Recoleta
GitHub
Source note

CoCoMUT: A Tool for Code-Context Mining and Automated Dataset Generation

Authors
Alessandro Botta, Shiven Garisa, Jaya Vardhini Akurathi, Ahsanul Ameen Sabit, Trey Woodlief, Soneya Binta Hossain
Source
arXiv
Published
Collection
Software Intelligence
Code Context MiningJava Static AnalysisDataset GenerationCall Graph ReconciliationSoftware Engineering AI
Open arXivMarkdown source

Summary

CoCoMUT extracts method-level Java context and writes versioned JSONL datasets for code-intelligence research. It matters because LLM and learned software-engineering tools often need callers, callees, class context, docs, and metadata that are hard to collect consistently.

Problem

  • Software assistants need context beyond a single method body, including enclosing class details, Javadoc, callers, callees, type hierarchy, and structural metrics.
  • Java context extraction is hard to reproduce because source declarations, bytecode signatures, build metadata, dependencies, overloads, generics, nested types, and synthetic methods do not line up cleanly.
  • Task-specific data extractors make comparisons harder and can introduce hidden assumptions about method identity and context boundaries.

Approach

  • CoCoMUT builds a Spoon source model and records stable source method URIs, Javadoc, annotations, hierarchy data, source positions, fields, overloads, sibling methods, and metrics.
  • It builds a SootUp static call graph from compiled project bytecode and dependencies, using RTA by default or CHA when selected.
  • It matches bytecode call targets to source methods only when there is a unique match; ambiguous and unmatched targets keep the bytecode target_uri and explicit resolution metadata.
  • It writes one deterministic JSONL record per selected method, with source, local class, documentation, caller/callee, provenance, and confidence fields.

Results

  • On 20 real Java repositories, split into 10 Maven and 10 Gradle projects, CoCoMUT completed build, bytecode availability, call-graph construction, and JSONL emission for all 20.
  • It emitted 56,512 method-context records and 386,048 serialized caller/callee entries.
  • Every caller/callee entry preserved a bytecode target_uri; 294,242 of 300,743 recognized project targets were linked to a source method_uri.
  • Source-bytecode reconciliation reached 97.8% overall, with 98.4% on Maven projects and 93.8% on Gradle projects; CoCoMUT abstained on 6,501 project targets.
  • Runtime across repositories was 9/65/275 seconds for min/average/max; Maven averaged 88 seconds and Gradle averaged 42 seconds.
  • In a manual audit of 200 records across 10 repositories and 406,312 production SLOC, 198 records passed all applicable checks, giving a 99.0% pass rate; annotator agreement was 100.0% with Cohen’s κ = 1.00.

Link