I have about 500 documents (1 page) that have been mapped to about 3000 short paragraphs (1-2 sentences). These paragraphs describe how the document needs to be reviewed. Each document can and is typically mapped to several paragraphs.
For example, if the document is about the procedures to follow for a certain production process, the paragraphs are about who needs to review the document, what needs to be reviewed, what is the goal of the review, how frequently it should be done etc.
I want to develop a model that can suggest the possible paragraphs from a given document. I have chosen to follow the below approach:
- Prepare the data (tokenize, remove stop words, lemmatize etc.)
- Consider all the paragraphs as one single output, i.e. concatenate them
- Use a sequence to sequence model (tensorflow encoder/decoder RNN model) to map the document to the concatenated paragraphs
- Use the outputted sequence to find the closest paragraphs as a suggestion
Given the small sample size, the model in step 3 does not converge.
I am trying to improve the modelling approach (e.g. one can map each sentence in the document to each paragraph to increase the sample size) or to find alternative approaches. What are some typical models for such a problem?