How I Used Deep Learning to Train a Chatbot to Talk Like Me | Hacker News Reader

How I Used Deep Learning to Train a Chatbot to Talk Like Me

adeshpande3.github.io

220 pointsadeshpandedsfd33 comments

It's always nice to read the details of people building real applications, even if it's not a homerun.

BTW: just blurring people's names using a gaussian blur operator is not enough, if your radius is too small. From the screenshots of chats, you can tell names like "Arvind Sankar", "Manu Saravanan", etc.

Also, although difficult, Gaussian blurs can be in certain cases reversed (even though this one is so obvious there was no point in blurring the names). I prefer a fat black box when censoring.

If you want a more robust parser I made scripts to export Messenger and Hangouts chat logs to dataframes for a previous project: https://github.com/MasterScrat/ChatShape

It will give you rows with : [timestamp, interlocutorName, messageSenderName, text]

Thanks! This is definitely useful

This reminds me of this year's xyzzy award-winning interactive fiction game, The Mary Jane of Tomorrow. https://emshort.blog/2016/06/05/the-mary-jane-of-tomorrow/ (It won Best Single Puzzle and Best NPC.)

No mention of Black Mirror?

Also: https://www.theverge.com/a/luka-artificial-intelligence-memo...

Just watched the episode today LOL Insane stuff

Great explanation! I also worked on a similar project (https://medium.com/towards-data-science/personality-for-your...).

Also feel free to check my Conversation Analyzer (https://github.com/5agado/conversation-analyzer), which includes scraper and parser for Facebook Messanger conversations.

Nice writeup! Definitely seems to be a lot more realistic than my bot LOL The parser is really helpful too

Is it possible to make it spew out random personal information like addresses and phone numbers if you give it the right input? That's usually what happens when you train a RNN and it overfits.

Yeah you're right it is a possibility. Overfitting was definitely a problem in this project. I think just a larger dataset and maybe adding some regularizers would have helped.

Luckily I dont think I have too much compromising information in my dataset LOL

Doesn't seem like it's really any better than eliza or the emacs doctor or whatever stuff didn't use the buzzword tech of the year.

Ha! Just yesterday I had the narcisstic idea of gifting someone I've been chatting regularly with for years a chatbot that answers his messages like I would. This is just what I needed, thanks!

Awesome! Let me know if you get good results or find a way to improve my Seq2Seq model

I have this idea pretty regularly. Human interactions are quite often annoying :).

Really cool project! Curious to see what type of changes you could make to get more realistic results

while(0) {cout << 'hurr. durr.\n'; sleep(1000);}

So exit immediately?

Hi. Welcome to chatbot/n.

Reply on news.ycombinator.com