Rasa NLU GQ

Rasa NLU (Natural Language Understanding) 是一个自然语义理解的工具，举个官网的例子如下：

"I'm looking for a Mexican restaurant in the center of town"

And returning structured data like:

  intent: search_restaurant
  entities: 
    - cuisine : Mexican
    - location : center

Introduction

原来的项目在分支0.2.7上，可自由切换。这个版本的修改是基于最新版本的rasa，将原来rasa_nlu_gao里面的component修改了下，并没有做新增。并且之前做法有些累赘，并不需要在rasa源码中修改。可以直接将原来的component当做addon加载，继承最新版本的rasa，可实时更新。

New features

目前新增的特性如下（请下载最新的rasa-nlu-gao版本）(edit at 2019.06.24)：

新增了实体识别的模型，一个是bilstm+crf，一个是idcnn+crf膨胀卷积模型，对应的yml文件配置如下：

  language: "zh"

  pipeline:
  - name: "JiebaTokenizer"
  - name: "CountVectorsFeaturizer"
    token_pattern: "(?u)\b\w+\b"
  - name: "EmbeddingIntentClassifier"
  - name: "rasa_nlu_gao.extractors.bilstm_crf_entity_extractor.BilstmCRFEntityExtractor"
    lr: 0.001
    char_dim: 100
    lstm_dim: 100
    batches_per_epoch: 10
    seg_dim: 20
    num_segs: 4
    batch_size: 200
    tag_schema: "iobes"
    model_type: "bilstm" # 模型支持两种idcnn膨胀卷积模型或bilstm双向lstm模型
    clip: 5
    optimizer: "adam"
    dropout_keep: 0.5
    steps_check: 100

新增了jieba词性标注的模块，可以方便识别名字，地名，机构名等等jieba能够支持的词性，对应的yml文件配置如下：

  language: "zh"

  pipeline:
  - name: "JiebaTokenizer"
  - name: "CRFEntityExtractor"
  - name: "rasa_nlu_gao.extractors.jieba_pseg_extractor.JiebaPsegExtractor"
    part_of_speech: ["nr", "ns", "nt"]
  - name: "CountVectorsFeaturizer"
    OOV_token: oov
    token_pattern: "(?u)\b\w+\b"
  - name: "EmbeddingIntentClassifier"

新增了根据实体反向修改意图，对应的文件配置如下：

  language: "zh"

  pipeline:
  - name: "JiebaTokenizer"
  - name: "CRFEntityExtractor"
  - name: "JiebaPsegExtractor"
  - name: "CountVectorsFeaturizer"
    OOV_token: oov
    token_pattern: '(?u)\b\w+\b'
  - name: "EmbeddingIntentClassifier"
  - name: "rasa_nlu_gao.classifiers.entity_edit_intent.EntityEditIntent"
    entity: ["nr"]
    intent: ["enter_data"]
    min_confidence: 0

新增了bert模型提取词向量特征，对应的配置文件如下：

  language: "zh"

  pipeline:
  - name: "JiebaTokenizer"
  - name: "rasa_nlu_gao.featurizers.bert_vectors_featurizer.BertVectorsFeaturizer"
    ip: '127.0.0.1'
    port: 5555
    port_out: 5556
    show_server_config: True
    timeout: 10000
  - name: "EmbeddingIntentClassifier"
  - name: "CRFEntityExtractor"

新增了对CPU和GPU的利用率的配置，主要是EmbeddingIntentClassifier和ner_bilstm_crf这两个使用到tensorflow的组件，配置如下（当然config_proto可以不配置，默认值会将资源全部利用）：

  language: "zh"

  pipeline:
  - name: "JiebaTokenizer"
  - name: "CountVectorsFeaturizer"
    token_pattern: '(?u)\b\w+\b'
  - name: "EmbeddingIntentClassifier"
    config_proto: {
      "device_count": 4,
      "inter_op_parallelism_threads": 0,
      "intra_op_parallelism_threads": 0,
      "allow_growth": True
    }
  - name: "rasa_nlu_gao.extractors.bilstm_crf_entity_extractor.BilstmCRFEntityExtractor"
    config_proto: {
      "device_count": 4,
      "inter_op_parallelism_threads": 0,
      "intra_op_parallelism_threads": 0,
      "allow_growth": True
    }

新增了embedding_bert_intent_classifier分类器，对应的配置文件如下：

  language: "zh"

  pipeline:
  - name: "JiebaTokenizer"
  - name: "rasa_nlu_gao.featurizers.bert_vectors_featurizer.BertVectorsFeaturizer"
    ip: '127.0.0.1'
    port: 5555
    port_out: 5556
    show_server_config: True
    timeout: 10000
  - name: "rasa_nlu_gao.classifiers.embedding_bert_intent_classifier.EmbeddingBertIntentClassifier"
  - name: "CRFEntityExtractor"

在基础词向量使用bert的情况下，后端的分类器使用tensorflow高级api完成，tf.estimator,tf.data,tf.example,tf.saved_model intent_estimator_classifier_tensorflow_embedding_bert分类器，对应的配置文件如下：

language: "zh"

pipeline:
- name: "JiebaTokenizer"
- name: "rasa_nlu_gao.featurizers.bert_vectors_featurizer.BertVectorsFeaturizer"
  ip: '127.0.0.1'
  port: 5555
  port_out: 5556
  show_server_config: True
  timeout: 10000
- name: "rasa_nlu_gao.classifiers.embedding_bert_intent_estimator_classifier.EmbeddingBertIntentEstimatorClassifier"
- name: "SpacyNLP"
- name: "CRFEntityExtractor"

Quick Install

pip install rasa-nlu-gao

Some Examples

具体的例子请看rasa_chatbot_cn

TODO