Hello,
Could you explain how we obtain the solutions for hands on 6 - exercise 6?
Thank you
Kasimir
Question VI.I:
We assume the limited scope for syntactic dependencies (1 neighbor) hypothesis here, which means that the tagging of "fast" depends on the tagging of "fly". In the same way, the tagging of "fly" depends on the tagging of "ducks", and the tagging of "ducks" depends on the tagging of "young". There is a chain of dependencies here. So, we can say that the tagging of "fast" depends on the tagging of "young". This can be done in both directions.
Question VI.II:
When tagging of "birds" is known (i.e. when the word "bird" is non-ambiguous), it breaks the above chain of dependencies. Hence, the tagging of "fast" no longer depends on the tagging of "young" and vice versa.
We assume the limited scope for syntactic dependencies (1 neighbor) hypothesis here, which means that the tagging of "fast" depends on the tagging of "fly". In the same way, the tagging of "fly" depends on the tagging of "ducks", and the tagging of "ducks" depends on the tagging of "young". There is a chain of dependencies here. So, we can say that the tagging of "fast" depends on the tagging of "young". This can be done in both directions.
Question VI.II:
When tagging of "birds" is known (i.e. when the word "bird" is non-ambiguous), it breaks the above chain of dependencies. Hence, the tagging of "fast" no longer depends on the tagging of "young" and vice versa.
Thanks Anu!
As a follow-up: maybe drawing the lattice (the data structure involved in the Viterbi algorithm) could help to see this: (for an order-1 HMM), if all the paths go through one single node (non-ambiguous word), then right and left part of the optimal path are independent.
Hope it makes sense.
As a follow-up: maybe drawing the lattice (the data structure involved in the Viterbi algorithm) could help to see this: (for an order-1 HMM), if all the paths go through one single node (non-ambiguous word), then right and left part of the optimal path are independent.
Hope it makes sense.
Hi Professor, I have a question about order 1-HMM.
If an HMM is of order 1, P(T_i/T_1, ..., T_i-1) = P(T_i/T_i-1), it means that the tag of a specific node only depends on the node on the left and the node on the right. If every node satisfies this situation, each node must be a non-ambiguous word, then why can we say the sentence in Question VI.I, where all the words are potentially ambiguous, can meet the “limited scope for syntactic dependencies (1 neighbor)” hypotheses?
I'm not sure where I went wrong.
Thanks in advance.
If an HMM is of order 1, P(T_i/T_1, ..., T_i-1) = P(T_i/T_i-1), it means that the tag of a specific node only depends on the node on the left and the node on the right. If every node satisfies this situation, each node must be a non-ambiguous word, then why can we say the sentence in Question VI.I, where all the words are potentially ambiguous, can meet the “limited scope for syntactic dependencies (1 neighbor)” hypotheses?
I'm not sure where I went wrong.
Thanks in advance.
Thank you for your answer! What confused me was that it can be done in both directions.
I would have answered that fly does not depend on fast and that young does not depend on fast either.
With Mr. Chappeliers answer I see why we have to consider both directions.
I would have answered that fly does not depend on fast and that young does not depend on fast either.
With Mr. Chappeliers answer I see why we have to consider both directions.
> What confused me was that it can be done in both directions.
Don't forget the starting point/the original objective, which is to maximize P(T_1, ..., T_n | w_1, ..., w_n) : there is no "direction" in this target formula.
It's not because we choose to parametrize it in one direction (typically the reading direction) that the original objective changes. It remains "direction-less".
