How to represent huge paragraphs of text so as to provide it as an input to a model that performs abstractive summarization? I have read about distributional methods like word2vec but are there any others? How useful are graphical representations as far as neural network based summarization methods are concerned?
If the article is simple English text then one of the simplest approach would be to use NLTK to do the pre-processing (stop-word removal, tokenization, sentence creation,etc). In the later stage, you could rate each word a score based on the maximum frequency. In final stage, you can pick up the sentences ,let say top 10 sentences with maximum score. Hopefully, it will summarize the whole article. However, this is influences by the corpus you take. I think Deep learning approaches can even further improve the results. But, for baseline, this can be implemented. I hope its helpful.
Dear Siddhartha Bhuyan
I would say: come to understand a system of representation (visual representations or visually-related representations in the Memories) AND, when looking at the "system", appreciate and also properly process the hierarchical relationships between major sets of such representations (active in a given set of circumstances, aka environmental [aspects]). [ I may seem to be "talking" mainly about the visual, but humans are VERY visual -- and it is hard to imagine much not related to that. ]
[ NOTE: "Hierarchical" is often not well-understood; it does not just refer to building on earlier "pieces" BUT some change, refinement, integration and consolidation of those "pieces"; __AND__ THEN also (in a true hierarchy) some new significant guiding and often occurring (re-occurring) sorts of aspects of concrete realities (qualitatively similar in any given level/stage change, and qualitatively similar across circumstances) EMERGE with ontogeny. This is as with sophisticated, well-contextualized innate guidance [(with each stage]) of perception then attention; these are good part of what it added for something to be (or forming) a true hierarchy. ]
If you can correctly model all this THAT IS SOMETHING and that seems clearly related to representing "large paragraphs of text for performing automatic abstractive summarization" in-reality -- if I get your meaning. This is what I, as a unique developmental psychologist and stance empiricist, propose AGI (true artificial intelligence) people incorporate as a processing into their robots.
One of the answers is word2vec. The harder question is how to generate the summary from the digest. One option is one of the varieties of Recurrent Neural Network. So the whole system will be a little like an autoencoder, except that the decoded text is shorter than the encoded text.
Before summarizing a long text, it is necessary to structure it, for example to determine subtopics. Known methods to do this are "Inverse Document Frequency", "Latent Dirichlet Allocation" or "Latent Semantic Indexing". The results can then be visualized with Gephi or Neo4J, for example. Texts can then be created based on this outline. We are currently working on this subject and are particularly interested in the writing process (Natural Language Generation). We have already published the analysis of a large number of texts in our IEEE article (open accesss), that can be downloaded here: https://ieeexplore.ieee.org/document/8718286
- Badges
- Science topic