I am creating a machine learning model that essentially returns the correctness of one text to another.
For example; “the cat and a dog”, “a dog and the cat”. The model needs to be able to identify that some words (“cat”/“dog”) are more important/significant than others (“a”/“the”). I am not interested in conjunction words etc. I would like to be able to tell the model which words are the most “significant” and have it determine how correct text 1 is to text 2, with the “significant” words bearing more weight than others.
It also needs to be able to recognise that phrases don’t necessarily have to be in the same order. The two above sentences should be an extremely high match.
What is the basic algorithm I should use to go about this? Is there an alternative to just creating a dataset with thousands of example texts and a score of correctness?
I am only after a broad overview/flowchart/process/algorithm.
I think TF-IDF might be a good fit to your problem, because:
Emphasis on words occurring in many documents (say, 90% of your sentences/documents contain the conjuction word 'and') is much smaller, essentially giving more weight to the more document specific phrasing (this is the IDF part).
Ordering in Term Frequency (TF) does not matter, as opposed to methods using sliding windows etc.
It is very lightweight when compared to representation oriented methods like the one mentioned above.
Big drawback: Your data, depending on the size of corpus, may have too many dimensions (the same number of dimensions as unique words), you could use stemming/lemmatization in order to mitigate this problem to some degree.
You may calculate similiarity between two TF-IDF vector using cosine similiarity for example.
EDIT: Woops, this question is 8 months old, sorry for the bump, maybe it will be of use to someone else though.
Hej guys,
I'm working on some ranking related research. I would like to index a collection of documents with Lucene, take the tfidf representations (of each document) it generates, alter them, put them back into place and observe how the ranking over a fixed set of queries changes accordingly.
Is there any non-hacky way to do this?
Your question is too vague to have a clear answer, esp. on what you plan to do with :
take the tfidf representations (of each document) it generates, alter them
Lucene stores raw values for scoring :
CollectionStatistics
TermStatistics
Per term/doc pair stats : PostingsEnum
Per field/doc pair : norms
All this data is managed by lucene and will be used to compute a score for a given query term. A custom Similarity class can be used to change the formula that generates this score.
But you have to consider that a search query is made of multiple terms, and the way the scores of individual terms are combined can be changed as well. You could use existing Query classes (e.g. BooleanQuery, DisjunctionMax) but you could also write your own.
So it really depends on what you want to do with of all this but note that if you want to change the raw values stored by lucene this is going to be rather hard. You'll have to write a custom lucene codec and probably most the query stack to take benefit of your new data.
One nice thing you should consider is the possibility to store an arbitrary byte[] payloads. This way you could store a value that would have been computed outside of lucene and use it in a custom similarity or query.
Please see the following tutorials: Getting Started with Payloads and Custom Scoring with Lucene Payloads it may you give some ideas.
I'm a little confused when I read this paper:Pairwise Document Similarity in Large Collections with MapReduce
http://www.umiacs.umd.edu/~jimmylin/publications/Elsayed_etal_ACL2008_short.pdf
In this paper, the author seems didn't consider word only appears in one document, but according to the definition of cosine similarity, we need to consider this situation, right?
The material I used is this: https://www.dropbox.com/s/nctb66hh84ab32c/postings-Reuters-data
The java code I used is this: https://www.dropbox.com/s/aklviixup4uulmu/CosineSimilarity.java
And the results I generated is this: https://www.dropbox.com/s/ea6ov7l7yut7yfj/part-00000
In the results, I see a lot of 1's and even number bigger than 1. I think it's kind of weird, could someone help me find out the reason? Thanks.
I have a Lucene indexed corpus of more than 1 million documents.
I am searching for named entities such as "Susan Witting" by using the the Lucene java API for queries.
I would like to expand my queries by also searching for "Sue Witting" for example but would like that term to have a lower weight than the main query term.
How can I go about doing that?
I found infos about the boosting option in the Lucene Manual. But it seems to be set at indexing and it needs fields.
You can boost each query clause independently. See the Query Javadoc.
If you want to give different weight to the words of a term. Then
Query#setBoost(float)
is not useful. A better way is:
Term term = new Term("some_key", "stand^3 firm^2 always");
This allows to give different weight to words in the same term query. Here, the word stand boosted by three but always is has the default boost value.
I have a lucene index, the documents are in around 20 different languages, and all are in the same index, I have a field 'lng' which I use to filter the results in only one language.
Based on this index I implemented spell-checker, the issue is that I get suggestions from all languages, which are irrelevant (if I am searching in English, suggestions in German are not what I need). My first idea was to create a different spell-check index for each language and than select index based on the language of the query, but I do not like this, is it possible to add additional column in spell-check index and use this, or is there some better way to do this?
Another question is how could I improve suggestions for 2 or more Terms in search query, currently I just do it for the first, which can be strongly improved to use them in combination, but I could not find any samples, or implementations which could help me solve this issue.
thanks
almir
As far as I know, it's not possible to add a 'language' field to the spellchecker index. I think that you need to define several search SpellCheckers to achieve this.
EDIT: As it turned out in the comments that the language of the query is entered by the user as well, then my answer is limited to: define multiple spellcheckers. As for the second question that you added, I think that it was discussed before, for example here.
However, even if it would be possible, it doesn't solve the biggest problem, which is the detection of query language. It is highly non-trivial task for very short messages that can include acronyms, proper nouns and slang terms. Simple n-gram based methods can be inaccurate (as e.g. the language detector from Tika). So I think that the most challenging part is how to use certainty scores from both language detector and spellchecker and what threshold should be chosen to provide meaningful corrections (e.g. language detector prefers German, but spellchecker has a good match in Danish...).
If you look at the source of SpellChecker.SuggestSimilar you can see:
BooleanQuery query = new BooleanQuery();
String[] grams;
String key;
for (int ng = GetMin(lengthWord); ng <= GetMax(lengthWord); ng++)
{
<...>
if (bStart > 0)
{
Add(query, "start" + ng, grams[0], bStart); // matches start of word
}
<...>
I.E. the suggestion search is just a bunch of OR'd boolean queries. You can certainly modify this code here with something like:
query.Add(new BooleanClause(new TermQuery(new Term("Language", "German")),
BooleanClause.Occur.MUST));
which will only look for suggestions in German. There is no way to do this without modifying your code though, apart from having multiple spellcheckers.
To deal with multiple terms, use QueryTermExtractor to get an array of your terms. Do spellcheck for each, and cartesian join. You may want to run a query on each combo and then sort based on the frequency they occur (like how the single-word spellchecker works).
After implement two different search features in two different sites with both lucene and sphinx, I can say that sphinx is the clear winner.
Consider using http://sphinxsearch.com/ instead of lucene. It's used by craigslist, among others.
They have a feature called morphology preprocessors:
# a list of morphology preprocessors to apply
# optional, default is empty
#
# builtin preprocessors are 'none', 'stem_en', 'stem_ru', 'stem_enru',
# 'soundex', and 'metaphone'; additional preprocessors available from
# libstemmer are 'libstemmer_XXX', where XXX is algorithm code
# (see libstemmer_c/libstemmer/modules.txt)
#
# morphology = stem_en, stem_ru, soundex
# morphology = libstemmer_german
# morphology = libstemmer_sv
morphology = none
There are many stemmers available, and as you can see, german is among them.
UPDATE:
Elaboration on why I feel that sphinx has been the clear winner for me.
Speed: Sphinx is stupid fast. Both indexing and in the serving search queries.
Relevance: Though it's hard to quantify this, I felt that I was able to get more relevant results with sphinx compared to my lucene implementation.
Dependence on the filesystem: With lucene, I was unable to break the dependence on the filesystem. And while their are workarounds, like creating a ram disk, I felt it was easier to just select the "run only in memory" option of sphinx. This has implications for websites with more than one webserver, adding dynamic data to the index, reindexing, etc.
Yes, these are just points of an opinion. However, they are an opinion from someone that has tried both systems.
Hope that helps...