implement intersect

[qindazhu]

No description provided.

[csukuangfj]

could you visualize the result and post them here?

[qindazhu]

could you visualize the result and post them here?

Sure!

Input FSA a

Input FSA b

Output FSA c = a^b.

[csukuangfj]

Should the output be trimmed (connected) ?

[qindazhu]

Should the output be trimmed (connected) ?

I think it's better to leave this to the caller to do that.

[csukuangfj]

OpenFST produces a connected fsa.

Please refer to http://www.openfst.org/twiki/bin/view/FST/IntersectDoc

[qindazhu]

OpenFST produces a connected fsa.

Please refer to http://www.openfst.org/twiki/bin/view/FST/IntersectDoc

OpenFst supports kinds of filters, so callers can decide what fsa they want. However, we don't support such filters, so maybe callers want do RmEplision then Compose, etc.

Thus, I think it's better to let the caller do this instead of doing such operations (we make the decision for them) in one function call.

[csukuangfj]

ok.

Another point: why is there a self-loop with label epsilon in the input fsa a and the output fsa?

[qindazhu]

ok.

Another point: why is there a self-loop with label epsilon in the input fsa a and the output fsa?

I guess what you are asking is why there is self-loop with label epsilon in the input fsa (as if there is one in the input, there may be such a state in the output).

Well, that's just because we support such input fsa, we only require that either a or b is epsilon-free, but not limit what labels can be on self-loop. Anyway, the caller can remove such arcs if they want.

[csukuangfj]

I think epsilon free means you can go to state b from a with the epslion label (i.e., no input).
But a should be different from b.

Does every node implicitly have a self-loop with label epsilon?

[qindazhu]

We treat the epsilon free fsa has a self loop with epsilon at each state. Say if a has epsilon arc and b is epsilon-free. When we match Arc(src_a, dest_a, epsilon) in a with Arc(src_b, dest_b, label), we always add a arc (src_a, src_b) -> (dest_a, src_b) with label epsilon (That means we add a self-loop in src_b with epsilon)

[csukuangfj]

Arc(src_a, dest_a, epsilon)

I think dest_a != src_a.

From wiki: https://en.wikipedia.org/wiki/Epsilon_transition

An epsilon transition[1] (also epsilon move[1] or lambda transition) allows an automaton to change its state spontaneously, i.e. without consuming an input symbol.

An epsilon transition -->changes<--- the state.

The input fsa a has a transition

3 3 0

which confuses me.

[qindazhu]

As I said before, the caller can remove there arcs (even for input fsa) if they want or they should. That's not the responsiblity of Intersect.

BTW, my point for implementing those algorithms is : we should take requirements as few as possible on the input.

[qindazhu]

If there no other issues, please help to merge.

[danpovey]

Regarding self-loop epsilon transitions: they may seem pointless, but in the transducer case they can do something nontrivial so they should be kept. (This code can be used for transducers, you can use those lists of source arcs to do that).

…

On Sat, May 2, 2020 at 1:47 PM Haowen Qiu ***@***.***> wrote: As I said before, the caller can remove there arcs (even for input fsa) if they want or they should. That's not the responsiblity of Intersect. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <https://github.com/danpovey/k2/pull/22#issuecomment-622673596>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLO2RN4NEKM4WZOHEDTDRPOXVRANCNFSM4MXQSRLQ> .