commons-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF GitHub Bot (JIRA)" <j...@apache.org>
Subject [jira] [Work logged] (TEXT-155) Add a generic SetSimilarity measure
Date Sat, 09 Mar 2019 12:01:00 GMT

     [ https://issues.apache.org/jira/browse/TEXT-155?focusedWorklogId=210522&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-210522
]

ASF GitHub Bot logged work on TEXT-155:
---------------------------------------

                Author: ASF GitHub Bot
            Created on: 09/Mar/19 12:00
            Start Date: 09/Mar/19 12:00
    Worklog Time Spent: 10m 
      Work Description: aherbert commented on issue #109: TEXT-155: Add a generic IntersectionSimilarity
measure
URL: https://github.com/apache/commons-text/pull/109#issuecomment-471171172
 
 
   Hi Bruno,
   
   I'll cover most things in this comment.
   
   1. Name
   
   Since this class computes the intersect I called it `IntersectionSimilarity`. However it
also computes the union. A separate class for compute the union does not makes sense since
the two things are intertwined. It is a classic [Venn diagram](https://en.wikipedia.org/wiki/Venn_diagram)
of two sets. If you know the size of each set then if you compute the intersect you can know
the union and vice versa.
   
   How about `OverlapSimilarity`?
   
   2. Computing the Jaccard + Sorenson Dice
   
   I had added it mainly during testing to check against known results. But there are a lot
more metrics than can be computed using the true-positive (intersect) and false-positive (size
A - intersect) and false negatives (size B - intersect). Since we are not going to add them
all then change to:
   
   `OverlapResult` - defines set size A, B, intersect and union (defined using the other 3
values).
   
   It could also define falsePositives (using A) and falseNegatives (using B) as utility methods.
Or leave them out as we do not need them at the moment.
   
   3. Shared functionality
   
   This class can then be used within existing similarity scores in the library (Jaccard)
and for the new Sorensen-Dice similarity. I thought that was the intension.
   
   It will be up to those classes to appropriately use the `OverlapResult` to compute their
score.
   
   Alex
   
   
   
 
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


Issue Time Tracking
-------------------

    Worklog Id:     (was: 210522)
    Time Spent: 3h 10m  (was: 3h)

> Add a generic SetSimilarity measure
> -----------------------------------
>
>                 Key: TEXT-155
>                 URL: https://issues.apache.org/jira/browse/TEXT-155
>             Project: Commons Text
>          Issue Type: New Feature
>    Affects Versions: 1.6
>            Reporter: Alex D Herbert
>            Priority: Minor
>          Time Spent: 3h 10m
>  Remaining Estimate: 0h
>
> The {{SimilarityScore<T>}} interface can be used to compute a generic result. I
propose to add a class that can compute the intersection between two sets formed from the
characters. The sets must be formed from the {{CharSequence}} input to the {{apply}} method
using a {{Function<CharSequence, Set<T>>}} to convert the {{CharSequence}}. This
function can be passed to the {{SimilarityScore<T>}} during construction.
> The result can then be computed to have the size of each set and the intersection.
> I have created an implementation that can compute the equivalent of the {{JaccardSimilary}}
class by creating {{Set<Character>}} and also the F1-score using bigrams (pairs of characters)
by creating {{Set<String>}}. This relates to [Text-126|https://issues.apache.org/jira/projects/TEXT/issues/TEXT-126]
which suggested an algorithm for the Sorensen-Dice similarity, also known as the F1-score.
> Here is an example:
> {code:java}
> // Match the functionality of the JaccardSimilarity class
> Function<CharSequence, Set<Character>> converter = (cs) -> {
>     final Set<Character> set = new HashSet<>();
>     for (int i = 0; i < cs.length(); i++) {
>         set.add(cs.charAt(i));
>     }
>     return set;
> };
> IntersectionSimilarity<Character> similarity = new IntersectionSimilarity<>(converter);
> IntersectionResult result = similarity.apply("something", "something else");
> {code}
> The result has the size of set A, set B and the intersection between them.
> This class was inspired by my look through the various similarity implementations. All
of them except the {{CosineSimilarity}} perform single character matching between the input
{{CharSequence}}s. The {{CosineSimilarity}} tokenises using whitespace to create words.
> This more generic type of implementation will allow a user to determine how to divide
the {{CharSequence}} but to create the sets that are compared, e.g. single characters, words,
bigrams, etc.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message