字里医间Paper to Practice 语言、数据与 NLP 的基础知识。
NLP 学习笔记 ·
从幻觉和过时信息出发，经过 FActScore、DoveScore 与 RAG，再到 RLHF、偏差的缓解和可解释性，理解大语言模型的问题从哪里来、又能怎样改进。
From hallucination and stale information through FActScore, DoveScore and RAG to RLHF, bias mitigation and interpretability: where the problems of large language models come from and how they can be reduced.
NLP 学习笔记 ·
从 token、更长的前文和注意力机制三点改进出发，经过 Transformer 与 GPT、Llama 两大家族，理解提示、上下文学习、思维链和专门训练来推理的模型。
From three improvements on neural language models (tokens, longer context and attention) through the Transformer and the GPT and Llama families: prompting, in-context learning, chain-of-thought and models trained to reason.
NLP 学习笔记 ·
从语言模型的基本功能出发，经过词嵌入、神经元、激活函数和反向传播，理解神经网络怎样学会预测下一个词。
From what every language model computes through embeddings, neurons, activation functions and backpropagation: how a neural network learns to predict the next word.
NLP 学习笔记 ·
从字母和单词的频率出发，经过条件概率与二元模型，理解简单语言模型怎样给句子打分、又怎样一个词一个词地生成句子。
From letter and word frequencies through conditional probability and bigrams: how a simple language model scores a sentence and writes one word at a time.
NLP 学习笔记 ·
从词位和词义关系出发，经过 WordNet 与 FrameNet，理解计算机怎样一步步建立知识图谱，以及 UMLS、MeSH 在医学中的作用。
From lexemes and word relations through WordNet and FrameNet: how computers build up to knowledge graphs, and what UMLS and MeSH do in medicine.
NLP 学习笔记 ·
从 token、句子边界到词性标注和命名实体识别，理解文本如何被切分并贴上标签。
From tokens and sentence boundaries to POS tagging and named entity recognition: how text is split and labeled.
NLP 学习笔记 ·
从词的结构、词形处理和正则表达式，理解文本如何成为可分析的材料。
Connect word structure, normalization and regular expressions to understand how text becomes material for analysis.
NLP 学习笔记 ·
从词句、语境到声音线索，理解自然语言处理能做什么，以及它与访谈研究的联系。
Words, context, and vocal cues: an introduction to natural language processing and its connection to interview research.
