Two main Dialogue Systems architectures:

• frame based systems: talk to users + accomplish specific tasks
• LLM: reasoning as agents

Dialogue Systems vs Chatbot

Previously, when we say Chatbot we mean task-based systems

humans and chat

humans tend to think of Dialogue Systems as human-like even if they know its not. this makes users more prone to share private information and worry less about its disclosure.

ELIZA

see ELIZA

LLM Chatbots

Training Corpus

C4: colossal clean crawled corpus

\(\chi^2\) is a test statistic for hypothesis testing.

motivation for chi-square

The motivation for chi-square is because t-test (means, “is the value significantly different”) and z-test (proportion, “is the incidence percentage significantly different”) all don’t really cover categorical data samples: “the categories are distributed in this way.”

Take, for instance, if we want to test the following null hypothesis:

\(\alpha = 0.05\). What do we use to test this??

Chiara Marletto is an physicist working on Quantum mechanics working in D. of Physics, Wolfson College, University of Oxford.

Subfield: constructor theory. She studies quantum theory.

I was digging through my OneDrive recently for work, and found this piece of writing.

There is naught but a small, dirt-filled puddle in front of this lawn. Yet only here – by the puddle – can Gary find a small, much-needed respite from the neverending work. Of course, without the hours he has committed to the sweatshop, his mother would have died ages ago from colora.

But how does it matter now? Rarely now – once every year – does he even earn the privilege to exit the heavily-guarded area to visit his mother; and how little time he has during such visits: each visit seems to just be a long walk, a knock, a kiss on the cheek – then back to the workhouse he goes.

DOI: 10.3389/fpsyg.2020.623237

One-Liner (thrice)

1. Used features extracted by VGGish from raw acoustic audio against a SVM, Perceptron, 1NN; got \(59.1\%\) classif. accuracy for dementia
2. Then, trained a CNN on raw wave-forms and got \(63.6\%\) accuracy
3. Then, they fine-tuned a VGGish on the raw wave-forms and didn’t report their results and just said “we discovered that audio transfer learning with a pretrained VGGish feature extractor performs better” Gah!

Novelty

Threw the kitchen sink to process only raw acoustic input, most of it missed; wanted 0 human involvement. It seems like last method is promising.