Skip to main content

DeepPavlov (Sentiment analysis per message)

In this article I'll show you how to setup sentiment analysis using https://deeppavlov.ai/ and my prepared docker image

Required 4.48v Live Helper Chat version.

Installing DeepPavlov​

git clone https://github.com/LiveHelperChat/sentiment-per-message && cd sentiment-per-message
docker-compose -f docker-compose.yml pull
wget https://livehelperchat.com/var/deep_sentence_v2.tgz
tar zxfv deep_sentence_v2.tgz
rm -f deep_sentence_v2.tgz

Run one time

docker-compose -f docker-compose.yml up

Run as a service

docker-compose -f docker-compose.yml up -d

Testing

curl -X 'POST' \
  'http://127.0.0.1:5058/model' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "sentences": [
    "I'\''m sad"
  ]
}'

If you did everything right you should see output like this

[
  [
    "negative"
  ],
  [
    [
      0.0346120223402977,
      0.8017117977142334,
      0.023267682641744614,
      0.14040860533714294
    ]
  ]
]

You can also point your browser to

http://127.0.0.1:5058/docs

Configuring Live Helper Chat​

Our requirements are

  • We should send set sentiment on chat close event
  • We should see visitor sentiment while chat is going on
  • Support ElasticSearch for sentiment analysis.

For that we will be using webhooks.

Configuring Rest API call​

In this scenario Rest API will be sending individual message to get a sentiment.

Download configuration

After import make sure you change host if you are not running it on local machine.

Configuring bot​

Download configuration you will need to set appropriate Rest API calls.

In the bot for simplicity we will have

  • Sentiment User Message - this trigger is executed once we receive a message from visitor/operator (chat.web_add_msg_admin,chat.chat_started,chat.addmsguser) events
  • Evaluate Sentiment Message - this trigger works with response from Rest API call
  • Sentiment Chat - this trigger does all the aggregation and set's main chat sentiment attributes (chat.close event)

In this configuration we set following sentiment attributes

  • Set sentiment for the operator and visitor.
  • Update visitor sentiment within each message.
  • Update visitor sentiment per chat

Main configuration for the Rest API Call​

I strongly suggest to use php-resque extensions and offload all Rest API calls...

If you webhooks worker is already resque type you can un-check. Send Rest API Call in the background. as process already will be running in the background.

Evaluate Sentiment Message​

This trigger stores sentiment withing chat message and recalculates sentiment for the chat as a whole.

More information about Messages aggregation

Sentiment Chat​

Calculates visitor messages sentiment using SUM as comparator and AVG as value

Calculates operator messages sentiment using SUM as comparator and AVG as value

Calculates ratio of positive messages in comparison to negative and positive messages.

Configuring webhook​

In webhook we define what event we want to listen.

Presently we are interested in for events

  • chat.web_add_msg_admin - listen for operator message and set sentiment
  • chat.chat_started - listen for very first visitor message
  • chat.addmsguser - listen for visitor message
  • chat.close - make summary of sentiment based on chat messages

Bonus​

Using this configuration you can also show custom icon based on chat sentiment.