Conversations From Make-Believe: An Attentive Encoder–Decoder Chatbot Trained on Scripted Dialogue

This paper builds a fully neural, open-domain chatbot that learns to respond like a conversational partner rather than a search engine, using an encoder-decoder with bidirectional recurrent memory and a token-level attention mechanism to track context across turns. A large corpus of fictional, face-toface exchanges is cleaned into paired utterances, tokenized with subword units, and used to train the model end-to-end, yielding fluent, on-topic replies without hand-crafted rules or retrieval templates. Training and validation use sequence likelihood objectives, while quality is assessed with both automatic indicators (e.g., perplexity and n-gram overlap) and qualitative probes that test specificity, coherence, and avoidance of generic ”safe” answers. A lightweight desktop interface demonstrates interactive behavior by surfacing multiple candidate responses from beam search and selecting among them for variety and fit. The study discusses common failure modes in open-domain chat (repetition, blandness, drift) and outlines practical remedies—data curation, decoding constraints, and post-training reward signals—to further align responses with human conversational expectations.