日本語版はこちら5.4. Correction¶
This section describes about the following correction features:
- How it works
- How to use
- How to learn
5.4.1. How it works¶
The correction feature uses three searches to compute corrected words:
- Cooccurrence search against learned data.
- Similar search against registered words. (optional)
5.4.1.1. Cooccurrence search¶
Cooccurrence search can find registered words from user's wrong input. It uses user submit sequences that will be learned from query logs, access logs and so on.
For example, there are the following user submissions:
| query | time |
|---|---|
| serach (typo!) | 2011-08-10T22:20:50+09:00 |
| search (fixed!) | 2011-08-10T22:20:52+09:00 |
Groonga creates the following correction pair from the above submissions:
| input | corrected word |
|---|---|
| serach | search |
Groonga treats continuous submissions within a minute as input correction by user. Not submitted user input sequence between two submissions isn't used as learned data for correction.
If an user inputs "serach" and cooccurrence search returns "search" because "serach" is in input column and corresponding corrected word column value is "search".
5.4.1.2. Similar search¶
Similar search can find registered words that has one or more the same tokens as user input. TokenBigram tokenizer is used for tokenization because suggest dataset schema created by groonga-suggest-create-dataset uses TokenBigram tokenizer as the default tokenizer.
For example, there is a registered query "search engine". An user can find "search engine" by "web search service", "sound engine" and so on. Because "search engine" and "web search engine" have the same token "search" and "search engine" and "sound engine" have the same token "engine".
"search engine" is tokenized to "search" and "engine" tokens. (Groonga's TokenBigram tokenizer doesn't tokenize two characters for continuous alphabets and continuous digits for reducing search noise. TokenBigramSplitSymbolAlphaDigit tokenizer should be used to ensure tokenizing to two characters.) "web search service" is tokenized to "web", "search" and "service". "sound engine" is tokenized to "sound" and "engine".
5.4.2. How to use¶
Groonga provides suggest command to use correction. --type correct option requests corrections.
For example, here is an command to get correction results by "saerch":
Execution example:
suggest --table item_query --column kana --types correction --frequency_threshold 1 --query saerch
# [
# [
# 0,
# 1337566253.89858,
# 0.000355720520019531
# ],
# {
# "correct": [
# [
# 1
# ],
# [
# [
# "_key",
# "ShortText"
# ],
# [
# "_score",
# "Int32"
# ]
# ],
# [
# "search",
# 1
# ]
# ]
# }
# ]
5.4.3. How it learns¶
Cooccurrence search uses learned data. They are based on query logs, access logs and so on. To create learned data, groonga needs user submit inputs with time stamp.
For example, an user wants to search by "search" but the user has typo "saerch" before inputs the correct query. The user inputs the query with the following sequence:
- 2011-08-10T13:33:23+09:00: s
- 2011-08-10T13:33:23+09:00: sa
- 2011-08-10T13:33:24+09:00: sae
- 2011-08-10T13:33:24+09:00: saer
- 2011-08-10T13:33:24+09:00: saerc
- 2011-08-10T13:33:25+09:00: saerch (submit!)
- 2011-08-10T13:33:29+09:00: serch (correcting...)
- 2011-08-10T13:33:30+09:00: search (submit!)
Groonga can be learned from the input sequence by the following command:
load --table event_query --each 'suggest_preparer(_id, type, item, sequence, time, pair_query)'
[
{"sequence": "1", "time": 1312950803.86057, "item": "s"},
{"sequence": "1", "time": 1312950803.96857, "item": "sa"},
{"sequence": "1", "time": 1312950804.26057, "item": "sae"},
{"sequence": "1", "time": 1312950804.56057, "item": "saer"},
{"sequence": "1", "time": 1312950804.76057, "item": "saerc"},
{"sequence": "1", "time": 1312950805.76057, "item": "saerch", "type": "submit"},
{"sequence": "1", "time": 1312950809.76057, "item": "serch"},
{"sequence": "1", "time": 1312950810.86057, "item": "search", "type": "submit"}
]