在Vowpal Wabbit中使用LDA的输出格式

Question

I used VowpalWabbit.LDA to generate topics for some document collection. Output file looks like:

Version 7.7.0

Min label:0.000000

Max label:1.000000

bits:18

0 pairs: 

0 triples: 

rank:0

lda:10

0 ngram: 

0 skip: 

options: --lda 10

0 21407.330078 1.025429 0.648226 0.917246 0.451278 0.801456 11463.415039 0.876181 1.105704 0.785956 

1 39210.687500 0.814911 0.389153 0.473620 0.391765 0.688513 0.708061 0.526936 0.719026 0.659338 

2 41573.523438 1.161345 0.583391 0.918144 0.318337 0.543920 0.704812 0.987455 0.633980 0.890918 

3 2.759077 1.114242 0.662993 1.113668 0.632519 0.707388 26730.898438 1.074518 0.974116 0.909262 

4 1.476383 1.263869 0.552380 0.838780 0.500615 0.529077 24156.128906 0.689529 1.400310 0.530180 

5 1.083310 0.746087 0.539263 1.152820 0.496213 0.726304 17391.972656 0.809698 1.682978 0.925061 

6 4.601943 1.551102 0.541617 1.532858 0.418091 1.432069 10.024081 1.992290 12924.787109 1.202141 

I supposed to see identifier of each word and the probabilities of belonging it to each topic. But I see some huge numbers like 21407.330078. Does anybody know how to transform this output format to the format I want to see?

Answer 1

It seems you are looking into predictions output file. It contains "the inferred per-document topic weights" with following format: "Each line corresponds to a document d. Each column corresponds to a topic k".

If you need information about words you shall add "--readable_model topics.dat" parameter to the command line. This will give you the topics in human-readable format with following content: "Each line corresponds to a topic k. Each column corresponds to a word w" Please refer to https://github.com/JohnLangford/vowpal_wabbit/wiki/lda.pdf

Answer 2

If you run vw with -a (audit) you can see the mappings from words to hash values. I assume you need to normalize each line in the model output and then find the top words for each topic.