1. What is hyphen analysis?
Usually, you index only words into Elasticsearch, and the indexing does not care about other symbols like delimiters, punctuation, hyphens, apostrophs, or other caracters which can be found within words or at word boundaries.
But that does not always work out right. Entity names (like organization names, person names, or book titles) may carry symbols as part of the identifying function that makes the entity name unique, but are often ignored or handled sloppily when search terms are being entered for that names.
Or in german language, there are many Bindestrichwörter which consist of word parts that are connected by a hyphen symbol. In many cases, sloppy searchers are not using in-word hyphens correctly and will therefore not get correct search results.
For examples in german, see https://de.wikipedia.org/wiki/Viertelgeviertstrich
For example, indexing entity names like
U.S.A., O’Grady, Corinna’s Cause or Programming with C++
are a challenge when being searched for.
Often it is preferable to have successful searches for related terms like usa, ogrady, corinnas cause.
On the other hand, you want to avoid false hits, when searching for Programming with C.
So, C++ must not be indexed as C.
To achieve that, we use hyphen tokenizing together with a special character-based hyphen symbol detection that
allows for indexing multiples forms of the same word in the token chain.
Corinna’s will be indexed as Corinna, Corinnas, and Corinna’s to generate hits when searching for that forms.
Entity names like "C++" or "AT&T" can be protected by the keyword_marker filter available in Elasticsearch.
That means, they are preserved throughout the process.
2. Lucene standard tokenization follows Unicode tokenization rules
Lucene 4+ is using a new tokenzation by default and switched from a grammar-based tokenzation (now called the classic tokenization) to Unicode-based tokenization (also known as UAX#29).
Lucene default tokenization before Lucene 4 was focusing on european languages only. The "classic" tokenizer does not work well with asian languages for example.
The challenge is now, while Lucene is now adhering to international standards, which is a good thing, it does no longer treat hyphens as word part delimiters and ignores them:
The correct interpretation of hyphens in the context of word boundaries is challenging. It is quite common for separate words to be connected with a hyphen: “out-of-the-box,” “under-the-table,” “Italian-American,” and so on. A significant number are hyphenated names, such as “Smith-Hawkins.” When doing a Whole Word Search or query, users expect to find the word within those hyphens. While there are some cases where they are separate words (usually to resolve some ambiguity such as “re-sort” as opposed to “resort”), it is better overall to keep the hyphen out of the default definition. (http://unicode.org/reports/tr29/)
The solution this plugin offers is an improved implementation of the old classic tokenizer by a rewritten JFlex grammar that takes more composed word fragments together. A tricky aspect is to detect superfluous adjunct characters but keep them in acronyms like "U.S.A." so the character will be dropped correctly before indexing.
The hyphen tokenizer uses the Unicode "Pd" (Punctuation Dash Category) to detect punctuation in words.
The price this solution pays is that it does not conform to Unicode tokenization rules. So it is recommended to use hyphen tokenization on european language fields only.
3. Hyphen tokenizer example for Elasticsearch 5.x
In this example, it is demonstrated how the token "E-Book" is indexed. It generates tokens so that "E-Book", "EBook", and "Book" will match.
While the hyphen tokenizer cares about the comma and suppresses the character, the hyphen token filter cares about creating "EBook" and "Book" tokens.
PUT /test
{
"settings": {
"index": {
"analysis": {
"analyzer": {
"my_analyzer": {
"type": "custom",
"tokenizer": "hyphen",
"filter" : [ "hyphen", "lowercase" ]
}
}
}
}
},
"mappings": {
"docs": {
"properties": {
"text": {
"type": "text",
"analyzer": "my_analyzer"
}
}
}
}
}
GET /test/_settings
GET /test/_mapping
PUT /test/docs/1
{
"text" : "Read this E-Book, or you will miss the best"
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "E-Book",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "EBook",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "ebook",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "book",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}4. The hyphen token filter
The hyphen token filter can be used in conjunction with any tokenizer. It focuses on a single word and examines them for generating one or more tokens. Connected word fragments are detected by using the Unicode "L" (Letter Category). The token filter does not use the improved JFlex grammar technique.
Here is an example to demonstrate the hyphen analyzer at work. The whitespace tokenizer is used here but
it does not guarantee to remove adjunct punctuation. Therefore, in realw rold data, you should always use the
hyphen tokenizer accompanying the hyphen token filter.
4.1. Hyphen token filter example for Elasticsearch 5.x
PUT /test
{
"settings": {
"index": {
"analysis": {
"filter": {
"hyphen": {
"type": "hyphen",
"hyphens": "+-'",
"respect_keywords": true
},
"marker": {
"type": "keyword_marker",
"keywords": [
"C++",
"AT&T"
]
}
},
"analyzer": {
"my_analyzer": {
"type": "custom",
"tokenizer": "whitespace",
"filter": [
"marker",
"hyphen"
]
}
}
}
}
},
"mappings": {
"docs": {
"properties": {
"text": {
"type": "text",
"analyzer": "my_analyzer"
}
}
}
}
}
GET /test/_settings
GET /test/_mapping
PUT /test/docs/1
{
"text" : "Corinna's Cause"
}
PUT /test/docs/2
{
"text" : "U+002B"
}
PUT /test/docs/3
{
"text" : "Programming C++"
}
PUT /test/docs/4
{
"text" : "Build a career with AT&T"
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "Corinna Cause",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "Corinnas Cause",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "Corinna's Cause",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "002B",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "U\\+002B",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "Programming C\\+\\+",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "Programming C",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "Build a career with AT&T",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "Build a career with ATT",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}5. The Hyphen Analyzer
For convenience, this plugin provides a hyphen analyzer which is a custom analyzer with a hyphen tokenizer.
PUT /test
{
"settings": {
"index": {
"analysis": {
"analyzer": {
"my_analyzer": {
"type": "hyphen"
}
}
}
}
},
"mappings": {
"docs": {
"properties": {
"text": {
"type": "text",
"analyzer": "my_analyzer"
}
}
}
}
}
GET /test/_settings
GET /test/_mapping
PUT /test/docs/1
{
"text" : "Read this E-Book, or you will miss the best"
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "E-Book",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "EBook",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}
POST /test/docs/_search
{
"query": {
"simple_query_string": {
"query" : "Book",
"fields" : [ "text" ],
"default_operator": "and"
}
}
}6. Options
These options can be used for the hyphen tokenizer.
| max_token_length |
maximum length of a single token that will be indexed. Default is 255 (StandardAnalyzer.DEFAULT_MAX_TOKEN_LENGTH) |
These options can be used for the hyphen token filter.
| hyphens |
a string containing characters that should be used for detection. Default is |
| subwords |
if subwords should be generated as tokens. Default is |
| respect_keywords |
if |
7. Gradle test report
The current test report is here