S2KG ReadMe

Semi-Supervised Knowledge-Grounded Pre-training for Task-Oriented Dialog Systems

We present our models for Track 2 of the SereTOD 2022 challenge, which is the first challenge of building semi-supervised and reinforced TOD systems on a large-scale real-world Chinese TOD dataset MobileCS. We build a knowledge-grounded dialog model, S2KG to formulate dialog history and local KB as input and predict the system response.

This paper has been accepted at the SereTOD 2022 Workshop, EMNLP 2022

System Performance

Our system achieves the first place both in the automatic evaluation and human interaction, especially with higher BLEU (+7.64) and Success (+13.6%) than the second place. The evaluation results for both Track 1 and Track 2, which can be accessed via this this link.

🔥 News

[2022-11-19]: We release our S2KG-base model for knowledge-grounded dialogue generation !
[2022-10-10]: S2KG has been accepted at the SereTOD 2022 Workshop, EMNLP 2022 !
[2022-09-10]: 🏆 Achieved the 1st rank on SereTOD 2022 track 2 !

Model List

We released the S2KG-base model for knowledge-grounded dialogue generation. You can import this by using HuggingFace's Transformers.

Model
AndrewZeng/S2KG-base

Setup environment

pip install -r requirements.txt

Quick Start

from transformers.trainer_utils import speed_metrics
from transformers.file_utils import is_torch_tpu_available
from transformers import T5Tokenizer
from transformers import MT5ForConditionalGeneration

def convert_KB_to_sequences(KB):
    temp_seq = []
    for e in KB:
        if e == 'NA':
            NA_temp = []
            for na in KB['NA']:
                NA_temp.append(na+'：'+KB['NA'][na])
            temp_seq.append('；'.join(NA_temp))
        else:
            ent_info = KB[e]
            ent_temp = []
            for ent in ent_info:
                if ent == 'name':
                    ent_temp.append('名称：'+ent_info[ent])
                elif ent == 'type':
                    ent_temp.append('类型：'+ent_info[ent])
                else:
                    ent_temp.append(ent+'：'+ent_info[ent])
            temp_seq.append('；'.join(ent_temp))
    temp_seq = ' '.join(temp_seq)
    return temp_seq
    
local_KB = {
      "NA": {
        "欠费": "欠费,欠了_八块二毛八,八块多,八块二毛八",
        "持有套餐": "五十八的,它,活动"
      },
      "ent-1": {
        "name": "五十八的",
        "type": "业务",
        "业务费用": "五十八"
      },
      "ent-2": {
        "name": "它,活动",
        "type": "业务",
        "业务规则": "返了二十五块钱话费,每月返的二十五"
      }
    }
    
nlg_tokenizer = T5Tokenizer.from_pretrained(
        'AndrewZeng/S2KG-base',
        extra_ids=0,
    )  
nlg_model = MT5ForConditionalGeneration.from_pretrained(
        'AndrewZeng/S2KG-base',
    )
nlg_model.resize_token_embeddings(len(nlg_tokenizer))
KB_seq = convert_KB_to_sequences(local_KB)
dialog_his = '知识:'+KB_seq+' 用户0:' + '你好！' +' ' + ' 用户:'+'欠了多少钱？'
dialog_inputids = tokenizer(dialog_his, return_tensors="pt").input_ids
outputs = model.generate(dialog_inputids, num_beams=7, max_length=150)
outputs_txt = tokenizer.decode(outputs[0], skip_special_tokens=True)

Reproduce Results

You can reproduce our response generation results by:

cd ./response_gen
bash test_seq2seq_1B_submit.sh

Interact

You can interact with system using:

python3 interact.py

You can modify local KB in interact.py manually.

Zeng-WH/S2KG