Fix finalize error in ContextGraph implementation.

[maxwellzh]

For Aho-Corasicks algorithm, at the end of beam search, we should subtract scores of edges state.node_score only if current state is not an end of word.
Fix is simple.
cc @pkufool

[pkufool]

@maxwellzh Thanks!, but the fix is not correct. I have some major improvements in this PR #1428 .

[maxwellzh]

@pkufool plz correct me if I'm wrong,

1. node.node_score is the cumulative bonus of all edges from root to current node. This score is used to encourage beam search going along word token path
2. And there is an output_score as bonus for word completion. This score encourages beam search to get completed word.

So at finalize(), why you would like to subtract all bonus of edges for a completed word?

e.g.
For a graph with single edge (1, 2, 3), weight 1.0

At finalize(), path [1,2,3] is expected to have 1.0+1.0 bonus, which without the fix would be 1.0 instead. Or is this by design?

[pkufool]

@maxwellzh Sorry, I am not quite follow your example. The idea is in current implementation all bonus are contained in the node.output_score already, you can have a look at the test cases in context_graph.py, it will draw the graph into pdf format, you can go through all cases step by step.

[maxwellzh]

@pkufool Sry I misunderstood the implementation. Thanks for the clarification :)