The ISTI group tried to white list all the tokens that appeared at least once on a random sample of 50 documents of the dataset when generating the synthetic news. This was done to try and match the probability distribution of the words in the real article with the one of the fake ones.
While not entirely successful it succeded in worsening their Roberta classifier (if the first 100 tokens or so of the text were skipped).
It was proposed to plot the number of tokens whitelisted in relation to a growing cutoff number (50, 100, 1000 documents, etc. ) to see where should te trace the line. Also, instead of choosing random documents, I proposed to use the top documents with the highest number of different tokens in it.
We left by saying what to do:
- Re-evaluate the dataset and, maybe, changing it with one already used to do MGT detection;
- Do a UD profiling on: original cut at 256, synth and synth lexically controlled;
- Prepare a dataset for DPO where the preference is expressed by selecting texts where an important feature for a SVM trained to classify MGT, has a large gap, expressing the preference towards the sentence that the SVM would classify as the correct one. Prepare a single dataset with a balanced number of pairs for each of the top x most important features.
- At the same time, prepare prompts with verbalized feature, to explicitly ask the model to write generations with certain linguistical feature we know characterize the real texts with respect to the synth ones.
- Generate with both the model that gets prompted, and the model that has done DPO, both with the “lexical informed decoding” (aka the white list) a new set of generations (iteration 1) and evaluate them both with RoBERTa, the SVM and human judgments.
- Repeat until the original and the synth texts are very similar.