Paul has proposed working together to publish a paper on data science. I think this is a really good idea, so I should try to come up with some ideas that we could work together on.
It’s desirable for the problem to be an open problem that we are both interested in. Given that Aykut has suggested that Computer Vision has had a lot of research effort invested in it, perhaps something that focuses on a different area than CV would be good.
We all work with text a lot so a natural language problem would make the most sense. I’ve been interested in summarization and semantics for a while now. Maybe something in that area would be good.
Logical Inference
To determine the quality of a summarizer it must be possible to make logical statements about a text. All logical statements that a text infers could be generated and then the logical statements for a summary could be generated. If the summary statements are a subset of the full text statements then the summary is faithful to the original text.
The kind of logical inference I am thinking of are one step relationship statements. Given the text:
The cat sat on the mat
You can make the statement “the mat is below the cat” which is true, and “the cat is brown” which is false (closed world assumption).
Logical inference could be achieved by taking statements from the text and then asking a language model a question to determine if the relationship holds. The idea would be that the language model can reason it’s way through the different ways that the text structures the information, and that such an approach would be better than a purely rule based system.
There already exists rules based systems for this, I’ve already found one in Standford NLP. Here is a paper on Natural Logic for Textual Inference. I also recall a deep learning model being shared in the data-science channel that did closed world logical inference, unfortunately I cannot find that right now.
Music Genre
Another idea might be the automatic determination of the genre of a provided musical piece. This could be an interesting way to recommend musical pieces to someone based on the qualities of the music which are the same instead of the collaborative filtering approach that is commonly used. What is nice is it wouldn’t suffer from the cold start problem for new pieces of music, because the qualities of a piece of music could be determined without anyone hearing it.
Music Generation
A lot of work has been done on music generation and the consensus seems to be that it is bad. Perhaps the problem here is that the music generation is done at one “level” when music tends to have sequences and themes that occur both over the entire piece and during a small window. If an ensemble of models was used, where some of them operate over long time scales and then provide part of the input to others, then music which is more varied over time could be achieved.
Mapping Text to Emoji
To determine the emotional content of a snippet of text is hard. Perhaps we can provide a mapping from some text to the emoji which most represents that text.
We might then want to label the clusters of emoji. The label for the emoji could be the emotive text that best fits in that cluster?