How to add "context" to Elastic Search suggestions - elasticsearch

Question

I'm building a Enterprise social network.
I want to suggest people to add as friend, based on their title.
For example, the value can be: developer, blogger, singer, barber, bartender ...
My users are saved into ElasticSearch, their titles are saved in the field 'title'.
The current mapping is:
title: {
type: 'text',
analyzer: 'autocomplete_analyzer',
search_analyzer: 'autocomplete_analyzer_search'
}
and the query is:
should: [
{
match: {
title: {
query: user.title,
minimum_should_match: '90%',
boost: 2
}
}
}
]
and the analyzers definitions are:
indexConfig: {
settings: {
analysis: {
analyzer: {
autocomplete_analyzer: {
tokenizer: 'autocomplete_tokenizer',
filter: ['lowercase', 'asciifolding']
},
autocomplete_analyzer_search: {
tokenizer: 'lowercase',
filter: ['asciifolding']
},
phrase_analyzer: {
tokenizer: 'standard',
filter: ['lowercase', 'asciifolding', 'fr_stop', 'fr_stemmer', 'en_stop', 'en_stemmer']
},
derivative_analyzer: {
tokenizer: 'standard',
filter: ['lowercase', 'asciifolding', 'derivative_filter', 'fr_stop', 'fr_stemmer', 'en_stop', 'en_stemmer']
}
},
tokenizer: {
autocomplete_tokenizer: {
type: 'edge_ngram',
min_gram: 2,
max_gram: 20,
token_chars: ['letter', 'digit']
}
},
filter: {
derivative_filter: {
type: 'word_delimiter',
generate_word_parts: true,
catenate_words: true,
catenate_numbers: true,
catenate_all: true,
split_on_case_change: true,
preserve_original: true,
split_on_numerics: true,
stem_english_possessive: true
},
en_stop: {
type: 'stop',
stopwords: '_english_'
},
en_stemmer: {
type: 'stemmer',
language: 'light_english'
},
fr_stop: {
type: 'stop',
stopwords: '_french_'
},
fr_stemmer: {
type: 'stemmer',
language: 'light_french'
}
}
}
}
}
I tested it, the relevance is very good, but they are not enough users matched by this, because of the '90%' criteria.
A quick and dirty solution is to lower this criteria to 50% of course.
However, If I do that, I suppose that Elastic will search titles based on the concordance of the letters in the title, rather that the relevance of the proximity between titles.
For example, If my user is a 'barber', ElasticSearch might suggest 'bartender', because they have in common: b,a,r,e,r
Hence, I have two questions:
1 - is my assumption correct ?
2 - what can I do to add more relevance on my titles search ?

Answer 1

The problem with your search is following - it uses autocomplete_analyzer, which is basically creates a huge index with a lot of n-grams.
Example for bartender would be something like ba, bar, bart, etc.
As you could see, for barber you will have a bit similar n-grams, which would make a match.
Regarding your questions, if you would lower the minimum_should_match you will get more results, but that's just because the following matching procedure will lead to partial matches.
To increase the relevancy - I would recommend to use another analyzer, since this n-gram analyzer is usually suitable only for autosuggest functionality, which isn't the case. There could be several choices from keeping it simple to keyword analyzer, or whitespace one.
What would be more important is to properly construct the query. For example if user searches for partial title, e.g bar, you may use prefix query. However, if you're searching just by full match (e.g. developer or bartender) it would be more important to just normalize title field properly. E.g. to use lowercase analyzer with some stemming.