前往小程序,Get更优阅读体验!
立即前往
首页
学习
活动
专区
工具
TVP
发布
社区首页 >专栏 >Week 4 Language Use in the Wild

Week 4 Language Use in the Wild

作者头像
杨丝儿
发布2022-03-01 12:57:36
2090
发布2022-03-01 12:57:36
举报
文章被收录于专栏:杨丝儿的小站杨丝儿的小站

Lecture 10.1 Communication

Key linkage

  • Communication is a goal: to transfer information from source to destination.
  • If M and F are independent, H(M∣F)=H(M)H(M|F) = H(M)H(M∣F)=H(M).
  • If F deterministically map to M, H(M∣F)=0H(M|F) = 0H(M∣F)=0.

Key terms

communication, information, source, destination, computational level goal, necessary subgoal

source, probability distribution over meaning, transmitter, encoder, function, meaning, form, receiver, decoder, noise source, reference resolution

information theory, information content, surprise, entropy, uncertainty, outcome, conditional entropy, mutual information, deterministic, KL divergence

Equation or Formula or method collection

  • Surprisal: I(x)=−logxP(x)I(x)=-log_xP(x)I(x)=−logx​P(x)
  • Entropy: H(X)=∑xP(x)I(x)H(X)=\sum_x P(x)I(x)H(X)=∑x​P(x)I(x)
  • Mutual information: the information shared between two variables I(M;F)=H(M)−H(M∣F)I(M; F) = H(M) - H(M|F)I(M;F)=H(M)−H(M∣F)
  • Kullback-Leibler (KL) Divergence: The information gained by using distribution P compared to distribution Q KL(P∣∣Q)=−∑xP(x)logQ(x)P(x)KL(P||Q) = -\sum_xP(x)log\frac{Q(x)}{P(x)}KL(P∣∣Q)=−∑x​P(x)logP(x)Q(x)​

Lecture 10.2 Computational Psycholinguistics

Key terms

tool, frequent, homophone, word sense, information processing effect, reading time,

Key linkage

  • Language as a communication system seems optimized for efficiency.
  • Highly needed words are:
    • short in length
    • close to each other in sentences
    • versatile
  • Homophones: Same sounds but different meanings
  • Word Sense: A single meaning of a word

Lecture 11 Linguistic Diversity & Efficient Communication

Key linkage

  • Sapir-Whorf Hypothesis
    • Languages carve up the world in different ways.
    • Does this influence our conceptual system?
  • You cannot minimize two superlatives without specifying how they trade-off.
  • Objective Function
    • Minimize Description Length (Algorithmic Complexity)
    • Minimize Communicative Cost: KL[M∣∣M^]KL[M||\hat{M}]KL[M∣∣M^]
  • There are a variety of fields that study linguistic diversity.
    • Sociolinguistics
    • Linguistic Anthropology
    • Linguistic Typology
  • Semantic typology studies how different languages carve up the world.
  • Recently, information theory has helped characterize and explain the diversity of semantic systems as communicative efficiency trade-offs.
  • This is in line with the soft take of the Sapir-Whorf Hypothesis.

Key terms

linguistic diversity, Sapir-Whorf Hypothesis, World Color Survey, Munsell Color Chart, basic color term, efficiency trade-off, principle of least effort, force of unification, force of diversification, communicative efficiency trade-offs, the information bottleneck, theoretical limit, observation, hypothetical variant, kinship system, objective function, description length, communication cost,

Equation or Formula or method collection

  • The information bottleneck, Fβ[q(w∣m)]=Iq(M;W)−βKL[M∣∣M^]=Iq(M;W)−βIq(W;U)F_β[q(w|m)]= I_q(M; W )-βKL[M||\hat{M}]= I_q(M; W )-βI_q(W ;U)Fβ​[q(w∣m)]=Iq​(M;W)−βKL[M∣∣M^]=Iq​(M;W)−βIq​(W;U)

Lecture 12.1 Vector Space Semantics: Intuition

Key terms

context vector, context words, target words, distributional similarity, syntactic categories, context window, wording meaning, LSA, dimensionality reduction step,

Lecture 12.2 Vector Space Semantics: Implementation

Key terms

learn, word embedding, neural network, guess a word from its context, representation of the context, representation of the word, words within a context window, count vector, Word2V Model, the input and output layers are one-hot encoded, each word is represented as a vector of size V, the number of words in the vocabulary, representation learning

Key linkage

  • after training, the hidden layer hih_ihi​ for target word yiy_iyi​ can be used as its context vector.

Lecture 12.3 Vector Space Semantics: Evaluation

Key terms

model evaluation, semantic priming, lexical decision experiment,

Key linkage

  • Strengths:
    • learns word representations automatically from raw text;
    • simple approach: all we need is a corpus and some notion of what counts as a word;
    • language-independent, cognitively plausible.
  • Weaknesses:
    • many ad-hoc parameters when creating the embeddings;
    • ambiguous words: their meaning is the average of all senses;
    • no representation of word order
本文参与 腾讯云自媒体同步曝光计划,分享自作者个人站点/博客。
原始发表:2021-02-23,如有侵权请联系 cloudcommunity@tencent.com 删除

本文分享自 作者个人站点/博客 前往查看

如有侵权,请联系 cloudcommunity@tencent.com 删除。

本文参与 腾讯云自媒体同步曝光计划  ,欢迎热爱写作的你一起参与!

评论
登录后参与评论
0 条评论
热度
最新
推荐阅读
目录
  • Lecture 10.1 Communication
    • Key linkage
      • Key terms
        • Equation or Formula or method collection
        • Lecture 10.2 Computational Psycholinguistics
          • Key terms
            • Key linkage
            • Lecture 11 Linguistic Diversity & Efficient Communication
              • Key linkage
                • Key terms
                  • Equation or Formula or method collection
                  • Lecture 12.1 Vector Space Semantics: Intuition
                    • Key terms
                    • Lecture 12.2 Vector Space Semantics: Implementation
                      • Key terms
                        • Key linkage
                        • Lecture 12.3 Vector Space Semantics: Evaluation
                          • Key terms
                            • Key linkage
                            领券
                            问题归档专栏文章快讯文章归档关键词归档开发者手册归档开发者手册 Section 归档