Linguistic Analogies in Word Embeddings: Where Are They?

Word Embedding has greatly improved Natural-Language Processing. In word-embedding models, words are represented as vectors in a multi-dimensional space; these vectors are trained through neural networks, by means of very large corpora of textual documents. Linguistic analogies are claimed to be encoded within word-embedding models, in such a way that they can be dealt with through simple vector-offset operations. This paper aims to give an answer to the following research question: given a word-embedding model, are linguistic analogies really present? It seems rather unrealistic that complex semantic relationships are encoded within a word-embedding model, which is trained to encode positional relationships between words. The investigation methodology is presented, and the results are discussed. This leads to the following question: “Linguistic analogies in Word Embeddings: where are they?”.