Minute and thoughts of the launch conference of THUKC (清华大学人工智能研究院知识智能研究中心).
Mainly about the part of Commonsense-Awared Natural Language Generation (常识知识感知的语言生成, 黄民烈)
Commonsense Extraction
- what is Commonsense Knowledge (CS Know.)
- what is the Boundary of commonsense
- Commonsense Extraction
- From Embedding
- Extracting Commonsense Properties from Embeddings with Limited Human Guidance. 2018
- Commonsense Knowledge base completion
- Commonsense Knowledge base Completion. 2018
- From RAW data(text, image)
- Automatic Extraction of Commonsense LocatedNear Knowledge. 2018
- From Embedding
Commonsense Knowledge in RC
- 代表工作:Knowledgeable Reader
Commonsense Knowledge to Intent, Reaction, Emotion, etc
- 代表工作:
- Event2Mind: Commonsense Inference on Events, Intents and Reaction
- Modeling Naive Psychology of Characters in Simple Commonsense Stories
- AI2 实验室做了大量这方面的工作,相关调研:link
Commonsense to Controllable TG
- 黄民烈老师组的一些很好的工作(黄老师称这些工作是常识知识感知的语言生成的初步探索):
- Three Fundamental problems in current Neural Language Generation Models
- semantics(real understanding)
- Consistency(long text generation)
- Logic(reasonable and making sense)
- 在评价方面,关于如何评价生成的文本具有逻辑性还是个挑战(相关研究点:文本生成评价指标(BLEU、ROUGE等)的研究)
- New Architecture:
- symbolic knowledge + planning + neural computing
发布的重要资源
知识智能研究中心:http://ai.tsinghua.edu.cn/kirc/#
- 中英文跨语言百科知识图谱XLORE — 世界知识
- 特点:
- 大规模跨语言百科知识图谱
- 通过融合维基百科和百度百科,并对百科知识进行结构化和跨语言链接构建而成。
- 以结构化形式描述客观世界中的概念、实例、属性及其丰富语义关系。
- XLORE目前包含约247万概念、44.6万属性/关系、1628万实例和260万跨语言链接。
- 项目地址:https://xlore.org/
- 特点:
- 基于义原的开放语言知识库 OpenHowNet — 语言知识、常识知识
- !HowNet 核心数据首次开源
- 项目地址:https://openhownet.thunlp.org/
- 1.https://juejin.im/post/5b6a9e085188251a8d37136d ↩