Using Rasa to build a FAQ bot 中文版本
Rasa is a powerful dialog framework that we can easily use it to build a practical dialog system.
FAQ(Frequently Asked Questions) bot is the most important part of the smart customer service products, which help users to solve general problems.
Usually we collect common questions and answers from actual scenarios and we call these standard question-and-answer pairs for the knowledge base. FAQ-bot will intelligently answer user questions based on this knowledge base.
So, in this project, we will use Rasa to build a smart faq-bot!
1. Run the shell and activate bert-service or computer sentence embedding with sentence-transformers
We store the knowledge in data/nlu/faq.json，which includes lots of question-and-answer pairs. We use bert-service to calculate the similarity between the user question and the questions in the knowledge base. Then we choose the most similar question and return the corresponding answer to the user.
- Specific operation：
- Install bert-serving-server and bert-serving-client，More information about bert-as-service
- Download pre-trained BERT model，unzip it(for example, cased_L-24_H-1024_A-16.zip).
- Next, you need to change "BERT_ENGLISH_MODEL_DIR" into your model path
-bert-serving-start \ -pooling_layer -4 -3 -2 -1 \ -model_dir=BERT_ENGLISH_MODLE_DIR \ -num_worker=8 \ -max_seq_len=16
- Run the shell
./run_bert_service.sh
- Alternatively, you can also choose to use
sentence transformerto coputer sentence embedding. Later, sentence embeeding can be used to calculate the similarity between the user question and the questions in the knowledge base and thus you can skip the earlier optionbert-as-service. The original repo of sentence transformer is available at here.
- In order to use this, please modify the
action.pyfile. Enable thesentence_transformer_selectflag as True and chosepretrained_modelfor sentence embedding with different pooling schemes are available, please refer to this link for more details.
sentence_transformer_select=True
pretrained_model='bert-base-nli-mean-tokens'-
There are several options for pre-trained BERT models with different pooling schemes are:
- bert-base-nli-mean-tokens: BERT-base model with mean-tokens pooling.
- bert-base-nli-max-tokens: BERT-base with max-tokens pooling.
- bert-base-nli-cls-token: BERT-base with cls token pooling.
- bert-large-nli-mean-tokens: BERT-large with mean-tokens pooling.
- bert-large-nli-max-tokens: BERT-large with max-tokens pooling.
- bert-large-nli-cls-token: BERT-large with CLS token pooling.
- Also, note that you can modify the matching score threshold criteria by changing the
score_thresholdinactions.pyfile based on your need.
2. Run Rasa custom actions
We design our own customer actions in actions.py which which is how the dialog responds to the user when it receives the customer's information. If the user's intent is faq, then go to the knowledge base to find the best match and give the user answers. If there is no match to the right question, we will also tell the user "Sorry, this question is beyond my ability."
- Change the port in endpoints.yml to keep the port the same as the port of rasa actions(default port is 5055):
action_endpoint:
url: "http://localhost:5055/webhook"- Run rasa actions：
rasa run actionsThen you can get a log like this:
│2019-08-09 11:10:32 INFO rasa_sdk.endpoint - Starting action endpoint server...
│(1000, 3072)
│2019-08-09 11:10:33 INFO rasa_sdk.executor - Registered function for 'action_get_answer'.
│2019-08-09 11:10:33 INFO rasa_sdk.endpoint - Action endpoint is up and running. on ('0.0.0.0', 5055)3. Run Rasa X
-
Install Rasa X
-
Run Rasa X
rasa xAnd you can get a log like this:
Starting Rasa X in local mode... �🚀
The server is running at http://localhost:5005/login?username=me&password=zrjV0BwYSzYPCopy the link into your browser, then you can access your rasa x page.
4. Note
- Sometimes when the service is started, the port is already in use. At this time, it is better to kill the process of the relevant port directly.
- If the error is about the database is locked，just delete rasa.db and tracker.db.
- You'd better use compatible rasa and rasa x，this project uses rasa 1.2.2 and rasa x 0.20.1(they are compatible)
- Use your own data to train a model:
- Follow the data foemat in data/nlu/faq.json，replace it by your own data
- Run process.py(Default data size is at most 1000 pieces，you can modify the process.py on your own)
- Run actions.py
- Please tune the matching threshold score as per your requirement.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent