Server

KoGPT2

• Key model for chatbots that generate answers to users' chats

Baseline & Dataset

Baseline code : Korean Language Model for Wellness Conversation

Dataset : 주제별 텍스트 일상 대화 데이터

Process

• Data Preparation
• Data Preprocessing
• Using pretrained weight
• Train the model
• Validate the model

How to make AIFriend?

1. Data Preparation

• Download the dataset

• This project used only Kakao data in the source data compression(원천데이터 압축파일) in the Training/Validation folder in dataset.
  If you proceed in the same environment as the project, you only need to use Kakao data.

• If necessary, you can proceed with the learning by using additional refining.

(Project folder)/data/kakao - Place the prepared KAKAO.txt files (Collect all the txt files in the /data/kakao/)

---

2. Data Preprocessing

• In (Project folder)/preprocess/split_new.py code, Check the path that exists at the top, set it correctly, and run it.
  It preprocesses data, Kogpt2 can be fine-tuned to AIFriend.

• At this time, exceptions are processed so that certain words such as politics/ideologies are not learned for a document, but you can also modify the project file directly to make better exceptions.

---

3. Using pretrained weight

• Use the Word2Vec pretrained model for extracting interest categories(download link)
  
  Place the files to (Project folder)/core/AIfriend/Keybert/

• Use the kogpt2 pre-trained model provided by SKT (provided by transformers library)

• The pretrained weight of the AIfriend that we used in the project(download link)
  (Please note that the model weight may have learned the bias or incorrect knowledge of the data)

---

4. Train the model

• In (Project folder)/train/run_auto_regressive.py code, Check the path that exists at the top, set it correctly, and run it.
  Automatically load the skt pre-trained kogpt2 model and training with preprocessed data.

• Hyperparameters such as learning rate and epoch can be adjusted as you want.

---

5. Validate the model

• You can test directly on the code before testing on the server and client.

• In (Project folder)/example/kogpt2-text-generation.py code, AIfriend can be tested using trained kogpt2 before it is used on the server.

Server with Firestore(DB)

• Connect between the server and the client by network socket and using thread to make multiple connections
• Manage conversations with chatbots and expansion of interests channel, the one of the core of the AIfriend project

Demo.py

Load various pre-trained models

#model loading
root_path = '../..'   
...   
model_W2V = word2vec.Word2Vec.load('./Keybert/model_W2V')

• Load the pretrained model that contains KoGPT2, Word2Vec, Tokenizer, SentenceTransformer.
• Please write the correct path

---

Modify the category

# Adding category in here.
category = ['여행', '음악', '게임', '동물', '옷', '음식', '운동', '독서', '요리']

• Modify the category that you want.
  If you change the category, you should make a new fav, Board (Reference)

---

Main loop

try:
    while True:
        print('>> Waiting for a new connection')
        client_socket, addr = server_socket.accept()  
        user_sockets.append(client_socket)
        print("Current user : ", len(user_sockets))
        start_new_thread(threaded, (client_socket, addr))
except ...
finally ...

• Generate a thread by receiving user information from the network socket

---

Thread server

def threaded(client_socket, addr):

• When a new user accesses the server, it is managed through a new thread.

...
data = client_socket.recv(2048)
...
original_data = data.decode('utf-8')
...
if original_data[:6] == 'AIchat':
                    uid = original_data[6:]
                    KoGPT(uid, db, model, tokenizer, push_service, model_ST, model_W2V, category)
...

• The client sends a prefix AIchat to have a conversation with chatbot.
  Additionally, receive a uid to access firestore.

---

functions.py

Main function(KoGPT)

• Called by network socket prefix AIchat, Create AIfriend answers and make judgments about adding interests.

def KoGPT(...):
...
# Document search
    document_name = db.collection(u'AIChat').where('uid', 'array_contains', uid).get()[0].id
    AIchat_ref = db.collection(u'AIChat').document(document_name).collection('Chats')

• Chatting log document is loaded by Document search

def KoGPT(...):
...
    tokenized_indexs = tokenizer.encode(user_chat_list[0])
    input_ids = torch.tensor([tokenizer.bos_token_id, ] + tokenized_indexs + [tokenizer.eos_token_id]).unsqueeze(0)
    sample_output = model.generate(input_ids=input_ids)
    answer = tokenizer.decode(sample_output[0].tolist()[len(tokenized_indexs) + 1:], skip_special_tokens=True)

• Use tokenizer and trained KoGPT2 to generate answers to user chat

def KoGPT(...):
...
    # In each 'fav_max_count', server starts extracting interests
    if keybert_check != -1:
        keybert_check = keybert_check % fav_max_count
    if keybert_check == 0:
        # Category connecting each 'fav_max_count' user chatting
        category_connect(uid, db, model_ST, model_W2V, category)

• In each 'fav_max_count', call category_connect(...) to connect the user and interest board

---

Main function(category_connect)

def category_connect(...):
    bert_keyword = key_bert(uid, db, model_ST, model_W2V, category)

• Call key_bert(...) to get keyword in user chatting.

def category_connect(...):
...
    if bert_keyword in category:
        email = db.collection(u'user').where('uid', '==', uid).get()[0].id
        check = db.collection("fav").document(bert_keyword).get().to_dict()['users']
...
            db.collection("fav").document(bert_keyword).update({"users": firestore.ArrayUnion([email])})
            AIchat_ref.add({'message': bert_keyword + '에 관심있구나! 내가 비슷한 취향을 가진 친구들을 소개시켜줄게! 내 관심사 탭에 가볼래?', 'time': firestore.SERVER_TIMESTAMP, 'uid': 'AIfriend'})

• If a user's interest keyword is in a category, then add the user to the interest board.

---

Sub functions

def getToken(...):

• Function used to acquire token of client.

def sendMessage(...):

• Use the token of the client to raise the push alarm.

def chatting_delay(...):

• Delay function to make AIfriend feel like a person

def key_bert(...):

• Checking the reference.