>>34748696I'm sorry I cannot link you the paper because it would doxx me. But in short, it uses a supervised
combination of different unsupervised textual similarities. These textual similarities
rely on the introduction of a relation matrix in the classical cosine similarity between bag-of-words, so as to get a softcosine that takes into account relations between words. According to the type of relation matrix embedded in the soft-cosine, semantic or lexical relations can be considered.
(tldr : To evaluate the similarity between A and B, you're looking for the level of relationship between each pair of words)
While this is a surprisingly simple approach, especially compared to large language models that dominate everything today, it works very well! In some specific cases better than BERT models, because it's cheap to train on your own data.
Now I must say I don't really know why these three would be grouped together. I remade the algorithm from scratch at home as an exercise, it's possible I messed up somewhere desu. My guess would be during the tokenization or lemmatization phase.
