Journey into word2vec
Tomoya Ishibashi, Kento Niikura, Koki Sone
VR experience to travel into the head of language AI
Journey into word2vec is a VR experience to travel into the head of language AI. This work asks, "What are the language that we use?" and "What is the world of words that machines see?"
"word2vec" is a method of analyzing a large quantity of text data by utilizing a neural network and quantifying the meaning of words. It was announced by Google in 2013, and its high degree of accuracy has had a huge impact on the world. For example, in the sentences "I went to the hospital" and "I went to the police station", it can be understood that hospital and police station are words that are used in a similar way. word2vec utilizes a neural network to analyze a large quantity of text to digitize the meaning of words with high precision. This makes it possible to calculate the meaning of words, such as the words "king + woman" to output the word "queen".
In this work, the high-dimensional information of word2vec is compressed three-dimensionally in order to be perceived by humans and is reconstructed as a VR space. The player wears a head-mounted display and enters a space of words arranged by the order of the machine. The player is free to explore this space while looking for words of interest. As the words are passed by, they are read aloud in a small whisper. When two words of interest are caught, word2vec calculates in real time and warps to a word with their meanings added (for example, we obtained the result that adding "social" to "insect" warps to "honey bees"). By performing this process several times, the player gradually understands the world of words ordered by the machine.
In this work, traveling through an ordered space of words from the perspective of AI will disassimilate and relativize the words we usually use unconsciously, and change the way we view language.