4.27 CIPS SMP Notes

1 minute read

Published:

4.27 中国中文信息学会 社会媒体处理(SMP)专委会 首届“社交机器人”论坛

Pretraining in NLP and CV

ULMFit

double fine-tune

CoVe

Supervised Language Model

MS: KnowledgeBERT for Semantic Parsing

MSPars: A Multi-Perspective Semantic Parsing Dataset for Knowledge-based Question Answering

Pre-training-based Natural Language Generation

BERT-output -> draft(Text Summarization) -> Transformer Decoder -> output

Pre-training in ImageNet/Multi-Modal

ImageBERT Image object + text -> BERT

VideoBERT(Google)

videos -> frames -> token -> BERT

Application: Video QA, Summarization and Chat.

Video span like machine reading comprehension.

NLU in Task-Oriented Dialog System

intent detection, slot filling, state tracking.

Intent Detection

Slot Filling

Dialog Management

Reinforcement Learning

Dialog Generation

based on Pattern/LM/Seq2Seq

对话技术平台

希望中小型开发者只需要上传自己的数据,即可开发自己的对话系统。

哈工大聊天机器人“笨笨”/任务型对话系统

Few-shot Learning

小样本的训练

标注数据的自动扩充 Seq2Seq/Pre-training

How to do few-shot learning on sequence labeling(slot filling) task?

Draw-back of traditional dialong system

Depend on previous dialog.

Joing Training: Intent detection + Slot filling based on Stack Propagation(important) and Multi-Task

Evaluation of Dialog System Techniques

ECDT2017-2018

Alibaba

Asememble Learning Hybrid CNN
冷启动快速端到端测试 Deep reinforcement Learning

MRC for Unstructured Data

based on SLQA -> EMNLP

Open-Domain Non-Oriented Dialog System

Retrieval-Based Chatbot

Multi-View: Relevence, Interestingness, Informativeness,

Non-Sentential Utterance Resolution

Retrieval from Non-Dialogue Corpus

  • 和阅读理解有什么区别? 专业知识可能会粒度特别的细,所以需要开放域(大概是这个意思)

Neural Responce Generation

  • The “Bland Response” Problem: I dont’t know/Well/Great/Fine, Jiwei Li
    • Adversarial Training
    • discriminator生成的是比较细粒度的东西,用一个单一的score作为reward去回传会不会有问题,所以把????和embedding直接乘起来, 剩下的没记下来orz 好像组会讲过XD
  • The “Myopia Problem” of Beam Search

Child Friendly Social Chatbot

  • 避免一些对小孩不合适的话题,谈恋爱,结婚生子等
  • 先用一个用儿童语料库训练的语言模型去过滤
  • 生成模型本身比较保守,双重过滤后概率会远小于真实世界的概率

Q&A

Constraint on Dialog Generation