Raisin Poem feat. Langston Hughes

It Depends


To be frank, this assignment was kind of frustrating. I was very excited to utilize word2vec and discover interesting relationships between words that maybe I wasn’t expecting. I also thought relationships between words I thought would be clear would have a relationship even when manipulating vectors. However, that is not what happened. I was met with strange answers that I couldn’t creatively make into a poem. I had trouble finding linear relationship, especially when trying to do incremental operations. For example, I thought that adding a positive word like happy to a negative word like tragedy iteratively in my second poem would gradually converge to a positive word. However, the results either did not point to a positive word or the vector maxed out and kept producing the same word. I did end up discovering that only changing some of dimension resulted in more variable answers. This was useful for me in my first poem where I converted an existing poem into its negative version by multiplying 25 of the 50 dimensions by -1. Sonifying the text however was a lot more fun. Playing with how the tone changed when word2vec provided a similar word challenged me to shape the sound. I found that rhythmic sounds that matched when the words appeared made sense, but also have long tones when writing a phrase of words was soothing. I tried to experiment a little bit with how tempo affected the moods of the poems as well. Overall, the concept was cool and it’s interesting to see what word2vec produced. However, it’s not really clear how it’s comparing numbers, which makes it hard to work in a consistent way.