Databases
StackOverflow2013
Postsdbo
POText classification using SVM works with unigrams but not higher order n-grams
 

Note that there are some explanatory texts on larger screens.

plurals
  1. POText classification using SVM works with unigrams but not higher order n-grams
    primarykey
    data
    text
    <p>I'm using LibSVM (in Java fwiw) to classify text samples into one of two categories: english or spanish language. I'm training on three texts in each language, for a total of roughly 50,000 words each. I'm then testing on a number of shorter texts and checking for appropriate classification. Some of the testing data is drawn from the training data (trivial, but essentially done as a sanity check) and the rest is new.</p> <p>To build the SVM vectors, I have been parsing text into ngrams, and then hashing these ngrams to get numerical representations. For instance, the following vector:</p> <pre><code>2.0 1:9.0 2:3.0 3:1.0 4:7.0 5:4.0 ... </code></pre> <p>with label 2 implies 9 ngrams hashed to value 1, 3 ngrams hashed to value 2, and so on.</p> <p>This has been working well for me when using unigrams, but for some reason as soon as I switch to bigrams or higher order n-grams, classification entirely fails. Can you think of any reason why this might be the case? The size of my feature set is bounded at 4999 (ie. I mod each hash so that it is no bigger than 4999). I've tried increasing and decreasing this bound but to no avail.</p> <p>Does anybody know where the problem might be coming from? Might my corpora be too small, or is it a problem with my approach to tokenization / building feature vectors?</p> <p>Thanks in advance for your help.</p>
    singulars
    1. This table or related slice is empty.
    1. This table or related slice is empty.
    1. PTQuestion
    1. This table or related slice is empty.
    1. USGeoffroy
    plurals
    1. This table or related slice is empty.
    1. This table or related slice is empty.
    1. This table or related slice is empty.
    1. This table or related slice is empty.
    1. VO
      singulars
      1. POText classification using SVM works with unigrams but not higher order n-grams
      1. USRyan R. Rosario
      1. VTFavorite
    2. VO
      singulars
      1. POText classification using SVM works with unigrams but not higher order n-grams
      1. USFrank Visaggio
      1. VTFavorite
    1. COMy first thought is that perhaps you are getting too many hash conflicts? What hash algorithm are you using? I would sanity check the hash and see how often you get two different n-grams hashing to the same value.
      singulars
      1. POText classification using SVM works with unigrams but not higher order n-grams
      1. USkarenu
    2. COThe function I'm using is relatively trivial: `Math.abs(text.hashCode() % _hash_size)` where `_hash_size` is the bound on the feature set size (ie. 4999). I haven't actually measured collision rate, but I assume the collisions are pretty low before I do the mod operation. As for quantifying the impact of the mod operation, I tried increasing the feature set size (ie. to ~100k) and this didn't any difference.
      singulars
      1. POText classification using SVM works with unigrams but not higher order n-grams
      1. USGeoffroy
    3. COSure, by modding the result you are causing collisions but the collisions are only partially related to the feature size, without the mod if you have over 50,000 different n-grams with java's hashCode you will start getting collisions. I'm assuming if you have several texts with over 50,000 words, you have more n-grams than that. Hash functions are not all created equal, I would use a cryptographic hash function like MD5 which has a much higher collision rate. I also wouldn't cut the feature set down by modding it like that - use a feature selection algorithm.
      singulars
      1. POText classification using SVM works with unigrams but not higher order n-grams
      1. USkarenu
 

Querying!

 
Guidance
A row detail

Detail views are divided into sections. All the information in the data section comes from columns in the selected row. The other sections display data from other, related rows.

Related data can be related in a to-one or a to-many fashion. Captions of data related in a to-many fashion link to a list view showing a filtered view of the table.

Try moving around until you find a non-empty to-many entry and click on the label to get to one. You can move back to the root by clicking on the database name in the header.

SQuiL has stopped working due to an internal error.

If you are curious you may find further information in the browser console, which is accessible through the devtools (F12).

Reload