CollocAgent: A Corpus-Grounded System for Auditable Collocation Feedback
Year: 2026
Type: software
Abstract: CollocAgent is a standalone system for collocation feedback in second-languageacademic writing. It separates the task into three components with an explicitevidence boundary: a diagnostic component that locates a candidate span andassigns a construction type and severity, a retrieval component that queries alocal corpus index and ranks replacement candidates by attestation, within-windowpair frequency and log-Dice, and an explanation component that may only reportvalues produced by retrieval. Every result carries a provenance record containing the corpus name, theco-occurrence window, the token and sentence counts, the retrieved counts and aSHA-256 fingerprint of the indexed source files. When the index contains nosupport for a recommendation, the system returns an explicit insufficient-evidencestatus instead of a number. The implementation uses only the Python standard library, requires Python 3.10 orlater, and runs from a command-line interface, a local JSON API and a browserclient. Container files are included. Verification of this release: nine unit tests passed, and ten constructed fixtureswere each detected with the expected top-ranked candidate. These are softwarechecks establishing that the components execute as specified. They are not alinguistic evaluation, and no precision, recall, F-score, hallucination rate orhuman rating is claimed. The bundled corpus is synthetic and exists only to verifyexecution; no licensed corpus text or frequency table is included. This deposit contains the software artifact only.
Authors: Nour, Mohammed Abdessamed, Nasri, Kamal
Institution: Centre Universitaire Nour Bachir El Bayadh
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204261589