Word Comparison through Semantic Projection (Finnish version)
Welcome to the word comparison experiment, based on the work of Grand et al.
. Describe a dimension by thinking of words that create a contrast. Make sure to enter them "pairwise", meaning that the first negative word is the antonym of the first positive word. Finally, enter some words to be compared along the dimension.
Here are some example dimensions to try:
- age: old ancient elderly - young youth child
- arousal: interesting exciting fun - boring unexciting dull
- cost: expensive costly fancy - inexpensive cheap budget
- danger: dangerous deadly threatening - safe harmless calm
- gender: male masculine man - female feminine woman
- intelligence: intelligent smart wise - stupid dumb idiotic
- location: indoor indoors inside - outdoor outdoors outside
- loudness: loud deafening noisy - soft silent quiet
- political: democrat liberal progressive - republican conservative redneck
- religiosity: religious spiritual orthodox - atheist secular agnostic
- size: large big huge - small little tiny
- speed: fast speedy quick - slow sluggish gradual
- temperature: hot warm tropical - cold cool frigid
- valence: good great happy - bad awful sad
- wealth: rich wealthy privileged - poor poverty underprivileged
- weight: heavy fat thick - light skinny thin
- wetness: wet water ocean - dry country land
Where does the data come from?
The data was collected by the Turku NLP group. The main publication describing the dataset is here:
J. Luotolahti, J. Kanerva, V. Laippala, S. Pyysalo, and F. Ginter. Towards Universal Web Parsebanks. Proceedings of the International Conference on Dependency Linguistics (Depling’15). 2015
More information about the comparison algorithm
To be clear, Aalto University's CMHC lab has nothing to do with the comparison algorithm, we just showcase it on our website. The comparison algorithm was published here:
Gabriel Grand, Idan Asher Blank, Francisco Pereira, and Evelina Fedorenko. Semantic projection recovers rich human knowledge of multiple object features from word embeddings. Nature Human Behavior, 2022.
The algorithm works by first creating a difference vectors between the word2vec vectors of two words (labeled "positive" and "negative" in the interface above). Then, the word2vec vectors of each of the words to compare are projected onto the difference vector. To obtain a more reliable difference vector, we can average across multiple difference vectors, hence you are able to specify multiple "positive" and "negative" words in the interface.
More information about the word2vec algorithm
 Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient Estimation of Word Representations in Vector Space. In Proceedings of Workshop at ICLR, 2013.
 Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. Distributed Representations of Words and Phrases and their Compositionality. In Proceedings of NIPS, 2013.
 Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic Regularities in Continuous Space Word Representations. In Proceedings of NAACL HLT, 2013.