Sentence merge algorithm

[ABarpanda]

As of now, the sentence merging works on a threshold based system. But it has some problems

• It is very subjective to the input text data (story)
• If the story is large and threshold is also high, then none of the sentences will be merged. This may cause the program to create huge number of images.

Idea -

• Create a system that will collect the similarity values
• Then it will select which sentences to merge
  • It can select a certain percentage (20-30%) and the top percentage of similarity will be merged
  • Create a distinction based system
    • Suppose the values are (0.8, 0.76, 0.78, 0.2, 0.1, 0.15); then the algorithm must able to figure out to select the first 3

[Dhruta25]

like you are saying you have limited token so when someone giving promp it should be in limited manner? we can do semantic search it understand the context also do cosine seach and get result?

[ABarpanda]

Yes, something like that. Also if we generate 2 images - "The man stopped with a dramatic pause. The man spoke up with excitement". These sentence are related very highly. So it doesn't make sense to generate 2 images for almost the same sentence.

[Dhruta25]

can u provide me the location of you code i can solve this issue .i use semantic similarity filtering like i use embedding model and with cosine similarity generate a score after that if similarity>.9 then merge then into one this will generate one image

[ABarpanda]

https://github.com/ABarpanda/Animegen/blob/main/src/scene_split.py