Elasticsearch 新近发布的 7.6 版本里面包含了很多激动人心的功能，而最让我感兴趣的是利用机器学习来自动检测语言的功能。
检测文本语言本身不是什么稀奇事，之前做爬虫的时候，就做过对网页正文进行语言的检测，有很多成熟的方案，而最好的就属 Google Chrome 团队开源的 CLD 系列（全名：Compact Language Detector）了，能够检测多达 80 种各种语言，我用过CLD2，是基于 C++ 贝叶斯分类器实现的，而 CLD3 则是基于神经网络实现的，无疑更加准确，这次 Elasticsearch 将这个非常小的功能也直接集成到了默认的发行包里面，对于使用者来说可以说是带来很大的方便。
相信很多朋友，在实际的业务场景中，对碰到过一个字段同时存在多个语种的文本内容的情况，尤其是出海的产品，比如类似大众点评的 APP 吧，一个餐馆下面，来自七洲五湖四海的朋友都来品尝过了，自然要留下点评语是不，德国的朋友使用的是德语，法国的朋友使用的是法语，广州的朋友用的是粤语，那对于开发这个 APP 的后台工程师可就犯难了，如果这些评论都存在一个字段里面，就不好设置一个统一的分词器，因为不同的语言他们的切分规则肯定是不一样的，最简单的例子，比如中文和英文，设置的分词不对，查询结果就会不精准。
相信也有很多人用过这样的解决方案，既然一个字段搞不定，那就把这个字段复制几份，英文字段存一份，中文字段存一份，用 MultiField 来做，这样虽然可以解决一部分问题，但是同样会带来容量和资源的浪费，和查询时候具体该选择哪个字段来参与查询的挑战。
而利用 7.6 的这个新功能，可以在创建索引的时候，可以自动的根据内容进行推理，从而影响索引文档的构成，进而做到特定的文本进特定的字段，从而提升查询体验和性能，关于这个功能，Elastic 官网这里也有一篇博客2，提供了详细的例子。
看上去不错，但是鲁迅说过，网上得来终觉浅，觉知此事要躬行，来， 今天一起跑一遍看看具体怎么个用法。
首先，这个功能叫 Language identification，是机器学习的一个 Feature，但是不能单独使用，要结合 Ingest Node 的一个 inference ingest processor 来使用，Ingest processor 是在 Elasticsearch 里面运行的数据预处理器，部分功能类似于 Logstash 的数据解析，对于简单数据操作场景，完全可以替代掉 Logstash，简化部署架构。
Elasticsearch 在 7.6 的包里面，默认打包了提前训练好的机器学习模型，就是 Language identification 需要调用的语言检测模型，名称是固定的 lang_ident_model_1，这也是 Elasticsearch 自带的第一个模型包，大家了解一下就好。
那这个模型包在什么位置呢，我们来解刨一下：
$unzip /usr/share/elasticsearch/modules/x-pack-ml/x-pack-ml-7.6.0.jar$/org/elasticsearch/xpack/ml/inference$ tree.|-- ingest| |-- InferenceProcessor$Factory.class| `-- InferenceProcessor.class|-- loadingservice| |-- LocalModel$1.class| |-- LocalModel.class| |-- Model.class| `-- ModelLoadingService.class`-- persistence|-- InferenceInternalIndex.class|-- TrainedModelDefinitionDoc$1.class|-- TrainedModelDefinitionDoc$Builder.class|-- TrainedModelDefinitionDoc.class|-- TrainedModelProvider.class`-- lang_ident_model_1.json3 directories, 12 files
可以看到，在 persistence 目录就有这个模型包，是 json 格式的，里面有个压缩的二进制编码后的字段。
我们还可以通过新的 API 来获取这个模型信息，以后模型多了之后会比较有用：
GET _ml/inference/lang_ident_model_1{"count" : 1,"trained_model_configs" : [{"model_id" : "lang_ident_model_1","created_by" : "_xpack","version" : "7.6.0","description" : "Model used for identifying language from arbitrary input text.","create_time" : 1575548914594,"tags" : ["lang_ident","prepackaged"],"input" : {"field_names" : ["text"]},"estimated_heap_memory_usage_bytes" : 1053992,"estimated_operations" : 39629,"license_level" : "basic"}]}
好了，基本的了解就到这里了，我们开始动手吧，既然要和 Ingest 结合使用，自然免不了要定义 Ingest Pipeline，也就是说定一个解析规则，索引的时候会调用这个规则来处理输入的索引文档。Ingest Pipeline 的调试是个问题，好在Ingest 提供了模拟调用的方法，我们测试一下：
POST _ingest/pipeline/_simulate{"pipeline":{"processors":[{"inference":{"model_id":"lang_ident_model_1","inference_config":{"classification":{"num_top_classes":5}},"field_mappings":{}}}]},"docs":[{"_source":{"text":"新冠病毒让你在家好好带着，你服不服"}}]}
上面是借助 Ingest 的推理 Process 来模拟调用这个机器学习模型进行文本判断的方法，第一部分是设置 processor 的定义，设置了一个 inference processor，也就是要进行语言模型的检测，第二部分 docs 则是输入了一个 json 文档，作为测试的输入源，运行结果如下：
{"docs" : [{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"text" : "新冠病毒让你在家好好带着，你服不服","ml" : {"inference" : {"top_classes" : [{"class_name" : "zh","class_probability" : 0.9999872511022145,"class_score" : 0.9999872511022145},{"class_name" : "ja","class_probability" : 1.061491174235718E-5,"class_score" : 1.061491174235718E-5},{"class_name" : "hy","class_probability" : 6.304673023324264E-7,"class_score" : 6.304673023324264E-7},{"class_name" : "ta","class_probability" : 4.1374037676410867E-7,"class_score" : 4.1374037676410867E-7},{"class_name" : "te","class_probability" : 2.0709260170937159E-7,"class_score" : 2.0709260170937159E-7}],"predicted_value" : "zh","model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T15:58:44.783736Z"}}}]}
可以看到，第一条返回结果，zh 表示中文语言类型，可能性为 0.9999872511022145，基本上无限接近肯定了，这个是中文文本，而第二位和剩下的就明显得分比较低了，如果你看到是他们的得分开头是 1.x 和 6.x 等，是不是觉得，不对啊，后面的得分怎么反而大一些，哈哈，你仔细看会发现它后面其实还有 -E 啥的尾巴呢，这个是科学计数法，其实数值远远小于 0。
简单模拟倒是证明这个功能 work 了，那具体怎么使用，一起看看吧。
首先创建一个 Pipeline：
PUT _ingest/pipeline/lang_detect_add_tag{"description": "检测文本，添加语种标签","processors": [{"inference": {"model_id": "lang_ident_model_1","inference_config": {"classification": {"num_top_classes": 2}},"field_mappings": {"contents": "text"}}},{"set": {"field": "tag","value": "{{ml.inference.predicted_value}}"}}]}
可以看到，我们定义了一个 ID 为 lang_detect_add_tag 的 Ingest Pipeline，并且我们设置了这个推理模型的参数，只返回 2 个分类结果，和设置了 content 字段作为检测对象。同时，我们还定义了一个新的 set processor，这个的意思是设置一个名为 tag 的字段，它的值是来自于一个其它的字段的变量引用，也就是把检测到的文本对应的语种存成一个标签字段。
这个 Pipeline 创建完之后，我们同样可以对这个 Pipeline 进行模拟测试，模拟的好处是不会实际创建索引，方便调试。
POST /_ingest/pipeline/lang_detect_add_tag/_simulate{"docs": [{"_index": "index","_id": "id","_source": {"contents": "巴林境内新型冠状病毒肺炎确诊病例累计达56例"}},{"_index": "index","_id": "id","_source": {"contents": "Watch live: WHO gives a coronavirus update as global cases top 100,000"}}]}
返回结果：
{"docs" : [{"doc" : {"_index" : "index","_type" : "_doc","_id" : "id","_source" : {"tag" : "zh","contents" : "巴林境内新型冠状病毒肺炎确诊病例累计达56例","ml" : {"inference" : {"top_classes" : [{"class_name" : "zh","class_probability" : 0.999812378112116,"class_score" : 0.999812378112116},{"class_name" : "ja","class_probability" : 1.8175264877915687E-4,"class_score" : 1.8175264877915687E-4}],"predicted_value" : "zh","model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:21:26.981249Z"}}},{"doc" : {"_index" : "index","_type" : "_doc","_id" : "id","_source" : {"tag" : "en","contents" : "Watch live: WHO gives a coronavirus update as global cases top 100,000","ml" : {"inference" : {"top_classes" : [{"class_name" : "en","class_probability" : 0.9896669173070857,"class_score" : 0.9896669173070857},{"class_name" : "tg","class_probability" : 0.0033122788575614993,"class_score" : 0.0033122788575614993}],"predicted_value" : "en","model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:21:26.981261Z"}}}]}
可以看到，两个文档分别都正确识别了语种，并且创建了对应的 tag 字段，不过这个时候，文档里面的 ml 对象字段，就显得有点多余了，可以使用 remove processor 来删除这个字段。
PUT _ingest/pipeline/lang_detect_add_tag{"description": "检测文本，添加语种标签","processors": [{"inference": {"model_id": "lang_ident_model_1","inference_config": {"classification": {"num_top_classes": 2}},"field_mappings": {"contents": "text"}}},{"set": {"field": "tag","value": "{{ml.inference.predicted_value}}"}},{"remove": {"field": "ml"}}]}
那索引的时候，怎么使用这个 Pipeline 呢，看下面的例子：
POST news/_doc/1?pipeline=lang_detect_add_tag{"contents":"""On Friday, he added: "In a globalised world, the only option is to stand together. All countries should really make sure that we stand together." Meanwhile, Italy—the country worst affected in Europe—reported 41 new COVID-19 deaths in just 24 hours. The country's civil protection agency said on Thursday evening that 3,858 people had been infected and 148 had died."""}GET news/_doc/1
上面的这个例子就不贴返回值了，大家自己试试。
那回到最开始的场景，如果要根据检测结果来分别存储文本到不同的字段，怎么做呢，这里贴一下官网博客的例子：
POST _ingest/pipeline/_simulate{"pipeline": {"processors": [{"inference": {"model_id": "lang_ident_model_1","inference_config": {"classification": {"num_top_classes": 1}},"field_mappings": {"contents": "text"},"target_field": "_ml.lang_ident"}},{"rename": {"field": "contents","target_field": "contents.default"}},{"rename": {"field": "_ml.lang_ident.predicted_value","target_field": "contents.language"}},{"script": {"lang": "painless","source": "ctx.contents.supported = (['de', 'en', 'ja', 'ko', 'zh'].contains(ctx.contents.language))"}},{"set": {"if": "ctx.contents.supported","field": "contents.{{contents.language}}","value": "{{contents.default}}","override": false}}]},"docs": [{"_source": {"contents": "Das leben ist kein Ponyhof"}},{"_source": {"contents": "The rain in Spain stays mainly in the plains"}},{"_source": {"contents": "オリンピック大会"}},{"_source": {"contents": "로마는 하루아침에 이루어진 것이 아니다"}},{"_source": {"contents": "授人以鱼不如授人以渔"}},{"_source": {"contents": "Qui court deux lievres a la fois, n’en prend aucun"}},{"_source": {"contents": "Lupus non timet canem latrantem"}},{"_source": {"contents": "This is mostly English but has a touch of Latin since we often just say, Carpe diem"}}]}
返回结果：
{"docs" : [{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"contents" : {"de" : "Das leben ist kein Ponyhof","default" : "Das leben ist kein Ponyhof","language" : "de","supported" : true},"_ml" : {"lang_ident" : {"top_classes" : [{"class_name" : "de","class_probability" : 0.9996006023972855,"class_score" : 0.9996006023972855}],"model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:31:36.211596Z"}}},{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"contents" : {"en" : "The rain in Spain stays mainly in the plains","default" : "The rain in Spain stays mainly in the plains","language" : "en","supported" : true},"_ml" : {"lang_ident" : {"top_classes" : [{"class_name" : "en","class_probability" : 0.9988809847231199,"class_score" : 0.9988809847231199}],"model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:31:36.211611Z"}}},{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"contents" : {"default" : "オリンピック大会","language" : "ja","ja" : "オリンピック大会","supported" : true},"_ml" : {"lang_ident" : {"top_classes" : [{"class_name" : "ja","class_probability" : 0.9993823252841599,"class_score" : 0.9993823252841599}],"model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:31:36.211618Z"}}},{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"contents" : {"default" : "로마는 하루아침에 이루어진 것이 아니다","language" : "ko","ko" : "로마는 하루아침에 이루어진 것이 아니다","supported" : true},"_ml" : {"lang_ident" : {"top_classes" : [{"class_name" : "ko","class_probability" : 0.9999939196272863,"class_score" : 0.9999939196272863}],"model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:31:36.211624Z"}}},{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"contents" : {"default" : "授人以鱼不如授人以渔","language" : "zh","zh" : "授人以鱼不如授人以渔","supported" : true},"_ml" : {"lang_ident" : {"top_classes" : [{"class_name" : "zh","class_probability" : 0.9999810103320087,"class_score" : 0.9999810103320087}],"model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:31:36.211629Z"}}},{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"contents" : {"default" : "Qui court deux lievres a la fois, n’en prend aucun","language" : "fr","supported" : false},"_ml" : {"lang_ident" : {"top_classes" : [{"class_name" : "fr","class_probability" : 0.9999669852240882,"class_score" : 0.9999669852240882}],"model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:31:36.211635Z"}}},{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"contents" : {"default" : "Lupus non timet canem latrantem","language" : "la","supported" : false},"_ml" : {"lang_ident" : {"top_classes" : [{"class_name" : "la","class_probability" : 0.614050940088811,"class_score" : 0.614050940088811}],"model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:31:36.21164Z"}}},{"doc" : {"_index" : "_index","_type" : "_doc","_id" : "_id","_source" : {"contents" : {"en" : "This is mostly English but has a touch of Latin since we often just say, Carpe diem","default" : "This is mostly English but has a touch of Latin since we often just say, Carpe diem","language" : "en","supported" : true},"_ml" : {"lang_ident" : {"top_classes" : [{"class_name" : "en","class_probability" : 0.9997901768317939,"class_score" : 0.9997901768317939}],"model_id" : "lang_ident_model_1"}}},"_ingest" : {"timestamp" : "2020-03-06T16:31:36.211646Z"}}}]}
可以看到 Ingest Processor 非常灵活，且功能强大，所有的相关操作都可以在 Ingest processor 里面进行处理，再结合脚本做一下规则判断，对原始的字段重命名即可满足我们的文档处理需求。
今天我们聊了聊 Language Identity 这个功能，也聊了聊 Ingest Pipeline 的使用，怎么样，这个功能是不是很赞呀，如果有类似使用场景的朋友，可以自己试试看。另外值得注意的是，如果文本长度太小可能会识别不准，CLD3 设计的文本长度要超过 200 个字符。
相关链接：
CLD2: https://github.com/CLD2Owners/cld2
CLD3: https://github.com/google/cld3
Multilingual search using language identification in Elasticsearch ：https://www.elastic.co/blog/multilingual-search-using-language-identification-in-elasticsearch
ML Lang Ident 手册：https://www.elastic.co/guide/en/machine-learning/7.6/ml-lang-ident.html
Ingest Processor 手册：https://www.elastic.co/guide/en/elasticsearch/reference/7.6/inference-processor.html
