
 - A research which focus to differentiate the ability of humans and machines to detect the text generated by TGM.The important outcomes of the obervations are

 - Human detetors are capable of indentiying contradictions and semantic errors due to the posession of extensive sematic knowledge.
 - Machine detectors are capable of finding out high frequency words.


Differences in human and machine detector

GLTR tool can identify unexpected and out of context words which can help untrained humans to detect machine generated texts.

Supporting untrained humans using detectors

RoFT tool is used to quantify the human identification of texts generated by TGM. Humans need to identify the edge of text transition from human generated to machine generated.

Real or Fake Text (RoFT) tool 

A system that leverages the use of human visual skills and common sense.

- Differences in human and machine detector
 - Supporting untrained humans
 - Real or Fake Text (RoFT) tool     


San Jose State University

Machine learning models that can generate text  have attained human level of perfection in the areas of grammar, world knowledge, coherence and fluency.

Automatic Detection of Machine Generated Text

Automatic Detection of Machine Generated Text: A Critical Survey
by Ganesh Jawahar, Muhammad Abdul-Mageed, Laks V.S. Lakshmanan
https://arxiv.org/pdf/2011.01314.pdf

Automatic Detection of Machine Generated Text: A Critical Survey


 - Story generation 
 - Conversational response generation
 - Code auto-completion
 - Radiology report generation


Uses of Text Generative Models(TGMs)

 - Fake news generation 
 - Fake product reviews generation 
 - Spamming and phishing


Misuse of Text Generative Models(TGMs)

 - accurate
 - data-efficient
 - generalizable
 - interpretable
 - robust

Qualities of an ideal TGM detector

TGM is a language model, specifically a neural language model estimating the probability distribution of a token or word. 

TGM estimates parameters by minimising the following objective function

$ \mathcal L(p_{\theta},\mathcal D)  = - \sum_{j=1}^{|\mathcal{D}|} \sum_{t=1}^{|\textbf {x}^{(j)}|}    log p_{\theta}( x_t^{(j)} \mid x_1^{(j)},...,x_i^{(j)},....,x_{(t-1)}^{(j)})  $ 

$\textbf{x}_i \in \mathcal{V}$ - Vocabulary of words

$\textbf{x} = (x_1,....,x_{|\textbf{x} |})$ - text sequence

$p_*(\textbf{x})$ - reference distribution

$\mathcal D$ - a finite set of text sequences from $p_*$

$p_{\theta}( x_t \mid x_1,...,x_i,....,x_{(t-1)})$ - Probability of next token given the previous tokens in a given sentence


Training a text generative model

 - Bias
 - Beneficial usage
 - Malicious usage


Social impacts of TGMs

Most of the latest TGM’s used Transformer model architecture ,it helps to learn long range dependency information and the generated texts are correct ,coherent and uses world knowledge.

Text Generative Model Architecture

It is possible to control the topic and sentiment of text generated ,two methods by which we can control the text generated are

 - Training with control tokens
 - Control using attribute classifier


Controllable Text Generative Models

Detectors are used to identify machine generated text from human generated text. Classification of detectors are based on

- Classifiers trained from scratch 
 - Zero-shot classifier 
 - Fine-tuning NLM 
 - Human-machine collaboration


Learn Before

Related

Learn After