Ai.ra - The AIRES Expert

Ai.ra is a chatbot (or chatterbot, though we could also call Ai.ra a natural language processor) — that is, a software application used to hold an online text-based chat conversation. Ai.ra was designed to simulate how a human expert would behave during a round of questions and answers (Q&A).

We can classify this type of system (CUS - Conversation Understanding System) into “open domain systems” and “closed domain systems.” A closed-domain system, also known as domain-specific, focuses on a particular set of topics and has limited answers. An open-domain system, on the other hand, in principle understands any topic. For example, GPT-3 — the NLP model produced by OpenAI — is able to “talk about virtually anything.”

Ai.ra is a closed-domain chatbot, so don’t even try asking her the square root of 25 — she won’t be able to help you (but your calculator can!). Ai.ra was developed to provide definitions and answer questions on topics related to artificial intelligence (AI), machine learning, AI ethics, and AI safety, and this is “her domain.”

Ai.ra has three iterations. The first and second were trained via machine learning (using a Bayesian neural network and an LSTM, both trained through supervised learning), while the third was built from predefined rules (n-gram analysis + dictionary search). For its LSTM version, we used the word embeddings made available by the NILC Word Embeddings Repository (Interinstitutional Center for Computational Linguistics).

If you’re not satisfied with Ai.ra’s performance, would like to help us improve our system’s capabilities, or want to report any kind of message produced by Ai.ra, please contact nicholas@airespucrs.org.

Ai.ra was developed by Nicholas Kluge and Carolina Del Pino. For more information visit this repository. To access Ai.ra, follow this link!