Abstractive Summary or Plain Abstract?

Abstractive summary - that is what I have been working on for the past many months. Researching, learning, feeling frustrated over, racking my brain, learning, researching - it was a pendulum. It refers to summaries that are generated the way we speak, and I suppose we speak in 'abstractive' ways. In other words, it is a summary that retrieves the gist - not just repeating what was already said in the original text.

Last week, as I was fine-tuning, rather actually testing to see if any of my work worked, I realized it needed a little work. Maybe, I am understating this.

The abstractive summarizer I was building was supposed to be able to create or generate phrases that made sense in English, not nonsensical gibberish. After it was trained on, I admit, a very small dataset, with even more conservative model construction for fear of cpu-death, it was given a sentence to summarize. The output was so bizarre that bizarre is not a bizarre enough word to describe it. All the words were English, which was good news, but together, they made no sense. It was as if the model had randomly selected words out of a dictionary that was very limited and that was how it generated the summary - not very smart, nor showing the capacity to learn.

Clearly, what my model will receive points for is creativity. When we learn languages, or anything, we become biased, and I think it forces us to think only in a certain way. Could we or would we ever speak the way my model did, using random words that don't make sense in any way possible? Probably not. If we do, we make assumptions and jump to conclusions - either they have not learnt the language or are new, they are trying to be silly, or they are like James Joyce and have gone insane, all judged by their sentence formation. It is often not taken as creativity or artistic skill. Of course, a machine is probably not creative, but it could be, if we teach it to be. However, how can we teach a machine to be creative if we do not learn to appreciate different styles of thinking ourselves without arriving at conclusions automatically?

To pivot from abstraction, I was torn apart. I had no idea what to do. My model was speaking in a different language that used English words, and had sentences that were much longer than the actual phrase it was supposed to summarize. I was forced to look in a dictionary myself to make sure that the meaning of summary meant to shorten the original phrase and return the gist. What seemed hilarious in this exercise was a discovery of myself. My teachers and parents alike always laughed at me or sighed in exasperation when they saw my writing, especially my summaries. On timed essays, I would write so much I would run out of time to finish them. On essays with word limits, I was always above without fail, and struggled with desperation to bring it to the limit - an impossible feat. Most of what I wrote was almost always redundant, unnecessarily complicated, or simply over-detailed. When it came to summaries, my summaries became essays, and then novels - they clearly defeated the purpose of a summary.

Me and my model operate the same way. I picked abstractive summarization, as destiny would have it, I think to help me find appreciation in shorter lengths, as this blog post is not. By teaching a machine what summarization means, I think it would help me better understand what it means as well, so I could create summaries that weren't too abstractive-ly lengthy.