LSM(1) Latent Semantic Mapping LSM(1)
NAME
lsm - Latent Semantic Mapping tool
SYNOPSIS
lsm lsm_command [command_options] map_file [input_files]
DESCRIPTION
The Latent Semantic Mapping framework is a language independent, Unicode based technology that builds maps and uses them to classify texts
into one of a number of categories.
lsm is a tool to create, manipulate, test, and dump Latent Semantic Mapping maps. It is designed to provide access to a large subset of the
functionality of the Latent Semantic Mapping API, mainly for rapid prototyping and diagnostic purposes, but possibly also for simple shell
script based applications of Latent Semantic Mapping.
COMMANDS
lsm provides a variety of commands (lsm_command in the Synopsis), each of which often has a wealth of options (see the Command Options
below). Command names may be abbreviated to unambiguous prefixes.
lsm create map_file input_files
Create a new LSM map from the specified input_files.
lsm update map_file input_files
Add the specified input_files to an existing LSM map.
lsm evaluate map_file input_files
Classify the specified input_files into the categories of the LSM map.
lsm cluster [--k-means=N | --agglomerative=N] [--apply]
Compute clusters for the map, and, if the --apply option is specified, transform the map accordingly. Multiple levels of clustering may
be applied for faster performance on large maps, e.g.
lsm cluster --k-means=100 --each --agglomerative=100 --agglomerative=1000 my.map
first computes 100 clusters using (fast) k-means clustering, computes 100 subclusters for each first stage cluster using agglomerative
clustering, and finally reduces those 10000 clusters to 1000 using agglomerative clustering.
lsm dump map_file [input_files]
Without input_files, dumps all words in the map with their counts. With input_files, dump, for each file, the words that appear in the
map, their counts in the map, and their relative frequencies in the input file.
lsm info map_file
Bypass the Latent Semantic Mapping framework to extract and print information about the file and perform a number of consistency checks
on it. (NOT IMPLEMENTED YET)
COMMAND OPTIONS
This section describes the command_options that are available for the lsm commands. Not all commands support all of these options; each
option is only supported for commands where it makes sense. However, when a command has one of these options you can count on the same
meaning for the option as in other commands.
--append-categories
Directs the update command to put the data into new categories appended after the existing ones, instead of adding the data to the
existing categories.
--categories count
Directs the evaluate command to only list the top count categories.
--category-delimiter delimiter
Specify the delimiter to be used to between categories in the input_files passed to the create and update commands.
group Categories are separated by a `;' argument.
file Each input_file represents a separate category. This is the default if the --category-delimiter option is not given.
line Each line represents a separate category.
string Categories are separated by the specified string.
--clobber
When creating a map, overwrite an existing file at the path, even if it's not an LSM map. By default, create will only overwrite an
existing file if it's believed to be an LSM map, which guards against frequent operator errors such as:
lsm create /usr/include/*.h
--dimensions dim
Direct the create and update commands to use the given number of dimensions for computing the map (Defaults to the number of
categories). This option is useful to manage the size and computational overhead of maps with large number of categories.
--discard-counts
Direct the create and update commands to omit the raw word / token counts when writing the map. This results in a map that is more
compact, but cannot be updated any further.
--hash
Direct the create and update commands to write the map in a format that is not human readable with default file manipulation tools like
cat or hexdump. This is useful in applications such as junk mail filtering, where input data may contain naughty words and where the
contents of the map may tip off spammers what words to avoid.
--help
List an overview of the options available for a command. Available for all commands.
--html
Strip HTML codes from the input_files. Useful for mail and web input. Available for the create, update, evaluate, and dump commands.
--junk-mail
When parsing the input files, apply heuristics to counteract common methods used by spammers to disguise incriminating words such as:
Zer0 1nt3rest l0ans Substituting letters with digits
W E A L T H Adding spaces between letters
m.o.r.t.g.a.g.e Adding punctuation between letters
Available for the create, update, evaluate, and dump commands.
--pairs
If specified with the create command when building the map, store counts for pairs of words as well as the words themselves. This can
increase accuracy for certain classes of problems, but will generate unreasonably large maps unless the vocabulary is fairly limited.
--stop-words stop_word_file
If specified with the create command, stop_word_file is parsed and all words found are excluded from texts evaluated against the map.
This is useful for excluding frequent, semantically meaningless words.
--sweep-cutoff threshold
--sweep-frequency days
Available for the create and update commands. Every specified number of days (by default 7), scan the map and remove from it any
entries that have been in the map for at least 2 previous scans and whose total counts are smaller than threshold. threshold defaults
to 0, so by default the map is not scanned.
--text-delimiter delimiter
Specify the delimiter to be used to between texts in the input_files passed to the create, update, evaluate, and dump commands.
file Each input_file represents a separate text. This is the default if the --text-delimiter option is not given.
line Each line represents a separate text.
string Texts are separated by the specified string.
--triplets
If specified with the create command when building the map, store counts for triplets and pairs of words as well as the words
themselves. This can increase accuracy for certain classes of problems, but will generate unreasonably large maps unless the vocabulary
is fairly limited.
--weight weight
Scale counts of input words for the create and update commands by the specified weight, which may be a positive or negative floating
point number.
--words
Directs the evaluate or cluster commands to apply to words, instead of categories.
--words=count
Directs the evaluate command to list the top count words, instead of categories.
EXAMPLES
"lsm evaluate --html --junk-mail ~/Library/Mail/V2/MailData/LSMMap2 msg*.txt"
Simulate the Mail.app junk mail filter by evaluating the specified files (assumed to each hold the raw text of one mail message)
against the user's junk mail map.
"lsm dump ~/Library/Mail/V2/MailData/LSMMap2"
Dump the words accumulated in the junk mail map and their counts.
"lsm create --category-delimiter=group c_vs_h *.c ';' *.h"
Create an LSM map trained to distinguish C header files from C source files.
"lsm update --weight 2.0 --cat=group c_vs_h ';' ../xy/*.h"
Add some additional header files with an increased weight to the training.
"lsm create --help"
List the options available for the lsm create command.
1.0 2011-11-07 LSM(1)