Skip to content

Latest commit

 

History

History
117 lines (77 loc) · 5.95 KB

discussion.md

File metadata and controls

117 lines (77 loc) · 5.95 KB
Raw

Discussion

Pros and cons

Pros:

  • NaifJs is a low-level framework, dedicated to nodejs/javascript developers (this is also in the cons section).

  • The naif command line program supply a micro-framework of utilities

  • enabling to develop quickly a dialog project.

  • Designer/coder must think, design and implement dialogues as "state machines", explicitly thinking about internal states of a conversation (this is also in the cons section).

  • Very fast run-time (especially using NaifJs as embedded library)

  • Dialog manager is agnostic with respect to inbound/outbound message types. Even if in practice any inbound message (coming from an end-user) is turned to a text, NaifJs can elaborate in theory any kind of inbound message data type and the elaboration and the state transition logic, are all in charge to developer.

    By example a state could take in input a speech (a voice audio file) doing some sort of in-state elaboration without any transcript (ASR) pattern matching, but maybe the audio file could be processed just as a audio object processing.

  • Well matches with stateful-based, rule-based applications as task-based workflows. E.g. Games, interactive stories, real-time/embedded (voice) assistants/cobots.

  • Good for slot-filling / complex slot validation, closed-domain / task completion dialogues.

  • No dependencies from external (cloud-based) "NLU" platforms (if default, but optional regexp pattern matching is acceptable).

  • NaifJs command line utilities help to quickly test dialogues from command line and from a Telegram bot.

  • No machine-learning magics. (this is also in the cons section).

Cons:

  • NaifJs is a low-level framework, dedicated to nodejs/javascript developers (this is also in the pros section).

  • Designer/coder must think, design and implement dialogues as "state machines", explicitly thinking about internal states of a conversation (this is also in the pros section).

  • Not a valid approach for FAQ-like question-answering or open-domain/chit-chat chatbots.

  • No machine-learning magics. (this is also in the pros section).

Food for thought

  • Among researcher in "Conversational AI", State-machines are currently (in 2020), evil! The old way of thinking. And I probably agree! The machine-learning based "language models" along with probabilistic approach that build "virtual"/internal states, seem now a very promising way to overcome the deterministic/hard-coded approach that is the basis of NaifJs. Good stuff has done by RASA opensource (RASA Core) and Alexa Conversations is a promising way to rethink how overcome explicit states development.

  • What seems to me a still valid in NaifJs is the implementation of a multi-turn dialog as a graph of function handlers that represents states of a state machine. I imagine NaifJs dialog units as an "assembler" language, acting as the transpilation of a possible high-level common programming language (that allows conversational designers to quick realize dialog systems).

BackStory

Few years ago, I started working in the idea to implement a human-machine multi-turn conversation using the conceptual model of a state machine. At time, I was focused on how to implement a conversational agent (chatbot) for accomplish workflows/task in e-commerce/e-shopping domain. In general I was focused to realize a simple tool for developers allowing conversational, closed-domain, task-oriented workflows.

In 2016 I realized first prototype in Ruby language (slides of my talk: Naif — Ruby micro framework to build dumb chat machines). The reason of name Naif was because I was aware that the state machine approach do not pretend to be a magic "conversational AI" solution.

As researcher at the Italian National Research Council Institute for Educational Technology (ITD-CNR), in 2019 I ported the original prototype from Ruby to NodeJs, building NaifJs as it is now. I used NaifJs as multi-user dialog manager engine of CPIAbot, a Telegram voice/multi-modal chatbot to help foreigners refugees, students of Italian public adult schools, in learning Italian language L2.

Related academic papers (introducing NaifJs usage):

  • F. Ravicchio, G. Robino, S. Torsani. Un assistente conversazionale a supporto dell’apprendimento dell’italiano L2 per migranti: CPIAbot. Sep 2020. Published in Italian Journal of Educational Technology (IJET). PDF https://doi.org/10.17471/2499-4324/1142

  • F. Ravicchio, G. Robino, G. Trentin. CPIAbot: un chatbot nell’insegnamento dell’Italiano L2 per stranieri. Published in Didamatica 2019 acts, Best Paper Award in section: BYOD. Mobile e Mixed Learning (ISBN 978-88-98091-50-8) PDF p. 77-86.

In September 2020 I decided to opensource my software with a MIT license, allowing people to use the framework and hopefully contribute to improve it.

Acknowledgments

  • In 2016 I supported ChatScript, Bruce Wilcox’s famous language/engine to build chatbots, contributing myself to the open-source project, rewriting ChatScript wiki documentation on the Github repo. ChatScript has been for me an huge source of inspiration.

  • Telegram is a beautiful platform to experiments multimodal bots, due to the excellent Telegram Bot API. I used Telegram also to develop CPIAbot before mentioned, and NaifJs supply an adapter to the platform and the rapid test tool: naif telegram. A special thanks goes to the author of nodejs library node-telegram-bot-api I used also in this project.

top | home | index