Development of an Academic Chatbot Based on Natural Language Processing for Student Information Services at Universitas Bina Darma

Abstract

Purpose: Repetitive academic inquiries and substantial linguistic variation in informal Indonesian student messages can reduce the consistency and timeliness of university information services. This study develops and evaluates an academic chatbot for student information services at Universitas Bina Darma through WhatsApp. Methods/Study design/approach: An applied engineering-research approach was employed. An exploratory questionnaire involving 100 active students was used to identify dominant information needs and representative query expressions, while verified institutional documents were used separately to construct controlled service responses. The implemented architecture integrates Baileys as the WhatsApp messaging gateway, Node.js/Express as the orchestration layer, Dialog flow ES for intent and entity recognition, and MySQL for conversation logging and monitoring. The final agent contains 129 custom intents, 22 entities, and 2,572 training-phrase entries. An operational confidence threshold of 0.70 was applied to route uncertain inputs to fallback handling. Evaluation was conducted at four layers: 13 black-box functional scenarios, 24 intent cases representing major service families and out-of-scope input, 15 response-conformity cases, and six end-to-end integration scenarios. Result/Findings: All black-box scenarios produced the expected outputs. Intent agreement reached 95.83%, response conformity reached 86.67%, and all integration scenarios were successful. The difference between intent agreement and response conformity indicates that correct intent recognition does not necessarily guarantee an equally specific or institutionally appropriate answer when response coverage is incomplete. Novelty/Originality/Value: The contribution of this study is therefore positioned as applied system engineering rather than a new NLP algorithm. It demonstrates a deployable, confidence-controlled architecture that separates informal-query recognition from verified institutional delivery and evaluates recognition, answer conformity, and technical integration as distinct quality dimensions. Broader intent coverage, human-user evaluation, threshold calibration, load testing, and dynamic integration with institutional information systems remain necessary for wider deployment.