Module 9: AI-Assisted Chatbot

** Objective and Relevance **

• Artificial Intelligence Assisted Counselling Chatbot (named as Rakshak)
• Languages: English, Hindi and Code-mixed
• Empathy-awareness
• Facebook, Whatsapp, Web Integration
• Victims may not express themselves freely to the investigative agencies or their parents or guardians
• Chatbot would facilitate the conversation with the victims
• Chatbot could provide useful suggestions to come out of the mental trauma by conversing in an empathetic way (understanding the victim’s emotions)
• Chatbot could also provide some legal suggestions and assistance

Requirements:

• Create environment for running trained model

  • conda create -n model

• Python

  • conda install python==3.8.0

• Torch

  • pip install torch==1.10.1
  • cuda==11.1

• Transformers

  • pip install transformers==4.36.2

• PEFT

  • pip install peft==0.7.1
  • When you installed PEFT, it will change torch version, so again install torch using given command(we did this because our GPU does not support torch 2.x)

• Pandas

  • pip install pandas==2.0.3

• Datasets

  • pip install datasets==2.16.1

• TRL

  • pip install trl==0.7.7

• Bitsandbytes

  • pip install bitsandbytes==0.41.3

• Create environment for RASA model

  • conda create -n rasa

• Python

  • conda install python==3.8.1

• RASA

  • pip install rasa==2.8.0
  • If getting “Descriptors cannot not be created directly” error while running rasa model, then install protobuf using below command
    • pip install protobuf==3.20.*

Hardware Requirement:

• GPU and cuda>=11.1

Running Generation model

• Activate the model environment
• Download the model from "https://drive.google.com/drive/folders/1n_uzrbSocH-gKI0h4E002jsaBnRt_0Qw?usp=drive_link"
• Put the model in the folder from where model_api.py can access the model
• Use "python model_api.py" for running the model file

Training RASA

• Use "Rakshak_English_Updated" or "Rakshak_Hindi_Updated" or "Rakshak_codemixed_Updated" rasa model
• Activate rasa environemnt
• Use "rasa train" for training RASA model

Running RASA

• rasa run -vv --enable-api --cors "*"
• rasa tun actions

Web Interface

• Activate model environemnt
• Run "python web.py"

Dialogue flow of conversations(used for training the model):

• Greeting("Hello Rakshak" or "हेलो रक्षक।")
• Name Information or Privacy concern from user
• Event narration
  • Online stalking
  • Online harassment
  • Online Masquerading
• Help information
  • Legal help
  • Law information
  • Organisation name
    • NCW
    • Cyber crime
    • Cyber cell
    • MWCD
  • Helpline numbers(nationwide numbers or state wise information)
• Counselling help
  • Women-in-distress
  • NCW
• Other queries(Charges, timing information etc.)
• Closing remark

Limitations

• Event narration
  • Grammatical error or wrong sentences formation
  • “He got my pictures along with my boyfriend and try to blackmailing me.”
    • Correct: He got my pictures along with my boyfriend and tried to blackmail me.”
  • Ambiguous words
    • I am uncomfortable
    • “वह मेरा फ़ायदा उठ रहा है”