Interactive Text Generation

rezaBarzgar commented 1 year ago

What is the problem?

Although non-interactive models are capable of producing texts of high quality, they may occasionally be incapable of generating the specific text that the user desires. This paper introduces a model incorporating user interaction to generate text that closely aligns with the user's desired output.

How do they solve the problem?

Due to the high cost, time consumption, and potential issues, such as biasing the model towards users' beliefs, using real individuals during the training and testing phase is not feasible. Therefore, the emulation of real users is employed in place of actual human participants.

The system comprises two key components: a document-generating model and an oracle (emulation of the real user) with access to the target document. In each step, the model generates a document while considering the oracle's edits. Subsequently, oracle makes further edits based on the previously generated document and the target document. Edits are categorized into three different groups.

insertion
Deletion
Substitution

To comprehend the functioning of this model, we can imagine a scenario in which a student is composing a text while receiving continuous feedback from their teacher.

Input:

a sequence of tokens (the first step)
edits on the generated sequence of tokens (from the second step until a threshold of similarity between generated tokens and target)

Output: a sequence of tokens highly similar to the target tokens.

Dataset: 64 token lengths sentences that are tokenized by BART tokenizer from CNN/DailyMail article summaries

Baselines: The authors assert that their study represents the first instance of an interactive text-generating model, and as such, no existing baseline can be directly compared to their approach. However, they did report BART scores for various iterations of their model.

It is highly recommended to take a look at Figure 1 and Figure 2 of the paper to understand how their system works.

Limitations

As they mentioned in the paper, their system has two major limitations.

using the emulation of real users instead of real users during the training and testing phase.
the input of the system is just sentences of 64 tokens. But they mentioned that their next contribution would be increasing the size of the input to paragraph.

How is this paper related to our work?

(to be completed)

Code: https://github.com/ffaltings/InteractiveTextGeneration Paper: https://arxiv.org/abs/2303.00908

rezaBarzgar commented 1 year ago

@hosseinfani Hi, please take a look at this summary. I got a little bit confused about how we could use their work after reading the paper. I think we can use their approach in our training phase (They model their task to Markov decision Process). But how can we implement their model in Osprey?

hosseinfani commented 1 year ago

@rezaBarzgar Have a look at t5 for classification task (target text is "True" or "False". So, they train text-2-text for classification tasks. I was thinking of using this model to generate the labels as the message comes in.

We can talk tomorrow.

fani-lab / Osprey