
Opportunity 5: New Code Search Tasks.
