The CMU-Cambridge Statistical Language Modeling Toolkit v2

Contents

If you want to get started making language models as quickly as possible, you should install the toolkit and then read the Typical Use section.


Introduction

Version 1 of the Carnegie Mellon University Statistical Language Modeling toolkit was written by Roni Rosenfeld, and released in 1994. It is available by ftp from here. Here is a excerpt from its README file:


                Overview of the CMU SLM Toolkit, Rev 1.0
                ========================================

  The Carnegie Mellon Statistical Language Modeling (CMU SLM) Toolkit
is a set of unix software tools designed to facilitate language
modeling work in the research community.

  Some of the tools are used to process general textual data into:
    - word frequency lists and vocabularies
    - word bigram and trigram counts
    - vocabulary-specific word bigram and trigram counts
    - bigram- and trigram-related statistics
    - various Backoff bigram and trigram language models

  Others use the resulted language models to compute:
    - perplexity
    - Out-Of-Vocabulary (OOV) rate
    - bigram- and trigram-hit ratios
    - distribution of Backoff cases
    - annotation of test data with language scores

Version 2 of the toolkit seeks to maintain the structure of version 1, to include all (or very nearly all) of the functionality of version 1, and to provide useful improvements in terms of functionality and efficiency. The key differences between this version and version 1 are described in the next section.


Changes from Version 1

Efficient pre-processing tools

The tools used to generate vocabulariesm, and to process the text stream which is used as training data into a id n-gram file to serve as input to idngram2lm have been completely re-written, in order to increase their efficiency.

All of the tools have been written in C, so there is no longer the reliance on shell scripts and UNIX tools such as sort and awk. The tools now run much faster, due to requiring much less disk I/O, although they do now require more RAM than the tools of version 1.

Multiple discounting strategies

Version 1 of the toolkit allowed only Good-Turing discounting to be used in the construction of the models. Version 2 allows any of the following discounting strategies:

Use of n-grams with arbitrary n

The tools in the toolkit are no longer limited to the construction and testing of bigram and trigram language models. As larger corpora, and faster machines with more memory become available, it is becoming more interesting to examine 4-grams, 5-grams, etc. The tools in version 2 of this toolkit enable these models to be constructed and evaluated.

Interactive language model evaluation

The program evallm is used to test the language models produced by the toolkit. Commands to this program are read in from the standard input after the language model has been read, so the user can issue commands interactively, rather than simply from the shell command line. This means that if the user wants to calculate the perplexity of a particular language model with respect to several different texts, the language model only needs to be read once.

Evaluation of ARPA format language models

Version 2 of the toolkit includes the ability to calculate perplexities of ARPA format language models.

Handling of context cues

In version 1, the tags <s>, <p>, and <art> were all hard-wired to represent context cues, and the tag <s> was required to be in the vocabulary. In version 2, one may have any number of context cues (or none at all), and they may be represented by any symbols one chooses. The context cues are a subset of the vocabulary, and are specified in a context cue file.

In order to produce the same behaviour from version 2 as from version 1, the context cues file should contain the following lines:

<s>
<p>
<art>

Compact data storage

The data structures used to store the n-grams are more compact than those of version 1, with the result that language models construction is a less memory intensive task. For example, for a trigram language model, version 1 required 12 bytes per bigram and 4 bytes per trigram. Version 2 requires only 8 bytes per bigram and 4 bytes per trigram.

Support for gzip compression

As well as the compress data compression utility used in version 1 of the toolkit, there is now also support for gzip.

Confidence interval capping

Confidence interval capping has been omitted from version 2 of the toolkit.

Forced back-off

The tool used for evaluating language models allows the user to specify a set of forced back-off parameters. There may be items in the vocabulary (especially context cues and the "unknown" symbol) from which we may want to back-off all the time. For example, if we see the word string A <s> B (where <s> is a context cue indicating a sentence boundary), then instead of predicting the probability of B based on the full context (P(B | A <s>)), we may wish to disregard the information before the sentence boundary. Therefore we might want to back-off to the bigram distribution P(B | <s>) (inclusive forced back-off) or even to the unigram distribution P(B) (exclusive forced back-off). Version 2 supports both types of forced back-off for arbitrary vocabulary items.

The evallm program allows the user to specify either inclusive or exclusive forced back-off, as well as a list of words from which to enforce back-off.

Installing the Toolkit

For "big-endian" machines (eg those running HP-UX, IRIX, SunOS, Solaris) the installation procedure is simply to change into the src/ directory and type

make install
The executables will then be copied into the bin/ directory, and the library file SLM2.a will be copied into the lib/ directory. For "little-endian" machines (eg those running Ultrix, Linux) the variable BYTESWAP_FLAG will need to be set in the Makefile. This can be done by editing src/Makefile directly, so that the line
#BYTESWAP_FLAG  = -DSLM_SWAP_BYTES
is changed to
BYTESWAP_FLAG  = -DSLM_SWAP_BYTES

Then the program can be installed as before.

If you are unsure of the "endian-ness" of your machine, then the shell script endian.sh should be able to provide some assistance.

In case of problems, then more information can be found by examining src/Makefile.

Before building the executables, it might be worth adjusting the value of STD_MEM in the file src/toolkit.h. This value controls the default amount of memory (in MB) that the programs will attempt to assign for the large buffers used by some of the programs (this value can, of course, be overridden at the command line). The result is that the final process sizes will be a few MB bigger than this value. The more memory that can be grabbed, the faster the programs will run. The default value is 100, but if the machines which the tools will be run on contain less, or much more memory than this, then this value should be adjusted to reflect this.


Terminology and File Formats

Name Description Typical file extension
Text stream An ASCII file containing text. It may or may not have markers to indicate context cues, and white space can be used freely. .text
Word frequency file An ASCII file containing a list of words, and the number of times that they occurred. This list is not sorted; it will generally be used as the input to wfreq2vocab, which does not require sorted input. .wfreq
Word n-gram file ASCII file containing an alphabetically sorted list of n-tuples of words, along with the number of occurrences .w3gram, .w4gram etc.
Vocabulary file ASCII file containing a list of vocabulary words. Comments may also be included - any line beginning ## is considered a comment. The vocabulary is limited in size to 65535 words. .vocab.20K, .vocab.60K etc., depending on the size of the vocabulary.
Context cues file ASCII file containing the list of words which are to be considered "context cues". These are words which provide useful context information for the n-grams, but which are not to be predicted by the language model. Typical examples would be <s> and <p>, the begin sentence, and begin paragraph tags. .ccs
Id n-gram file ASCII or binary (by default) file containing a numerically sorted list of n-tuples of numbers, corresponding to the mapping of the word n-grams relative to the vocabulary. Out of vocabulary (OOV) words are mapped to the number 0. .id3gram.bin, .id4gram.ascii etc.
Binary language model file Binary file containing all the n-gram counts, together with discounting information and back-off weights. Can be read by evallm and used to generate word probabilities quickly. .binlm
ARPA language model file ASCII file containing the language model probabilities in ARPA-standard format. .arpa
Probability stream ASCII file containing a list of probabilities (one per line). The probabilities correspond the the probability for each word in a specific text stream, with context-cues and OOVs removed. .fprobs
Forced back-off file ASCII file containing a list of vocabulary words from which to enforce back-off, together with either an 'i' or an 'e' to indicate inclusive or exclusive forced back-off respectively. .fblist

These files may all be written are read by all the tools in compressed or uncompressed mode. Specifically, if a filename is given a .Z extension, then it will be read from the specified file via a zcat pipe, or written via a compress pipe. If a filename is given a .gz, it will be read from the specified file via a gunzip pipe, or written via a gzip pipe. If either of these compression schemes are to be used, then the relevant tools (ie zcat, and compress or gzip) must be available on the system, and pointed to by the path.

If a filename argument is given as - then it is assumed to represent either the standard input, or standard output (according to context). Any file read from the standard input is assumed to be uncompressed, and therefore, all desired compression and decompression should take place in a pipe: zcat < abc.Z | abc2xyz | compress > xyz.Z


The Tools

Note that in addition to the command line options mentioned, all the tools also support -help and -version.

text2wfreq

Input : Text stream

Output : List of every word which occurred in the text, along with its number of occurrences.

Notes : Uses a hash-table to provide an efficient method of counting word occurrences. Output list is not sorted (due to "randomness" of the hash-table), but can be easily sorted into the user's desired order by the UNIX sort command. In any case, the output does not need to be sorted in order to serve as input for wfreq2vocab.

Command Line Syntax:

text2wfreq [ -hash 1000000 ]
           [ -verbosity 2 ]
           < .text > .wfreq

Higher values for the -hash parameter require more memory, but can reduce computation time.

wfreq2vocab

Input : A word unigram file, as produced by text2wfreq

Output : A vocabulary file.

Command Line Syntax:

wfreq2vocab [ -top 20000 | -gt 10]
            [ -records 1000000 ]
            [ -verbosity 2]
            < .wfreq > .vocab

The -top parameter allows the user to specify the size of the vocabulary; if the program is called with the command -top 20000, then the vocabulary will consist of the most common 20,000 words.

The -gt parameter allows the user to specify the number of times that a word must occur to be included in the vocabulary; if the program is called with the command -gt 10, then the vocabulary will consist of all the words which occurred more than 10 times.

If neither the -gt, nor the -top parameters are specified, then the program runs with the default setting of taking the top 20,000 words.

The -records parameter allows the user to specify how many of the word and count records to allocate memory for. If the number of words in the input exceeds this number, then the program will fail, but a high number will obviously result in a higher memory requirement.

text2wngram

Input : Text stream

Output : List of every word n-gram which occurred in the text, along with its number of occurrences.

Command Line Syntax:

text2wngram [ -n 3 ]
            [ -temp /usr/tmp/ ]
            [ -chars n ]
            [ -words m ]
            [ -gzip | -compress ]
            [ -verbosity 2 ]
            < .text > .wngram

The maximum numbers of charactors and words that can be stored in the buffer are given by the -chars and -words options. The default number of characters and words are chosen so that the memory requirement of the program is approximately that of STD_MEM, and the number of charactors is seven times greater than the number of words.

The -temp option allows the user to specify where the program should store its temporary files.

text2idngram

Input : Text stream, plus a vocabulary file.

Output : List of every id n-gram which occurred in the text, along with its number of occurrences.

Notes : Maps each word in the text stream to a short integer as soon as it has been read, thus enabling more n-grams to be stored and sorted in memory.

Command Line Syntax:

text2idngram -vocab .vocab
           [ -buffer 100 ]
           [ -temp /usr/tmp/ ]
           [ -files 20 ]
           [ -gzip | -compress ]
           [ -n 3 ]
           [ -write_ascii ]
           [ -fof_size 10 ]
           [ -verbosity 2 ]
           < .text > .idngram 

By default, the id n-gram file is written out as binary file, unless the -write_ascii switch is used.

The size of the buffer which is used to store the n-grams can be specified using the -buffer parameter. This value is in megabytes, and the default value can be changed from 100 by changing the value of STD_MEM in the file src/toolkit.h before compiling the toolkit.

The program will also report the frequency of frequency of n-grams, and the corresponding recommended value for the -spec_num parameters of idngram2lm. The -fof_size parameter allows the user to specify the length of this list. A value of 0 will result in no list being displayed.

The -temp option allows the user to specify where the program should store its temporary files.

In the case of really huge quantities of data, it may be the case that more temporary files are generated than can be opened at one time by the filing system. In this case, the temporary files will be merged in chunks, and the -files parameter can be used to specify how many files are allowed to be open at one time.

ngram2mgram

Input : Either a word n-gram file, or an id n-gram file.

Output : Either a word m-gram file, or an id m-gram file, where m < n.

Command Line Syntax:

ngram2mgram -n N -m M
          [ -binary | -ascii | -words ]
          < .ngram > .mgram

The -binary, -ascii, -words correspond to the format of the input and output (Note that the output file will be in the same format as the input file). -ascii and -binary denote id n-gram files, in ASCII and binary formats respectively, and -words denotes a word n-gram file.

wngram2idngram

Input : Word n-gram file, plus a vocabulary file.

Output : List of every id n-gram which occurred in the text, along with its number of occurrences, in either ASCII or binary format.

Note : For this program to be successful, it is important that the vocabulary file is in alphabetical order. If you are using vocabularies generated by the wfreq2vocab tool then this should not be an issue, as they will already be alphabetically sorted.

Command Line Syntax:

wngram2idngram -vocab .vocab
              [ -buffer 100 ]
              [ -hash 200000 ]
              [ -temp /usr/tmp/ ]
              [ -files 20 ]
              [ -gzip | -compress ]
              [ -verbosity 2 ]
              [ -n 3 ]
              [ -write_ascii ]
              < .wngram > .idngram

The size of the buffer which is used to store the n-grams can be specified using the -buffer parameter. This value is in megabytes, and the default value can be changed from 100 by changing the value of STD_MEM in the file src/toolkit.h before compiling the toolkit.

The program will also report the frequency of frequency of n-grams, and the corresponding recommended value for the -spec_num parameters of idngram2lm. The -fof_size parameter allows the user to specify the length of this list. A value of 0 will result in no list being displayed.

Higher values for the -hash parameter require more memory, but can reduce computation time.

The -temp option allows the user to specify where the program should store its temporary files.

The -files parameter is used to specify the number of files which can be open at one time.

idngram2stats

Input : An id n-gram file (in either binary (by default) or ASCII (if specified) mode).

Output : A list of the frequency-of-frequencies for each of the 2-grams, ... , n-grams, which can enable the user to choose appropriate cut-offs, and to specify appropriate memory requirements with the -spec_num option in idngram2lm.

Command Line Syntax:

idngram2stats [ -n 3 ]
              [ -fof_size 50 ]
              [ -verbosity 2 ]
              [ -ascii_input ]
              < .idngram > .stats

mergeidngram

Input : A set of id n-gram files (in either binary (by default) or ASCII (if specified) format - note that they should all be in the same format, however).

Output : One id n-gram file (in either binary (by default) or ASCII (if specified) format), containing the merged id n-grams from the input files.

Notes : This utility can also be used to convert id n-gram files between ascii and binary formats.

Command Line Syntax:

mergeidngram [ -n 3 ]
             [ -ascii_input ]   
             [ -ascii_output ]   
             .idngram_1 .idngram_2 ... .idngram_N > .idngram

idngram2lm

Input : An id n-gram file (in either binary (by default) or ASCII (if specified) format), a vocabulary file, and (optionally) a context cues file. Additional command line parameters will specify the cutoffs, the discounting strategy and parameters, etc.

Output : A language model, in either binary format (to be read by evallm), or in ARPA format.

Command Line Syntax:

idngram2lm -idngram .idngram
           -vocab .vocab
           -arpa .arpa | -binary .binlm
         [ -context .ccs ]
         [ -calc_mem | -buffer 100 | -spec_num y ... z ]
         [ -vocab_type 1 ]
         [ -oov_fraction 0.5 ]
         [ -linear | -absolute | -good_turing | -witten_bell ]
         [ -disc_ranges 1 7 7 ] 
         [ -cutoffs 0 ... 0 ]
         [ -min_unicount 0 ]
         [ -zeroton_fraction 1.0 ]
         [ -ascii_input | -bin_input ]
         [ -n 3 ]  
         [ -verbosity 2 ]
         [ -four_byte_counts ]
         [ -two_byte_bo_weights
            [ -min_bo_weight -3.2 ] [ -max_bo_weight 2.5 ] 
            [ -out_of_range_bo_weights 10000 ] ]

The -context parameter allows the user to specify a file containing a list of words within the vocabulary which will serve as context cues (for example, markers which indicate the beginnings of sentences and paragraphs).

-calc_mem, -buffer and -spec_num x y ... z are options to dictate how it is decided how much memory should be allocated for the n-gram counts data structure. -calc_mem demands that the id n-gram file should be read twice, so that we can accurately calculate the amount of memory required. -buffer allows the user to specify an amount of memory to grab, and divides this memory equally between the 2,3, ..., n-gram tables. -spec_num allows the user to specify exactly how many 2-grams, 3-grams, ... , and n-grams will need to be stored. The default is -buffer STD_MEM.

The toolkit provides for three types of vocabulary, which each handle out-of-vocabulary (OOV) words in different ways, and which are specified using the -vocab_type flag.

A closed vocabulary (-vocab_type 0) model does not make any provision for OOVs. Any such words which appear in either the training or test data will cause an error. This type of model might be used in a command/control environment where the vocabulary is restricted to the number of commands that the system understands, and we can therefore guarantee that no OOVs will occur in the training or test data.

An open vocabulary model allows for OOVs to occur; out of vocabulary words are all mapped to the same symbol. Two types of open vocabulary model are implemented in the toolkit. The first type (-vocab_type 1) treats this symbol the same way as any other word in the vocabulary. The second type (-vocab_type 2) of open vocabulary model is to cover situations where no OOVs occurred in the training data, but we wish to allow for the situation where they could occur in the test data. This situation could occur, for example, if we have a limited amount of training data, and we choose a vocabulary which provides 100% coverage of the training set. In this case, an arbitrary proportion of the discount probability mass (specified by the -oov_fraction option) is reserved for OOV words.

The discounting strategy and its parameters are specified by the -linear, -absolute, -good_turing and -witten_bell options. With Good Turing discounting, one can also specify the range over which discounting occurs, using the -disc_ranges option.

The user can specify the cutoffs for the 2-grams, 3-grams, ..., n-grams by using the -cutoffs parameter. A cutoff of K means that > n-grams occurring K or fewer times are discarded. If the parameter is omitted, then all the cutoffs are set to zero.

The -zeroton_fraction option specifies that P(zeroton) (the unigram probability assigned to a vocabulary word that did not occurred at all in the training data) will be at least that fraction of P(singleton) (the probability assigned to a vocabulary word that occurred exactly once in the training data).

By default, the n-gram counts are stored in two bytes by use of a count table (this allows the counts to exceed 65535, while keeping the data structures used to store the model compact). However, if more than 65535 distinct counts need to be stored (very unlikely, unless constructing 4-gram or higher language models using Good-Turing discounting), the -four_byte_counts option will need to be used.

The floating point values of the back-off weights may be stored as two-byte integers, by using the -two_byte_alphas switch. This will introduce slight rounding errors, and so should only be used if memory is short. The -min_alpha, -max_alpha and -out_of_range_alphas are parameters used by the functions for using two-byte alphas. Their values should only be altered if the program instructs it. For further details, see the comments in the source file src/two_byte_alphas.c.

binlm2arpa

Input : A binary format language model, as generated by idngram2lm.

Output : An ARPA format language model.

Command Line Syntax:

binlm2arpa -binary .binlm
           -arpa .arpa 
         [ -verbosity 2 ]

evallm

Input : A binary or ARPA format language model, as generated by idngram2lm. In addition, one may also specify a text stream to be used to compute the perplexity of the language model. The ARPA format language model does not contain information as to which words are context cues, so if an ARPA format lanaguage model is used, then a context cues file may be specified as well.

Output : The program can run in one of two modes.

Command Line Syntax:

evallm [ -binary .binlm | 
         -arpa .arpa [ -context .ccs ] ]

Notes: evallm can receive and process commands interactively. When it is run, it loads the language model specified at the command line, and waits for instructions from the user. The user may specify one of the following commands:

Since the commands are read from standard input, a command file can be piped into it directly, thus removing the need for the program to run interactively:

echo "perplexity -text b.text" | evallm -binary a.binlm

interpolate

Input : Files containing probability streams, as generated by the -probs option of the perplexity command of evallm. Alternatively these probabilities could be generated from a seperate piece of code, which assigns word probabilities according to some other language model, for example a cache-based LM. This probability stream can then be linearly interpolated with one from a standard n-gram model using this tool.

Output : An optimal set of interpolation weights for these probability streams, and (optionally) a probability stream corresponding to the linear combination of all the input streams, according to the optimal weights. The optimal weights are calculated using the expectation maximisation (EM) algorithm.

Command Line Syntax :

interpolate +[-] model1.fprobs +[-] model2.fprobs ... 
        [ -test_all | -test_first n | -test_last n | -cv ]
        [ -tag .tags ]
        [ -captions .captions ]
        [ -in_lambdas .lambdas ]
        [ -out_lambdas .lambdas ]
        [ -stop_ratio 0.999 ]
        [ -probs .fprobs ]
        [ -max_probs 6000000 ]

The probability stream filenames are prefaced with a + (or a +- to indicate that the weighting of that model should be fixed).

There are a range of options to determine which part of the data is used to calculate the weights, and which is used to test them. One can test the perplexity of the interpolated model based on all the data, using the -test_all option, in which case a set of lambdas must also be specified with the -lambda option (ie the lambdas are pre-specified, and not calculated by the program). One can specify that the first or last n items are the test set by use of the -test_first n or -test_last n options. Or one can perform two-way cross validation using the -cv option. If none of these are specified then the whole of the data is used for weight estimation.

By default, the initial interpolation weights are fixed as 1/number_of_models, but alternative values can be stored in a file and used via the -in_lambdas option.

The -probs switch allows the user to specify a filename in which to store the combined probability stream. The optimal lambdas can also be stored in a file by use of the -out_lambdas command.

The program stops when the ratio of the test-set perplexity between two successive iterations is above the value specified in the -stop_ratio option.

The data can be partitioned into different classes (with optimisation being performed seperately on each class) using the -tags parameter. The tags file will contain an integer for each item in the probability streams corresponding to the class that the item belongs to. A file specified using the -captions option will allow the user to attach names to each of the classes. There should be one line in the captions file for each tag, with each line corresponding to the name of the tag.

The amount of memory allocated to store the probability streams is dictated by the -max_probs option, which indicates the maximum number of probabilities allowed in one stream.

Note: For an example use and output of a previous version of this program (with slightly different syntax), see Appendix B of R. Rosenfeld Adaptive Statistical Language Modeling: A Statistical Approach PhD Thesis, School of Computer Science, Carnegie Mellon University, April 1994. Published as Techical Report CMU-CS-94-138


Typical Usage

Simplified toolkit framework - 8KB

Given a large corpus of text in a file a.text, but no specified vocabulary

Alternatively, some of these processes can be piped together:

cat a.text | text2wfreq | wfreq2vocab -top 20000 > a.vocab
cat a.text | text2idngram -vocab a.vocab | \
   idngram2lm -vocab a.vocab -idngram - \
   -binary a.binlm -spec_num 5000000 15000000
echo "perplexity -text b.text" | evallm -binary a.binlm 

Discounting Strategies

Discounting is the process of replacing the original counts with modified counts so as to redistribute the probability mass from the more commonly observed events to the less frequent and unseen events. If the actual number of occurrences of an event E (such as a bigram or trigram occurrence) is c(E), then the modified count is d(c(E))c(E), where d(c(E)) is known as the discount ratio.

Good Turing discounting

Good Turing discounting defines d(r) = (r+1)n(r+1) / rn(r) where n(r) is the number of events which occur r times.

The discounting is only applied to counts which occur fewer than K times, where typically K is chosen to be around 7. This is the "discounting range" which is specified using the -disc_ranges parameter of the idngram2lm program.

For further details see "Estimation of Probabilities from Sparse Data for the Language Model Component of a Speech Recognizer", Slava M. Katz, in "IEEE Transactions on Acoustics, Speech and Signal Processing", volume ASSP-35, pages 400-401, March 1987.

Witten Bell discounting

The discounting scheme which we refer to here as "Witten Bell discounting" is that which is referred to as type C in "The Zero-Frequency Problem: Estimating the Probabilities of Novel Events in Adaptive Text Compression", Ian H. Witten and Timothy C. Bell, in "IEEE Transactions on Information Theory, Vol 37, No. 4, July 1991".

The discounting ratio is not dependent on the event's count, but on t, the number of types which followed the particular context. It defines d(r,t) = n/(n + t), where n is the size of the training set in words. This is equivalent to setting P(w | h) = c / (n + t) (where w is a word, h is the history and c is the number of occurrences of w in the context h), for events that have been seen, and P(w | h) = t / (n + t) for unseen events.

Absolute discounting

Absolute discounting defines d(r) = (r-b)/r. Typically b=n(1)/(n(1)+2n(2)). The discounting is applied to all counts.

This is, of course, equivalent to simply subtracting the constant b from each count.

Linear discounting

Linear discounting defines d(r) = 1 - (n(1)/C), where C is the total number of events. The discounting is applied to all counts.

For further details of both linear and absolute discounting, see "On structuring probabilistic dependencies in stochastic language modeling", H. Ney, U. Essen and R. Kneser in "Computer Speech and Language", volume 8(1), pages 1-28, 1994.


Up-to-date Information

The latest news on updates, bug fixes etc. can be found here.

Feedback

Any comments, questions or bug reports concerning the toolkit should be addressed to Philip Clarkson.


Philip Clarkson - prc14@eng.cam.ac.uk