So far from our posts we have seen that there are several tools for text mining. For example, we used machine based learning with tagged corpus and ontology. Vast collection of text has been studied and prepared in this corpus and a comprehensive collection of words has been included in the ontology. This gives us great resource to work with any document. Next we define different distance vectors and use clustering techniques to group and extract keywords and topics. We have refined the distance vectors and data points to be more representative of the content of the text. There have been several ways to measure distance or similarity between words and we have seen articulation of probability based measures. We have reviewed the way we cluster these data points and found out methods that we prefer over others.
We want to remain focused on keyword extraction even though we have seen similar usages in topic analysis and some interesting areas as text segmentation. We don't want to resort to a large corpus for light weight application plugins but we don't mind a large corpus for database searches.  We don't want processing that is better than O(N^2) in working with the data to extract keywords and we have the luxury to have a pipeline of steps to get to the keywords.

Writing powershell commands
Powershell lets you invoke CmdLets on the command line. Custom CmdLets are an instance of a .Net class. A CmdLet processes its input from an object pipeline instead of text. A CmdLet processes one object at a time. CmdLets are attributed with a CmdLetAttribute and named with a  verb-noun pair. The class derives from PSCmdLet which gives you access to PS runtime. The custom cmdLet class could also derive from CmdLet in which case it's more light weight. CmdLets don't handle argument parsing and error handling. These are done consistently across all Powershell CmdLets.
CmdLets support ShouuldProcess Parameter which lets the class have access to runtime behavior parameters - Confirm and WhatIf. Confirm specifies whether user confirmation is required. WhatIf informs the user what changes would have been made when the CmdLet is invoked.
Common methods to override include BeginProcessing which provides pre-processing functionality for the cmdlet, ProcessRecord which can be called any number of times, EndProcessing for post-processing functionality and StopProcessing when the user stops the cmdLet asynchronously.
CmdLet parameters allow the user to provide input into the CmdLet. This is done by adding properties to the class that implements the CmdLet and adding ParameterAttribute to them.
ProcessRecord generally does the work of creating new entries for data.
Parameters must be explicitly marked as public.  Parameters can be positional or named. If the parameter is positional, only the value is provided with the CmdLet invocation.  In addition, parameters can be marked as mandatory which means that they have a value assigned.
Some parameters are reserved  and are often referred to as Common parameters.  Another group of parameters are called the ShouldProcess parameters which give access to the Confirm and WhatIf runtime support. Parameters Sets are also supported by Powershell which refers to a grouping of the parameters.
For exception handling, a try catch can be added to the class method invocation. These should be to add more information when the error happens. If you don't want to stop the pipeline on error, then do not throw with ThrowTerminatingError.
Results are reported through objects. Powershell is emphatic on the way results are displayed and there's a lot of flexibility in what you want to include in your result objects. WriteObject is what is used to emit the results. These results can be returned to the pipeline. As with parameters, there should be consistency in the usage of both results and parameters.
There should be support for diagnostics when things go wrong so that the problem can be identified quickly and resolved. There is builtin support to send messages to the host application which could be powershell.exe and that displays the messages to the pipeline.
CmdLets can also be grouped so that the parameters or results need not be repeated. This is very convenient when there are fine grained CmdLets required but they essentially belong to the same group.  A snap in can also be created with PSSnapIn so that the CmdLets are registered for usage. These are available from the System.Management.Automation namespace. Installing a snap in is done via InstallUtil.exe  which creates some registry entries. Make sure that System.Management.Automation.dll is available from the SDK or the Global Assembly Cache (GAC).

I learned today that expressions and queries should be treated different. Even though you can have the predicate in a query as an expression tree and vice versa, there are several reasons to use one or the other specifically in certain scenarios.

Nltk classifier modules
The nltk decision tree module. This is a classifier model. A decision tree comprises of non-terminal nodes for conditions on feature values and the terminal nodes for the labels. The tree evaluates to a label for a given token.
The module requires feature names and labeled feature sets. Different thresholds such as for depth cut off, entropy cut off and support cut off can also be specified. Entropy refers to degree of randomness or variations in the results while support refers to the number of feature sets used for evaluation.

The term feature is used to refer to some property of an unlabeled token. Typically a token is a word from a text that we have not seen before. If the text is seen before and has already been labeled, it is a training set. Training set helps train our model so that we can pick the labels better for the tokens we encounter. As an example, the proper nouns for names may be labeled male or female. We start with a large collection of already tagged names, we call training data. We build a model where we say if the name ends with a certain set of suffixes, the name is that of a male. Then we run our model on the training data to see how accurate we were and we adjust our model to improve our accuracy. Next we can run our model on a test data. If a name from the test data is labeled by this model as a male, we know its likelihood to be correct.

The property of labeled tokens are also helpful. We call these as joint-features and we distinguish it from the feature we just talked about by referring to the latter as input-features. So joint-features belong to training data and input-features belong to test data. For some classifiers such as the maxent classifier we refer to these as features and contexts respectively. The maxent stands for maximum entropy model where joint-features are required to have numeric values and input-features are mapped to a set of joint-features.

 There are other types of classifiers as well. For example, the mallet package uses the external mallet machine learning package. The megam module uses the external megam maxent optimization package. The naïve Bayes module is a module that assigns probability for a label. The P(label/features) is computed as P(label) * P(features/label) / P(features).  The 'naive' assumption is that all features are independent.  The positivenaivebayes module is a variant of the Bayes classifier

 that performs binary classification based on two complementary classes  where we have labeled examples only for one of the classes. The there are classifiers based exclusively on the corpus that they are trained on. The rte_classify module is a simple classifier for the RTE corpus.  It calculates the overlap in words and named entities between text and hypothesis. Most of the classifiers discussed are built on top of the scikit machine learning library.

Nltk cluster package
This module contains a number of basic clustering algorithms . Clustering is unsupervised machine learning to group similar items with a large collection There are the k-means clustering, E-M clustering and a group average agglomerative clustering. The K-means clustering starts with the k arbitrary chosen means and assigns each vector to the cluster with the closest mean. The centroid of the cluster is recalculated as the means of each cluster. The process is repeated until the clusters stabilize. This may converge to a local maximum so this method is repeated for other random initial means and the most common occurring output is chosen. The Gaussian EM clustering starts with k arbitrarily chosen means, prior probabilities and co-variance matrices which forms the parameters for the Gaussian source. The membership probabilities is then calculated for each vector in each of the clusters - this is the E-step. The parameters are then updated in the M-step using the maximum likelihood estimate from the clustering membership probabilities. This process continues until the likelihood of the data does not significantly increase.  The GAAC clustering starts with each of the N vectors as singleton clusters and then iteratively merges pairs of clusters which have the closest centroids. This continues until there is only one cluster. The order of merges is useful in finding the membership of a given number of clusters because earlier merges are lower than the depth c in the resulting tree.

Usage of WixSharp to build MSI for installing and uninstalling applications.
WixSharp makes it easy to author the logic for listing all the dependencies of your application for deployment. It converts the C# code to wxs file which in turn is compiled to build the MSI. There are a wide variety of samples in the WixSharp toolkit. Some of them require very few lines to be written for a wide variety of deployment time actions. The appeal in using such libraries is to be able to get the task done sooner with few lines of code. The earlier way of writing and editing WXS file was error prone and tedious.

nifty way to find file format

Here's something I found on the internet. If you wanted know the file format and there is little or no documentation of the proprietary format, you can look up the header of the file  and other data structures with the corresponding symbols from kernel PDB file. You don't need the source to look up format.
There are tools that can render the information in the PDB to be navigated through the UI. This is possible via the DIA SDKs. The format of PDB files is also not open hence their access is via debugger sdk. The SDK is available via COM. So you may have to register the DIA dll.
The debuggers make a local cache of the symbols when requested and they download the symbols from the symbol server, so you can expect the PDB to be made available to you by the debugger in the directory you specify.
If you look at the kernel PDB, you will find the structures we are looking for, start with the name MINIDUMP and these can be walked from the header onwards.
To find the stack trace, we follow the header to the directory or stream for the exception and read the exception record and address where the exception occurred. Both of these are given in the MINIDUMP_EXCEPTION data structure. The exception stream also gives the thread context. The context gives the processor specific register data. When we dump the stack pointer, we get the stack trace. We resolve the symbols of the stack trace with the pdbs either explicitly through DIA or implicitly through the managed mdbgeng library of debugger sdk. The minidump actually has all the information in the file itself. For example, you can list all the loaded modules with the mindump module list as well as the mindump module structures. Module name, function name, line number and stack frame are available via IMAGEHLP data structures. The various types of streams in the minidump are :
Thread list stream given by the MINIDUMP_THREAD_LIST structure
Module list stream given by the MINIDUMP_MODULE_LIST structure
Memory list stream given by the MINIDUMP_MEMORY_LIST structure
Exception stream given by the MINIDUMP_EXCEPTION_STREAM structure
System Info stream given by the MINIDUMP_SYSTEM_INFO structure
ThreadExList stream given by the MINIDUMP_THREAD_EX_LIST structure
Memory64 list stream given by the MINIDUMP_MEMORY64_LIST structure
Comment stream
Handle Data stream given by the MINIDUMP_HANDLE_DATA_STREAM structure
Function table stream given by the MINIDUMP_FUNCTION_TABLE structure
Unloaded module list stream given by the MINIDUMP_UNLOADED_MODULE_LIST structure
Misc Info List stream given by the MINIDUMP_MISC_INFO structure
Thread Info List stream given by the MINIDUMP_THREAD_INFO_LIST structure
Handle operation  list stream given by the MINIDUMP_HANDLE_OPERATION_LIST structure
For diagnostics, we may choose to display messages to the output or error stream. More features can be built into the tool that retrieves the stack trace from the dump. This can be done in an extensible manner where tool runs a set of commands from the user by way of command line. Internally we can have a command pattern to implement the different debugger like functionalities of the tool. Also the tool can be deployed via MSI. This ensures cleanliness during install and uninstall. 

Review : Paper on comparision of document clustering techniques by Steinbach, Karypis and Kumar from Univerity of Minnesota.
This paper compares k-means clustering methods to hierarchical clustering methods. The paper suggests that bisecting k-means technique is better than the standard k-means technique which is in turn better than hierarchical techniques.
Hierarchical clustering has quadratic time complexity where as K-means have a time complexity that's linear. There are mixed approaches too.
There are two metrics used for cluster quality analysis. Entropy is one which provides a measure of goodness for single level clusters. F-measure is the other which measures the effectiveness of hierarchical clustering.
The bisecting k-means clustering is explained as follows:
Step 1 pick a cluster to split
Step 2 find two clusters using the basic k-means algorithm
Step 3 Repeat step 2 for a fixed number of times and take the split that produces the clustering with the highest overall similarity
Step 4 Repeat step 1, 2 and 3 until the desired number of clusters are reached.
Splitting the largest cluster also works.

Agglomerative hierarchical clustering have the following variations:
1) Intra cluster similarity : This hierarchical clustering looks at the similarity of all documents in the cluster to the centroid where the similarity distance is given as the sum of cosines. The pair of clusters that when merged leads to smallest decrease in similarity.

2) Centroid similarity technique: This works in a similar way but it takes the similarity distance as the cosine between the centroids of the two clusters.

3) The UPGMA scheme is based on the cluster similarity measure which takes the sum of the cosine distances between two documents of different clusters divided by the product of the sizes of their clusters,.

An explanation is given for why the agglomerative hierarchical clustering performs poorly when compared with bisecting k-means that mentions that the former puts documents of the same class in the same cluster and this is done early on and generally not reversed.