The query example does not work about elasticsearch-learning-to-rank HOT 8 CLOSED

Thanks, I will try this out tonight

from elasticsearch-learning-to-rank.

Does it work to just run 'ltr' by itself? Does it only fail in a rescore context?

from elasticsearch-learning-to-rank.

Could you confirm the plugin is installed via GET _cat/plugins?

from elasticsearch-learning-to-rank.

Hey @beifeizhou that ... is expecting you to place your own initial query such as {"match": {"_all": "<your query here>"}}"

from elasticsearch-learning-to-rank.

Hi @softwaredoug

Thanks a lot for answering my questions. In the end, I have figured out that by using this query it is working:

GET shakespeare/_search
{
"query": {
"query_string": {
"query": "story"
}
},
"rescore": {
"window_size" : 50,
"query": {
"rescore_query": {
"ltr": {
"model": {
"stored": "dummy"
},
"features":[{
"match": {
"title": "romeo"
}
},{
"constant_score": {
"query": {
"match_phrase": {
"title": "juliet"
}
}
}
}]
}
}
}
}
}

I think the example provided from your documentation missed "rescore_query".

However, I have another question. "ltr" can pass a trained model "dummy", but how does it pass doc features by "fetures" here? In my model, I have calculated several features for each query as following which can be used by RankLib:

0 qid:10002 1:0.007477 2:0.000000 3:1.000000 4:0.000000 5:0.007470
0 qid:10002 1:0.603738 2:0.000000 3:1.000000 4:0.000000 5:0.603175
...
1 qid:10036 1:0.040161 2:0.000000 3:1.000000 4:0.000000 5:0.044177

How does "features" work here to calculate features from match title?
"features":[{
"match": {
"title": "romeo"
}
},{
"constant_score": {
"query": {
"match_phrase": {
"title": "juliet"
}
}
}
}]

Thanks,
Beifei

from elasticsearch-learning-to-rank.

Hi @beifeizhou this blog article might help clarify things, it details our full example:

http://opensourceconnections.com/blog/2017/02/14/elasticsearch-learning-to-rank/

from elasticsearch-learning-to-rank.

@softwaredoug Thanks for the link.
However, the features in the examples are not generic enough for LTR. Is it possible to add one functionality which allows to upload features for each doc from external resources(files, databases, etc)?

from elasticsearch-learning-to-rank.

However, the features in the examples are not generic enough for LTR. Is it possible to add one functionality which allows to upload features for each doc from external resources(files, databases, etc)?

Yes we certainly could use more documentation. It sounds like what you want is document-only (not query-dependent features). For example, popularity or some other factor. I would recommend adding those values as fields on the documents and using a Function Score Query, specifically field_value_factor, to retrieve their values directly to be used as features in ranking.

from elasticsearch-learning-to-rank.