Question : Why Recurrent Neural Network, when we have the word 2 vector?
Answer : Word 2 vector is able to capture the inter relationship, however where it fails is to capture the intra-relationship. RNN’s are able to capture this intra-relationship. e.g
Word 2 vector : Given a document “Capital of US is Washington”. “It’s capital since 1970”
Word2Vec can capture relationship that “Washington” is related to “Capital” and “US”.
However, Word2vec cannot find intra-relationship between the above two sentence i.e it is not able to predict the second sentence given the first sentence. This is where the RNN’s true potential lies.
Question : What are some of the cases of intra-relationship usefulness?
Answer : Some of the cases, where intra-relationship is extremely useful is
1. Language Translation: e.g Find intra-relation ship between language.
2. Speech Recognition: Find intra-relation between the voice and the words.
Question : Well, I seen in the screen shot that for two three lettered sentence, the output is of varying length i.e 2 and 3 words respectively? Is this varying length not a problem?
Answer: Yes, it is a problem since it implies that the n word sentence when translated can be of any arbitrary length “n-m” where m varies both in positive and negative direction.
This is extremely problematic for the word 2 vectors, since it requires the fixed length, however RNN’s are very effective at handling such varying length.