ngramshrink man page

ngramshrink — shrink n-gram model using pruning criteria

Synopsis

ngramshrink [Options] [in.fst [out.fst]]

Description

Shrinks an input n-gram model using given pruning criteria.

Options

Boolean options accept either true or false as a value, or the default value if neither value is specified.

--backoff_label=INT64

Backoff label.  Default: 0.

--check_consistency[=BOOLEAN]

Check model consistency.  Default: false.

--context_pattern=STRING

Pattern of contexts to prune.  Default: no pruning.

--count_pattern=STRING

Pattern of counts to prune.  Default: no pruning.

--end_symbol=STRING

Class label for sentence end.  Default: "</s>".

--fst_align[=BOOLEAN]

Write FST data aligned where appropriate.  Default: false.

--fst_compat_symbols[=BOOLEAN]

Require symbol tables to match when appropriate.  Default: true.

--fst_default_cache_gc[=BOOLEAN]

Enable garbage collection of cache.  Default: true.

--fst_default_cache_gc_limit=INT64

Set the cache byte size that triggers garbage collection.  Default: 1048576.

--fst_error_fatal[=BOOLEAN]

If true, FST errors are fatal.  Otherwise, returned objects are flagged as bad. For example, FSTs are returned with the kError property set to true, and FST weights set so that Member() returns false.  Default: true.

--fst_field_separator=STRING

Set the characters used as a separator between printed fields.  Default: " ".

--fst_read_mode=STRING

Default file reading mode for mappable files, "read" to use read(2) or "map" to use mmap(2).  Default: "read".

--fst_verify_properties[=BOOLEAN]

Verify fst properties queried by TestProperties.  Default: false.

--fst_weight_parentheses=STRING

Set the characters enclosing the first weight of a printed composite weight (e.g., pair weight, tuple weight, and derived classes) to ensure proper I/O of nested composite weights.  Must have size 0 (none) or 2 (open and close parenthesis).  Default: no parentheses.

--fst_weight_separator=CHARACTER

Set the character separator between printed composite weights.  Default: comma.

--help[=BOOLEAN]

Show usage information.  Default: false.

--helpshort[=BOOLEAN]

Show brief usage information.  Default: false.

--method=STRING

One of "context_prune", "count_prune", "relative_entropy", or "seymore". Default: "seymore".

--ngram_error_fatal[=BOOLEAN]

If true, NGram errors are fatal.  Otherwise, returned objects are flagged as bad; i.e., NGramModel::Error() returns true.  Default: true.

--norm_eps=FLOAT

Normalization check epsilon.  Default: 0.001.

--save_relabel_ipairs=FILENAME

Save input relabel pairs to file.  Default: do not save.

--save_relabel_opairs=FILENAME

Save output relabel pairs to file.  Default: do not save.

--shrink_opt=INT32

Optimization level, from 0 (fastest) to 2 (most accurate).  Default: 0.

--start_symbol=STRING

Class label for sentence start.  Default: "<s>".

--target_number_of_ngrams=INT64

Maximum number of ngrams to leave in the model after pruning.  Values less than zero mean no target number, just use theta.  Default: -1.

--theta=FLOAT

Pruning threshold theta.  Default: 0.0.

--tmpdir=PATHNAME

Set the temporary directory to use.  Default: /tmp.

--total_unigram_count=FLOAT

Total unigram count.  Default: -1.0.

--v[=INT32]

Set the verbosity level.  Default: 0.

See Also

ngramapply(1), ngramcontext(1), ngramcount(1), ngramhisttest(1), ngraminfo(1), ngrammake(1), ngrammarginalize(1), ngrammerge(1), ngramperplexity(1), ngramprint(1), ngramrandgen(1), ngramrandtest(1), ngramread(1), ngramsort(1), ngramsplit(1), ngramsymbols(1), ngramtransfer(1)

Referenced By

ngramapply(1), ngramcontext(1), ngramcount(1), ngramhisttest(1), ngraminfo(1), ngrammake(1), ngrammarginalize(1), ngrammerge(1), ngramperplexity(1), ngramprint(1), ngramrandgen(1), ngramrandtest(1), ngramread(1), ngramsort(1), ngramsplit(1), ngramsymbols(1), ngramtransfer(1).

1.3.2 OpenGrm NGram User Commands