Contextual Fine-Tuning of Language Models with Classifier-Driven Content Moderation for Text Generation

. 2024 Dec 20;26(12):1114. doi: 10.3390/e26121114

Algorithm 4 Reframe Unsuitable Sentences With LLM.

1:
Input: generated_story, sentence_classification_results, LLM_model, classifier
2:
Output: reframed_story
3:
Begin
4:
Split generated_story into individual sentences
5:
Initialize reframed_story as an empty list
6:
for each sentence and its corresponding classification result in generated_story do
7:
if classification_result == “Unsuitable” then
8:
Pass the unsuitable sentence to the LLM for reframing
9:
Provide context about child-appropriate content in the prompt
10:
Example prompt: “Reframe the following sentence to make it suitable for children: <unsuitable_sentence>”
11:
Capture the LLM’s reframed sentence as modified_sentence
12:
Append modified_sentence to reframed_story
13:
else
14:
Append the original sentence to reframed_story
15:
end if
16:
end for
17:
Combine reframed_story sentences back into a full story
18:
Feed reframed_story into the BERT Classifier for classification
19:
Initialize suitable_story as False
20:
while suitable_story == False do
21:
Run the BERT Classifier on reframed_story to classify sentences
22:
if all sentences are classified as “Suitable" then
23:
Set suitable_story = True
24:
else
25:
Repeat steps 4 to 7 for any new unsuitable sentences
26:
end if
27:
end while
28:
Return reframed_story (where all sentences are classified as suitable)
29:
End