How to interpret user search query (in Elasticsearch) - amazon-web-services

Question

I would like to serve my visitors the best results possible when they use our search feature.
To achieve this I would like to interpret the search query.
For example a user searches for 'red beds for kids 120cm'
I would like to interpret it as following:
Category-Filter is "beds" AND "children"
Color-filter is red
Size-filter is 120cm
Are there ready to go tools for Elasticsearch?
Will I need NLP in front of Elasticsearch?

Answer 1

Elasticsearch is pretty powerful on its own and is very much capable of returning the most relevant results to full-text search queries, provided that data is indexed and queried adequately.
Under the hood it always performs text analysis for full-text searches (for fields of type text). A text analyzer consists of a character filter, tokenizer and a token filter.
For instance, synonym token filter can replace kids with children in the user query.
Above that search queries on modern websites are often facilitated via category selectors in the UI, which can easily be implemented with querying keyword fields of Elasticsearch.
It might be enough to model your data correctly and tune its indexing to implement the search you need - and if that is not enough, you can always add some extra layer of NLP-like logic on the client side, like #2ps suggested.
Now let me show a toy example of what you can achieve with a synonym token filter and copy_to feature.
Let's define the mapping
Let's pretend that our products are characterized by the following properties: Category, Color, and Size.LengthCM.
The mapping will look something like:
PUT /my_index
{
"mappings": {
"properties": {
"Category": {
"type": "keyword",
"copy_to": "DescriptionAuto"
},
"Color": {
"type": "keyword",
"copy_to": "DescriptionAuto"
},
"Size": {
"properties": {
"LengthCM": {
"type": "integer",
"copy_to": "DescriptionAuto"
}
}
},
"DescriptionAuto": {
"type": "text",
"analyzer": "MySynonymAnalyzer"
}
}
},
"settings": {
"index": {
"analysis": {
"analyzer": {
"MySynonymAnalyzer": {
"tokenizer": "standard",
"filter": [
"MySynonymFilter"
]
}
},
"filter": {
"MySynonymFilter": {
"type": "synonym",
"lenient": true,
"synonyms": [
"kid, kids => children"
]
}
}
}
}
}
}
Notice that we selected type keyword for the fields Category and Color.
Now, what about these copy_to and synonym?
What will copy_to do?
Every time we send an object for indexing into our index, value of the keyword field Category will be copied to a full-text field DescritpionAuto. This is what copy_to does.
What will synonym do?
To enable synonym we need to define a custom analyzer, see MySynonymAnalyzer which we defined under "settings" above.
Roughly, it will replace every token that matches something on the left of => with the token on the right.
How will the documents look like?
Let's insert a few example documents:
POST /my_index/_doc
{
"Category": [
"beds",
"adult"
],
"Color": "red",
"Size": {
"LengthCM": 150
}
}
POST /my_index/_doc
{
"Category": [
"beds",
"children"
],
"Color": "red",
"Size": {
"LengthCM": 120
}
}
POST /my_index/_doc
{
"Category": [
"couches",
"adult",
"family"
],
"Color": "blue",
"Size": {
"LengthCM": 200
}
}
POST /my_index/_doc
{
"Category": [
"couches",
"adult",
"family"
],
"Color": "red",
"Size": {
"LengthCM": 200
}
}
As you can see, DescriptionAuto is not present in the original documents - though due to copy_to we will be able to query it.
Let's see how.
Performing the search!
Now we can try out our index with a simple query_string query:
POST /my_index/_doc/_search
{
"query": {
"query_string": {
"query": "red beds for kids 120cm",
"default_field": "DescriptionAuto"
}
}
}
The results will look something like the following:
"hits": {
...
"max_score": 2.3611186,
"hits": [
{
...
"_score": 2.3611186,
"_source": {
"Category": [
"beds",
"children"
],
"Color": "red",
"Size": {
"LengthCM": 120
}
}
},
{
...
"_score": 1.0998137,
"_source": {
"Category": [
"beds",
"adult"
],
"Color": "red",
"Size": {
"LengthCM": 150
}
}
},
{
...
"_score": 0.34116736,
"_source": {
"Category": [
"couches",
"adult",
"family"
],
"Color": "red",
"Size": {
"LengthCM": 200
}
}
}
]
}
The document with categories beds and children and color red is on top. And its relevance score is twice bigger than of its follow-up!
How can I check how Elasticsearch interpreted the user's query?
It is easy to do via analyze API:
POST /my_index/_analyze
{
"text": "red bed for kids 120cm",
"analyzer": "MySynonymAnalyzer"
}
{
"tokens": [
{
"token": "red",
"start_offset": 0,
"end_offset": 3,
"type": "<ALPHANUM>",
"position": 0
},
{
"token": "bed",
"start_offset": 4,
"end_offset": 7,
"type": "<ALPHANUM>",
"position": 1
},
{
"token": "for",
"start_offset": 8,
"end_offset": 11,
"type": "<ALPHANUM>",
"position": 2
},
{
"token": "children",
"start_offset": 12,
"end_offset": 16,
"type": "SYNONYM",
"position": 3
},
{
"token": "120cm",
"start_offset": 17,
"end_offset": 22,
"type": "<ALPHANUM>",
"position": 4
}
]
}
As you can see, there is no token kids, but there is token children.
On a side note, in this example Elasticsearch wasn't able, though, to parse the size of the bed: token 120cm didn't match to anything, since all sizes are integers, like 120, 150, etc. Another layer of tweaking will be needed to extract 120 from 120cm token.
I hope this gives an idea of what can be achieved with Elasticsearch's built-in text analysis capabilities!