C# (CSharp) Lucene.Net.Analysis.Shingle Namespace

A ShingleFilter constructs shingles (token n-grams) from a token stream. In other words, it creates combinations of tokens as a single token.

For example, the sentence "please divide this sentence into shingles" might be tokenized into shingles "please divide", "divide this", "this sentence", "sentence into", and "into shingles".

This filter handles position increments > 1 by inserting filler tokens (tokens with termtext "_"). It does not handle a position increment of 0.

An instance of this class is used to maintain the number of input stream tokens that will be used to compose the next unigram or shingle: #gramSize.

gramSize will take on values from the circular sequence { [ 1, ] #minShingleSize [ , ... , #maxShingleSize ] }.

1 is included in the circular sequence only if #outputUnigrams = true.

A ShingleMatrixFilter constructs shingles (token n-grams) from a token stream. In other words, it creates combinations of tokens as a single token.

For example, the sentence "please divide this sentence into shingles" might be tokenized into shingles "please divide", "divide this", "this sentence", "sentence into", and "into shingles".

Using a shingle filter at index and query time can in some instances be used to replace phrase queries, especially them with 0 slop.

Without a spacer character it can be used to handle composition and decomposition of words such as searching for "multi dimensional" instead of "multidimensional". It is a rather common human problem at query time in several languages, notably the northern Germanic branch.

Shingles are amongst many things also known to solve problems in spell checking, language detection and document clustering.

This filter is backed by a three dimensional column oriented matrix used to create permutations of the second dimension, the rows, and leaves the third, the z-axis, for for multi token synonyms.

In order to use this filter you need to define a way of positioning the input stream tokens in the matrix. This is done using a ShingleMatrixFilter.TokenSettingsCodec. There are three simple implementations for demonstrational purposes, see ShingleMatrixFilter.OneDimensionalNonWeightedTokenSettingsCodec, ShingleMatrixFilter.TwoDimensionalNonWeightedSynonymTokenSettingsCodec and ShingleMatrixFilter.SimpleThreeDimensionalTokenSettingsCodec.

Consider this token matrix:

 Token[column][row][z-axis]{ {{hello}, {greetings, and, salutations}}, {{world}, {earth}, {tellus}} }; 

 "hello_world" "greetings_and" "greetings_and_salutations" "and_salutations" "and_salutations_world" "salutations_world" "hello_earth" "and_salutations_earth" "salutations_earth" "hello_tellus" "and_salutations_tellus" "salutations_tellus" 

This implementation can be rather heap demanding if (maximum shingle size - minimum shingle size) is a great number and the stream contains many columns, or if each column contains a great number of rows.

The problem is that in order avoid producing duplicates the filter needs to keep track of any shingle already produced and returned to the consumer.

There is a bit of resource management to handle this but it would of course be much better if the filter was written so it never created the same shingle more than once in the first place.

The filter also has basic support for calculating weights for the shingles based on the weights of the tokens from the input stream, output shingle size, etc. See CalculateShingleWeight.

NOTE: This filter might not behave correctly if used with custom Attributes, i.e. Attributes other than the ones located in org.apache.lucene.analysis.tokenattributes.