Billions of parameters to be learnt.
Training data from raw text from a diverse set of sources (like Wikipedia, Reddit, and news sources


Training Data for TGM

Billions of parameters to be learned from millions of data which would need a humongous computational cost.

Training cost 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.

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

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


TGM Classifiers or Detectors

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     


Learn Before

Related

Learn After