Hi Minh,
On 26 Sep, 2011,at 09:20 AM, Nghiêm Quốc Minh <[email protected]> wrote:
The scrip runs smoothly But I have some questions related to the tree-to-tree translation.
1. What is the sentence length and tree height limitation of moses-chart?
I could train long sentences but when I tried to translate long sentences, there is an error in CalcScoreChart that I think it's related to the sentence length / tree height.
The model restricts sentence length and tree height via the word penalty and rule penalty features, and they will ultimately be limited by the grammar, but moses_chart doesn't have any hard limits.
Could you provide some more information about the error you're seeing? What makes you think that it is related?
2. Is the output tree follow the grammar of the language?
I saw that the output string and the output tree is not follow the grammar of the language, it needs some reordering and symbol changing.
Is it because of
+ the decoder uses itself rules or
+ the size of training data or
+ the non-terminal (next question) or
+ the language model (question 4)
3. How to define the non-terminal?
My source and target language use different "non-terminal" symbols. Without defining the non-terminal, the system uses the "Q" and "X" symbols for non-terminal. Because of that, the target tree is not a correct tree. I saw that we can define non-terminal symbols in a file but I do not know how to do this.
I'm not sure I understand the question here. Just to clarify, "Q" and "X" are used for the glue grammar. The main grammar non-terminals should come from the labels on the parsed parallel corpus. It's also possible to specify a file containing labels to use for unknown words. I think this is described in the Moses manual (search for 'unknown-lhs').
4. How about the language model of tree-to-tree translation?
I used the string (not the tree) of target language to train LM. Is there anyway to train LM with tree information?
There's been work on syntactic LMs for phrase-based models. In tree-based models, the target tree information should help produce better-formed output, though there are no guarantees...
Phil
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
