Showing posts with label nlp. Show all posts
Showing posts with label nlp. Show all posts

Sunday, January 29, 2023

Need for Detect ChatGPT to protect ChatGPT

Some Background

ChatGPT has attracted a lot of attention because of the ability to generate text which while being grammatically correct most of the time happens to also be very conveniencing  to an extent that unless you are _actually_ aware of the _reality_ you might believe what the machine learning model, which is estimated to have cost 1.2 billion USD to build with your and my contributed English text data, is factually correct.

This has spiked debate on if students will outsource their school homework to ChatGPT when they are asked to do their homework. Lets keep aside the fact that even before ChatGPT attracted our eyeballs there was homework outsourcing, it was to pay for the service of an actual _human_ do their tasks.

Note that using ChatGPT to do homework is interpreted as cheating. Teachers always think that students are out there to find an easy way to accomplish the task assigned to them;  never questioning their inability to frame tasks that would actually require the students to demonstrate their learning.

Consequently, there has been an engaging debate on identifying text generated by ChatGPT to _help_ teachers identify if the student is indeed the author of the text or if s/he used the services of a AI bot used (underlying assumption is if they were unfair in completing their task)!

Detect ChatGPT is an outcome of this, a portal (I was not aware there was a portal for this until now!) that will allow one to differentiate between human generated text and machine generated text. A plagiarism checker if you wish.

First we research and build ChatGPT and then we realize that this can be used unfairly (by students!) so we come up with a brand new research problem to detect ChatGPT generated text & newspaper headlines "Stanford introduces DetectGPT to help educators fight back against ChatGPT".

 

"Solved Research problems lend themselves to new research problems --Anonymous" 

can not be less true!

 Going Forward

 

 Note that according to openAI, the creators of ChatGPT, ChatGPT has digested almost all the visible English text data that was available on the internet. So at some point of time in the near future the current version of ChatGPT is going to lagging in terms of being informative. 
 
For example say if a company released a new oral vaccine for COVID19 after the ChatGPT was released, then it would not be aware of it because the information corresponding to that would not been digested by the AI model ChatGPT.
 
Sooner or later,  for the above stated reason, at some time in future, the creators of ChatGPT will see a need for their large language model to be upgraded. 
 
See the above picture; "A" is when the ChatGPT was released and say "E" is the point at which, the upgrade is deemed required; the distance between "A" and "E" is the time between two versions of ChatGPT! 
 
The period between A and E, the world is doing what it does best, generate lots of text data. Except that there is an additional source (you guessed right, ChatGPT) generating data. This is shown by the "red" line. It text generation starts with the introduction of ChatGPT (point "A") and it is likely to overtake the "usual modes of" text generation by humans in terms of volume of text generated. 
The hump on the red line indicates the hype - more and more people trying things out using ChatGPT and then putting the generated text out on the internet to shout their findings!
So at some point E, when a ChatGPT upgrade  is envisioned. The upgraded version would digest more of its own generated text (pictorially the area between the two red lines) that the text that has come from usual sources (human; pictorially the area between the two blue lines) of text generation. In some sense it is learning from itself! 
 
This scenario could lead the upgraded ChatGPT into chaos especially if the output of the current ChatGPT is bullshit (as mentioned in  this article). So if more  and more bullshit text generated by ChatGPT is used to train the next version of ChatGPT; the output is going to be more bullshit (what ever that means). Making the upgrade unusable ...

Unless
  1. After the initial hype the text generated by the ChatGPT is reduced (pictorially it is the area between the red dotted lines; denoted by length between points "H" and "I") compared to the usual modes of text generation (by humans; pictorially it is the area between the blue lines; denoted by length between points "G" and "J"). So that the next version of ChatGPT learns from the "new" and "usual" text data than data generated by itself.
  2.  there is a way to identify the text generated by ChatGPT, namely detect ChatGPT! This mechanism can then be used to filter out text generated by ChatGPT, thereby allowing the upgraded version of ChatGPT to only learn from human generated text and not be biased by the text generated by its older version. All this can happen if the detect ChatGPT works to perfection. However, like all explorations, the ability to achieve good accuracy takes time.

While time will tell about #1. There is a definite need put all research might behind exploring methods to detect ChatGPT generated text data. This is probably the only way that ChatGPT will survive.
 
What do you think?

Monday, May 09, 2022

Why is Speech Signal Processing more complex than Text Processing


At best a speech signal can be best described as 

 indiaisthelargestdemocracywelcometoindia

 and in reality it is 

either 

indiaesthelarzestdemocracyvelcometwondia  or indiaes thelarzestdemo cracyvel cometw ondia

Those of you who have paid attention to the 40 character signal would be able to see that it is actually

India is the largest democracy. Welcome to India.

This essentially is the difference between the input's seen in a text processing pipeline versus a speech processing pipeline. For this reason several ask to the speech processing community is (Speech Recognition)

Can you give us "India is the largest democracy. Welcome to India." from the speech signal "indiaes thelarzestdemo cracyvel cometw ondia"? 

so that we can go on with out text processing pipeline and do all the glittery stuff in Natural Language Processing. 

A speech processing researcher is looking at  "indiaes thelarzestdemo cracyvel cometw ondia" or "indiaesthelarzestdemocracyvelcometwondia"  while a text processing researcher is looking at "India is the largest democracy. Welcome to India."

The differences "capitalized alphabets, the punctuation's, the word separators, ..." are a boon, served on a silver plate, to a text processing engineer. While they are to be deciphered, with significant effort, from a speech signal for a speech engineer.

As an example, just imagine how easy it is to identify the number of times "India" occurs in  a text signal versus trying to identify "India" in the speech signal "indiaes thelarzestdemo cracyvel cometw ondia" (Keyword spotting).

But a speech signal come with a lot more information than just the linguistic content. It has information about the speaker gender (male, female), speaker age (adult, child), speaker identity (Sunil, ...), speaker state (happy, neutral, stressed, ....), speaker accent (Indian, Australian, ...), speaker dialect, speaker health (dysarthric, ...) ... which most often does not accompany the text signal "India is the largest democracy. Welcome to India." 

However for any transaction to happen between a machine and a human the crucial step is one of trying to get the linguistic content in the speech. For example, a robot in an airport trying to answer questions posed by the travelers requires the robot to convert speech into text perfectly. And for this reason the priority and focus, in the speech community, has been on exploiting only the linguistic content in the speech signal (Speech Recognition) and not so much on any other information that is abundantly and uniquely present only in the speech signal.

A robot that tries to answer information about Bangalore airport in India. The first part is the ability to understand what the human is "asking", namely to convert speech signal into text.
 

For this reason it is natural that concepts and algorithms that pass the test or performance on text strings (signal) are adopted (given a try!) in speech signal processing.

For those of who who are curious. For a visual feel as to why speech signal is complex, please see complexity of speech.


A Lesson Hidden in Two Simple Words

Sometimes profound ideas are hidden in the most ordinary places. Recently, while looking at two common English words, I stumbled upon an int...