- SEE ALSO
This document is out of date as of version 0.67. Please consult the documentation in measure2d.pm and measure3d.pm. The following is provided for historical purposes only.
How to create a new statistics package for the Ngram Statistics Package.
The following steps should be followed while creating a new statistic library package for NSP.
The filename should have an extension of .pm. Usually the name of the file should be Statistic.pm, where "Statistic" is the name of the particular statistic you are writing.
Let us say you have named your file Statistic.pm. The first line of the file should declare that its a package of the same name as the filename. Thus the first line of the file Statistic.pm should be...
You need to implement at least two functions in your package
i) initializeStatistic() ii) calculateStatistic()
Function initializeStatistic() is passed the following parameters:
1) The ngram size. eg: 2 (for bigrams), 3 (for trigrams) etc. 2) The total number of bigrams in the corpus. 3) The number of frequency combinations 4) An array containing the frequency combinations.
The fourth data structure above may be accessed as a double dimensioned array in which each row represents a single frequency combination. On a given row, the first element denotes the number of indices on this row, say 'n'. This is followed by 'n' numbers representing the 'n' indices that make up this frequency combination. (For details on frequency combinations, see README.pod).
Thus for example say we are passing the default frequency combinations for trigrams. There are 7 combinations in the default. Thus the third item passed above will be '7'. After this, the following two dimensioned array would be passed:
Row 1: 3 0 1 2 Row 2: 1 0 Row 3: 1 1 Row 4: 1 2 Row 5: 2 0 1 Row 6: 2 0 2 Row 7: 2 1 2
(The "Row X:" parts are for explanation purposes are not passed). Thus row 1 start with the number '3' which says that there are 3 more numbers on this row, 0, 1, and 2. Similarly, row two 2 starts with '1' and then has one number after it: 0. And so on.
This function is called before any calls to the function calculateStatistic() and can be used by the statistic library to set up any values that may be required for the calculations later. For example, many statistical measures require the corpus size, and so this would be a good place to save that value (the second item passed above). Also, since the frequencies passed to the calculateStatistic() function below follow the order defined through the frequency combination array passed above, it is important to note which indices are to be used for the calculation. See dice.pm for an example of one way to do this.
This function is not expected to return anything. If an error occurs, it can be reported through the mechanisms described below.
The other mandatory function is calculateStatistic(). This is passed an array containing the frequency values for an ngram as found in the input n-gram file. The size of this array is guaranteed to be exactly the same as the third number passed to the initializeStatistic() function above.
Function calculateStatistic() is expected to return a (possibly floating) value as the value of the statistical measure calculated using the frequency values passed to it.
When a library is loaded, statistic.pl checks for these two functions: if they are not implemented, then an error is reported and the program quits.
Program statistic.pl also supports three other functions that are not mandatory, but may be implemented by the user. These are:
i) errorCode() ii) errorString() iii) getStatisticName()
Function errorCode, if implemented, is called immediately after the call to function initializeStatistic() and immediately after every call to function calculateStatistic().
This function should:
a) return 0 to imply that the last operation was successful. b) return an integer starting with 1 to imply that the last operation was unsuccessful, that there has been a fatal error and that statistic.pl should abort. c) return an integer starting with 2 to imply that the last operation was unsuccessful but it is not a fatal error, just a warning. Program statistic.pl will not abort on error codes starting with 2. If there is a warning after a call to the function calculateStatistic(), then the ngram for which the warning was issued will be ignored by statistic.pl.
If a non-zero code is returned by function errorCode(), statistic.pl will print to STDERR the message "Error from statistic library!", if the error code starts with 1, or the message "Warning from statistic library" if the error starts with 2. Then, statistic.pl will print the actual error code returned. Finally, if function errorString() has been defined, this function will be called. This function may be implemented by the user to return a wordy description of the error or warning; the string returned by this function is then printed to STDERR.
Note that functions errorCode() and errorString() should be implemented in such a way that they reset the error code and the error message respectively after a call to the function. This will prevent mistakenly reporting a warning more than once.
The third function that may be implemented is getStatisticName(). If this function is implemented, it is expected to return a string containing the name of the statistic being implmented. This string is used in the formatted output of statistic.pl. If this function is not implemented, then the statistic file name entered on the commandline is used in the formatted output.
Note that all three functions described in this section are first checked for existence before being called. So, if the user elects to not implement these functions, no harm will be done. However, we strongly recommend the implementation of at least the function errorCode() since this is the only way for the statistic library to report errors to the user.
Having implemented the two mandatory functions (in point 3 above) and zero or more of the three non-mandatory functions (in point 4 above), one must make these functions available outside the package. To do so, one has to export them, thusly.
For this, first include the Exporter package by including the following line in the program
Now include the following line to inherit Exporter's functions:
@ISA = qw ( Exporter );
Now export the various functions implemented so that they are accessible outside this package, by adding the following line (assume that you have implemented only the two mandatory functions):
@EXPORT = qw( initializeStatistic calculateStatistic );
If you implement say the errorCode() and errorString() functions too, you may export them like so:
@EXPORT = qw( initializeStatistic calculateStatistic errorCode errorString );
Note that the user may implement other functions too, and may export them if he so wishes, but since statistic.pl is not expecting anything besides the five functions above, doing so would have no effect on statistic.pl.
Finally, at the end of everything, add the line
1; This will ensure that the LAST line of the file returns a true value, and is necessary so that when this package is loaded, it returns a TRUE value.
The filename does not end with a .pm.
The rest of the filename (besides the extension) does not match the package name (declared in the first line of the file). Remember its case sensitive!
The five functions (2 mandatory, 3 non-mandatory) must have their names match EXACTLY with those shown above. Again, names are all case sensitive.
The last line of the file is not "1;". This is necessary, and easily overlooked!
Ted Pedersen (firstname.lastname@example.org) Satanjeev Banerjee (email@example.com)
home page: http://www.d.umn.edu/~tpederse/nsp.html mailing list: http://groups.yahoo.com/group/ngram/
Copyright (C) 2000-2001, Satanjeev Banerjee and Ted Pedersen
Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.
Note: a copy of the GNU Free Documentation License is available on the web at http://www.gnu.org/copyleft/fdl.html and is included in this distribution as FDL.txt.