A chatbot that can do math (because neural nets can't)

What it does ChatLearner trains a conversational agent in TensorFlow 1.4–1.11 using the then-new NMT seq2seq architecture. The hook: it layers hand-coded rules on top so the bot can handle tasks neural models alone botch, like arithmetic, date lookups, and story retrieval. A curated “Papaya” dataset mixes handcrafted persona data with cleaned Cornell movie dialogs and Reddit comments. SOAP and REST APIs wrap the model, with a Java GUI client included.

The interesting bit The author treats deep learning as a language-modeling layer, not an oracle. The rules aren’t an afterthought—they’re the admission that “no matter how powerful a deep learning model can be, it cannot even answer questions requiring simple arithmetic calculations.” The dataset is also unusually opinionated: the bot is trained to play a polite, philosophical 9-year-old named Papaya.

Key highlights

• Custom “Papaya” dataset with persona-consistent handcrafted samples plus cleaned Cornell and Reddit data
• Rule integration for math, time/date, and random content retrieval (stories, jokes)
• In-graph lowercasing solution for TensorFlow’s tf.data TextLineDataset
• Both SOAP and REST web service wrappers, with a Java GUI reference implementation
• Legacy seq2seq branch available for comparison

Caveats

• Locked to TensorFlow 1.4–1.11; the tf.data API changes in 1.12 break compatibility
• Requires manual PYTHONPATH setup and careful vocab.txt consistency between training and inference
• The author warns that vocabulary size, not training sample count, is the real capacity bottleneck

Verdict Worth a look if you’re building a chatbot and need a practical example of hybrid neural/rule architecture—especially for constrained domains. Skip if you need modern TensorFlow or an out-of-the-box production system; this is a research/educational snapshot from 2017-era tooling.