Konnect: A Distributed Microservices Architecture for Real-Time Chat with Multi-Task NLP and Context-Aware Toxicity Dampening

This paper presented Konnect, a microservices-based messaging platform that integrates dual transformer classifiers with a post-classification dampening layer for real-time content moderation. The dampening layer addresses a practical gap in toxicity classification: the tendency of pre-trained models to assign elevated scores to colloquial negative expressions that carry no interpersonal hostility. By examining whether the input text targets a human referent, the layer attenuates false positives by 38 percent while preserving 97.2 percent of true positive detections. The Kafka-based pub/sub architecture enables horizontal scaling of the inference pipeline, with near-linear throughput gains demonstrated up to four consumer replicas. The composite moderation score, computed over a Redis-backed sliding window with a safety floor mechanism,  drives  a  four-tier  graduated  escalation framework that balances automated enforcement with moderator authority

[1]  A. Vaswani et al., “Attention Is All You Need,” in Proc. NeurIPS, 2017, pp. 5998–6008. [2]   J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” in Proc. NAACL-HLT, 2019, pp. 4171–4186. [3]  Y. Liu et al., “RoBERTa: A Robustly Optimized BERT Pretraining Approach,” arXiv preprint arXiv:1907.11692, 2019. [4]   L. Hanu and Unitary team, “Detoxify,” GitHub repository, 2020. [Online]. Available: https://github.com/unitaryai/detoxify [5]  F. Barbieri, J. Camacho-Collados, L. Neves, and L. Espinosa-Anke, “TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification,” in Findings of EMNLP, 2020, pp. 1644–1650. [6]  J. Kreps, N. Narkhede, and J. Rao, “Kafka: A Distributed Messaging System for Log Processing,” in Proc. NetDB Workshop, 2011. [7]  M. Kleppmann and J. Kreps, “Kafka, Samza and the Unix Philosophy of Distributed Data,” IEEE Data Engineering Bulletin, vol. 38, no. 4, pp. 4–14, 2015. [8]   K. Crawford and T. Gillespie, “What is a flag for? Social media reporting tools and the vocabulary of complaint,” New Media & Society, vol. 18, no. 3, pp. 410–428, 2016. [9]  S. Jhaver, I. Birman, E. Gilbert, and A. Bruckman, “Human-Machine Collaboration for Content Regulation,” ACM Trans. Comput.-Hum. Interact., vol. 26, no. 5, pp. 1–35, 2019. [10]   L. Dixon, J. Li, J. Sorensen, N. Thain, and L. Vasserman, “Measuring and Mitigating Unintended Bias in Text Classification,” in Proc. AAAI/ACM Conf. on AI, Ethics, and Society, 2018, pp. 67–73. [11] M. Sap, D. Card, S. Gabriel, Y. Choi, and N. A. Smith, “The Risk of Racial Bias in Hate Speech Detection,” in Proc. ACL, 2019, pp. 1668–1678.