Phrase match error in chinese

[gemini133]

documents. default settings.

curl -s http://localhost:7700/indexes/products/documents | rq
[
  {
    "id": "123",
    "title": "小化妆包"
  },
  {
    "id": "456",
    "title": "Ipad 包"
  }
]

search with phrase match "化妆". no result returned.

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"化妆\"" }' | rq

{
  "exhaustiveNbHits": false,
  "hits": [],
  "limit": 20,
  "nbHits": 0,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "\"化妆\""
}

search with phrase match "小化妆". no result returned.

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"小化妆\"" }' | rq
{
  "exhaustiveNbHits": false,
  "hits": [],
  "limit": 20,
  "nbHits": 0,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "\"小化妆\""
}

search with phrase match "化妆包". result returned.

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"化妆包\"" }' | rq
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "123",
      "title": "小化妆包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "\"化妆包\""
}

search with phrase match "小化妆包". result returned.

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"小化妆包\"" }' | rq
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "123",
      "title": "小化妆包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "\"小化妆包\""
}

Expected behavior
queries: "化妆", "小化妆", "化妆包", "小化妆包" all of them should return same result.

 {
      "id": "123",
      "title": "小化妆包"
    }

MeiliSearch version:

meilisearch-http 0.22.0

[gemini133]

I did a quick test with an english document: {"title": "adult face mask"} and utilize phrase match: "adult", "adult face", "face mask", document is returned as expected

[YikSanChan]

@gemini133 I guess this is because meilisearch uses jieba to tokenize the search word "小化妆包" and it is understood as "化妆包".

See https://docs.meilisearch.com/reference/under_the_hood/tokenization.html

[gemini133]

yep, I think it's related to the tokenizer too.
why does "化妆" returns nothing then?
not sure how to work around this. confusing...

[gemini133]

another test with just "包"

{
    "q": "\"包\""
}

{
    "hits": [
        {
            "id": "456",
            "title": "Ipad 包"
        }
    ],
    "nbHits": 1,
    "exhaustiveNbHits": false,
    "query": "\"包\"",
    "limit": 20,
    "offset": 0,
    "processingTimeMs": 0
}

[ManyTheFish]

Hello @gemini133, it seems that you are using phrase search in your queries putting the " around your word.
If you remove " do you have the expected behavior?

[gemini133]

@ManyTheFish yes, I'm using phrase/exact search because the search results with normal search is sort of irrelevant to me
so I tried phrase search instead.

Please check screenshot

[ManyTheFish]

@gemini133, I can't speak any Chinese language, sorry if I don't understand the complete problem.

But, if I'm right, only the second request gets an irrelevant response, because Ipad 包 should never be returned for 化妆包.

Am I right?

[gemini133]

sorry, and you are right. Ipad 包 should never be returned for 化妆包. is it related to ranking rules? I've tweaked it a couple times but still not what I want. So, I then go to phrase search

[ManyTheFish]

I will investigate,
thank you @gemini133 for your report. 👍

[curquiza]

Closed by #1711 containing milli v0.17.0

[gemini133]

here is the feedback of 0.23. Build it from the master branch

**Phrase search**

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"小\"" }' | rq
//correct
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "123",
      "title": "小化妆包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "\"小\""
}



curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"小化妆\"" }' | rq
// wrong, must return 小化妆包
{
  "exhaustiveNbHits": false,
  "hits": [],
  "limit": 20,
  "nbHits": 0,
  "offset": 0,
  "processingTimeMs": 1,
  "query": "\"小化妆\"" 
}

---- 

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"化妆包\"" }' | rq
// correct
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "123",
      "title": "小化妆包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "\"化妆包\""
}

----

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"小化妆包\"" }' | rq
// correct
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "123",
      "title": "小化妆包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "\"小化妆包\""
}

----

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "\"包\"" }' | rq
// wrong, must contain 小化妆包 too
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "456",
      "title": "Ipad 包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 1,
  "query": "\"包\""
}

---

**Normal Search**

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "包" }' | rq
// wrong, must contain 小化妆包 too
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "456",
      "title": "Ipad 包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "包"
}

----

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "化妆包" }' | rq
// not sure, if jieba splits into "化妆" 和 "包", then it should contain "Ipad 包"
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "123",
      "title": "小化妆包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 1,
  "query": "化妆包"
}

----


curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "化妆" }' | rq
// correct
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "123",
      "title": "小化妆包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "化妆"
}

----

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "小" }' | rq
// correct
{
  "exhaustiveNbHits": false,
  "hits": [
    {
      "id": "123",
      "title": "小化妆包"
    }
  ],
  "limit": 20,
  "nbHits": 1,
  "offset": 0,
  "processingTimeMs": 0,
  "query": "小"
}

[gemini133]

@curquiza @ManyTheFish

[ManyTheFish]

Hello @gemini133!
In your last comment, we have different reasons to "why Meilisearch doesn't return the good response?":

Tokenization & Phrase search

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": ""小化妆"" }' | rq
// wrong, must return 小化妆包
{
"exhaustiveNbHits": false,
"hits": [],
"limit": 20,
"nbHits": 0,
"offset": 0,
"processingTimeMs": 1,
"query": ""小化妆""
}

In Meilisearch we can only store 1 variation of a tokenized text, and this variation is given by Jieba.
here the text 小化妆包 is tokenized in two "words" as ["小", "化妆包"].
The search query is tokenized in two "words" as ["小", "化妆"], and because it's a phrase search, we enforce an exact match on the two words:

• "小" == "小" ---------> ✅
• "化妆" =/= "化妆包" ---> ❌

This does not match.

Tokenization & suffix search

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": ""包"" }' | rq
OR curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "包" }' | rq
// wrong, must contain 小化妆包 too
{
"exhaustiveNbHits": false,
"hits": [
{
"id": "456",
"title": "Ipad 包"
}
],
"limit": 20,
"nbHits": 1,
"offset": 0,
"processingTimeMs": 1,
"query": ""包""
}

In Meilisearch we can only store 1 variation of a tokenized text, and this variation is given by Jieba.
here the text 小化妆包 is tokenized in two "words" as ["小", "化妆包"] (we technically can't store more variation for now).
Moreover, we don't support suffix search, we only support prefix search. This means that the word 化妆包 is never a match of the word 包.

Tokenization & query Words

curl -s -X POST 'http://localhost:7700/indexes/products/search' --data '{ "q": "化妆包" }' | rq
// not sure, if jieba splits into "化妆" 和 "包", then it should contain "Ipad 包"
{
"exhaustiveNbHits": false,
"hits": [
{
"id": "123",
"title": "小化妆包"
}
],
"limit": 20,
"nbHits": 1,
"offset": 0,
"processingTimeMs": 1,
"query": "化妆包"
}

Meilisearch tokenizes 化妆包 as ["化妆包"] so 包 can't be a match for 化妆包.
Moreover, if Meilisearch would tokenize 化妆包 as ["化妆", "包"], we couldn't match Ipad 包 because the document doesn't contain the word "化妆".

I hope my explanations were clear. Thanks a lot @gemini133 for your detailed report! 👍

[gemini133]

Thanks for your detailed explanation! It's very helpful.
Honestly, I'm sort of disappointed as compared to the past experience with elasticsearch.

[ManyTheFish]

I'm sorry for that.
But I have some questions.

is it really relevant to respond Ipad 包 to the query 化妆包?

I used a traduction tool, and 化妆包 seems to be a Cosmetic bag. I don't personally find that an Ipad bag is a good response to a Cosmetic bag. 🤔
Is there another traduction to 化妆包 that makes Ipad 包 relevant?

What is the most disappointing thing in the responses given by Meilisearch?

Is there one of your examples that makes meilisearch irrelevant for your usage?

[gemini133]

no, no, no. Ipad 包 is not relevant to 化妆包, but when you search for "包", they both should be returned, right?

from the view of a normal user, when I search for "包", the search results should contains all products entitled with "包" , in this case, "ipad 包" 和 "化妆包".

[ManyTheFish]

Yes, we could tune Jieba to tokenize better for our users.
Unfortunately, I can't read/write any Chinese script, so if you know a bit about Jieba and how we could tune it to enhance your experience with meilisearch, I would gladly make the changes! 🙂

[gemini133]

@ManyTheFish hey man, really appreciate your help. I'll look into jieba and see what I can do for all of us

[gemini133]

@ManyTheFish unfortunately, didn't find a way to tweak jieba except for load_dict loading our own dictionary.

messense/jieba-rs#77