I am using word2vec/doc2vec to find text similarities of two documents. I studied that word2vec works on two approaches :
But I am stuck at understanding how these two approaches works in calculating the text similarities. Also which one is the better approach for the current task.
Word vectors just model individual words.
But, you can then use these per-word vectors to create vectors for larger texts, or similarity-calculations between larger texts.
A simple way to turn a text into a single fixed-width vector is to average the word-vectors of all the text's words. (This could also be a weighted average, based on some ideas of individual words' importance.) This sort of text-vector can often work well as a quick and simple baseline. For two texts, the cosine similarity of the two averages-of-all-their-word-vectors is then the similarity of two texts.
An algorithm like Doc2Vec (aka "Paragraph Vector") is an alternative way to get a vector for a text. It doesn't strictly combine word-vectors, but rather uses a process like what is used to create word-vectors to create per-text vectors instead.
If just working with word-vectors, another option for text-to-text similarity is "Word Mover's Distance" (WMD). Rather than averaging all word-vectors together, to create a single vector for the text, the WMD measure treats all the words of a text as "piles of meaning" at their various word-vectors' coordinates. The distance between texts is how much effort is required to "move" the mass of one text's word-vectors to the other's. It's expensive (since each such pairwise calc is an optimization problem among many possible word-to-word shifts) but retains a bit more distinction than just collapsing a text into a single summary vector.