Academic Concept Dictionary

1 Semester 1

1.1 General Psychology

• Scientific method（科学方法）：test claims that could be proven wrong using
  observable evidence. The dividing line in psychology is not "what it studies" but
  "how it studies" — can it talk with data.
  ⚠️ Often confused with：the everyday sense of "psychology" (mind reading, personality
  analysis) is not the same thing as academic psychology.
• Description / correlation / experiment（描述 / 相关 / 实验；三种研究方法）：
  Description only records "what is"; correlation says "two things change together";
  experiment is the only one that can say "who causes whom". Whether you can claim
  causation depends on whether you manipulated a variable + randomly assigned.
• Independent variable / dependent variable (IV / DV)（自变量 / 因变量）：the independent
  variable is the one you actively change (e.g. the colour of the ink), the dependent
  variable is the one you measure (e.g. reaction time). Memory hook: IV = cause, DV = effect.
• Random assignment（随机分配）：putting participants into conditions at random.
  Its job is not "to make the groups the same" but to make the groups expected to differ in
  no way on all the factors you never thought of — this is the key step that lets an
  experiment claim causation.
• Weber's law（韦伯定律）：the difference you can just barely notice is proportional to the
  original stimulus intensity.
  Example： holding 1 kg, adding 10 g feels different; holding 10 kg, adding 10 g feels like nothing.
• Absolute threshold / difference threshold（绝对阈限 / 差别阈限）：the absolute
  threshold = the smallest amount of stimulus you can just barely feel; the difference
  threshold = the smallest change you can just barely tell apart as "it changed". Weber's law
  is about the latter.
• Signal Detection Theory, SDT（信号检测论）：a framework that separates "how sharp
  you are" from "how fond you are of saying 'yes'". Because it found that: given the
  same sensation, whether you report "yes" or "no" also depends on your decision criterion.
• Hit / miss / false alarm / correct rejection（击中 / 漏报 / 虚报 / 正确拒绝）：the four
  cells of SDT. Hit = there is a signal and you say yes; miss = there is a signal but you say
  no; false alarm = there is no signal but you say yes; correct rejection = there is no signal
  and you say no.
• d′（辨别力指数）：the index in SDT that measures "true sensitivity" — the larger the
  more sensitive, and unaffected by how fond you are of saying yes. This is exactly what
  is clever about SDT.
• Criterion（判准）：the height of the bar at which you report "yes". Low bar →
  easy to false-alarm; high bar → easy to miss. The criterion shifts with motivation and
  consequences (e.g. in an exam, "only answer when you are sure").
• Sensation / perception（感觉 / 知觉）：sensation = receiving raw stimulation (light,
  sound); perception = the organisation and interpretation of sensation.
  Perception is not passive reception, it is active construction (→ the gorilla experiment).
• Inattentional blindness（非注意盲视）：what you do not attend to, you really do not
  "see".
  Example： while counting passes, half the people never see the gorilla walking across the screen.
• Dichotic listening（双耳分听）：a different message in each ear, and you repeat back only
  one ear. Used to study "at which step the filter of attention sits".
• Filter model / attenuation model / late selection / perceptual load（过滤器模型 / 衰减模型 / 后期选择 / 知觉负载）：
  four models of "at which step attention screens out information". Filter = screen
  everything out early; attenuation = only weaken it early, do not screen it out; late
  selection = process it all, then screen; perceptual load = it depends on how busy you
  are — if you are not busy, it leaks in. They overturn one another; they are not parallel
  alternatives.
• Automatic processing（自动化加工）：processing that stops costing attention once you
  have practised it (e.g. reading words). The Stroop effect is it fighting with controlled
  processing: you are asked to name the ink colour, but the meaning of the word automatically
  runs out and interferes.
• Stroop effect（Stroop 效应）：write "red" in blue ink and ask someone to say the ink
  colour → they slow down. Proves that "reading a word" is automatic and cannot be
  suppressed.
• Sensory memory / short-term memory / working memory（感觉记忆 / 短时记忆 / 工作记忆）：
  sensory memory = the instant afterglow of a stimulus that has just passed (under 1 second);
  short-term memory = seconds to tens of seconds, very small capacity; working memory =
  the part where you hold things and operate on them at the same time (the upgraded version).
• Chunk（组块）：packing several small units into one unit, which makes your capacity
  bigger.
  Example： "19491001" is easier to remember than "1-9-4-9-1-0-0-1". Experts are strong because
  their "chunks are big", not because their "RAM is big".
• 7±2 / 4：the number of units you can hold in mind at once over the short term. Miller
  said 7±2 (1956), Cowan revised it to about 4 (2001).
• Partial report（部分报告法）：flash 12 letters, and ask them to report only one row.
  The result is far better than "report everything" → showing that the information was all
  there, it just faded too fast.
• Reconstructive memory（重构记忆）：memory is not a video replay, it is reassembled each
  time.
  Misinformation effect: changing one word in a question after the fact (was it a "smashed"
  or a "bumped"?) can change the details you remember.
• Intrinsic motivation / extrinsic motivation（内在动机 / 外在动机）：intrinsic = doing
  it because you like it; extrinsic = doing it for a reward / to avoid punishment.
• Overjustification effect（过度理由效应）：adding a reward to something you already enjoy
  makes you enjoy it less. Because your reason shifted from "I like it" to "I want the prize",
  and once the reward is withdrawn, the reason is gone.
• Expectancy-value theory（期望-价值理论）：motivation = how likely you think success is ×
  how much this matters to you. If either end is 0, motivation is 0.
• Conservation（守恒）：a Piaget concept — knowing that "the shape changed but the amount
  did not".
  Example： pour water from a short glass into a tall thin one, and a preoperational child says
  there is "more".
• Big Five / OCEAN（大五人格）：describing personality along five dimensions: openness,
  conscientiousness, extraversion, agreeableness, neuroticism. It is the most widely accepted
  framework at present ("five words" describe a person fairly stably).
• Conformity（从众）：changing your own judgement because of group pressure (even when
  you know the answer). Asch's experiment: the lines are plainly the same length, but when the
  majority says otherwise, about 1/3 of participants go along and answer wrongly.
• Obedience（服从）：complying with an authority's commands. Milgram's experiment:
  65% of people turned the "shock" up to the highest setting, purely because a man in a white
  coat said "please continue".
• Observational learning（观察学习）：you can learn without experiencing reward and
  punishment yourself — just by watching others. Bandura's Bobo doll experiment: children who
  had watched an adult hit the doll hit it themselves. This directly refutes "you must be
  reinforced yourself to learn".

1.2 Principles of Education

• Is / ought（实然 / 应然）：is = how things actually are (testable with evidence);
  ought = how things should be (a value judgement).
  ⚠️ The single most central thing in this course: other courses teach you "what is", this
  course teaches you "what should be" — the two kinds of question are answered in completely
  different ways, so do not use data to answer a value question.
• The three basic elements of education（教育的三个基本要素）：the educator, the educated,
  and the educational influence (content and means). The relations among the three (who leads,
  who is active) are exactly what the various schools argue about.
• The nature of education（教育的本质）："what education actually is" is itself a disputed
  question — some say "transmitting social experience" (society-centred), some say "promoting
  human development" (individual-centred). There is no standard answer; it depends on your
  stance.
• Heredity / environment / education / subjective agency（遗传 / 环境 / 教育 / 主观能动性）：
  the four factors that influence human development. Education plays the leading role — but
  note: "leading" is not "omnipotent"; it depends on heredity and environment to provide the
  conditions, and in the end it only works through the person's own agency.
• Zone of Proximal Development, ZPD（最近发展区）：Vygotsky's idea — the stretch between
  "what you can do on your own" and "what you can do with someone's help".
  Teaching should aim at that stretch: too easy and there is no growth, too hard and it is
  out of reach.
• Scaffolding（支架）：temporary help given to the student inside the ZPD (hints,
  demonstrations, breaking the steps down), removed once they have got it. The name comes
  from the scaffolding on a building site.
• Individual-oriented / society-oriented views（个人本位 / 社会本位；教育目的之争）：
  individual-oriented = education for the sake of human development itself (Rousseau, Dewey);
  society-oriented = education for the sake of society's needs (Durkheim). This is a value
  choice, not a factual question.
• Moral education / knowing, feeling, willing, acting（德育 / 知情意行）：moral character has
  four aspects: knowing (cognition), feeling (emotion), willing (volition), acting
  (behaviour). The four are not necessarily in step — "knowing what one should do" is not
  the same as "doing it".
• Teacher professional development / teacher-student relationship（教师专业发展 / 师生关系）：
  a teacher is not a "technician" but a specialised profession that requires continuous
  growth; the teacher-student relationship is the background colour of education, not an
  optional extra.

1.3 Introduction to Linguistics

• Saussure（索绪尔）：the father of modern linguistics. His core contribution was turning
  "language" into a system that can be studied.
• Signifier / signified（能指 / 所指）：signifier = the sound / written form (the form of
  the sign); signified = the concept it stands for (the content of the sign). Only the two
  together make a "sign".
• Arbitrariness（任意性）："why is a dog called 'dog'?" — there is no reason, it is
  convention. The fact that the same thing has completely different names in different
  languages is the evidence.
• Langue / parole（语言 / 言语）：langue = the rule system shared by a society (abstract);
  parole = the actual utterances an individual produces (concrete). Saussure argued for
  studying the former.
• Synchronic / diachronic（共时 / 历时）：synchronic = cutting across one point in time to
  look at the system; diachronic = looking lengthwise at how it evolved. Saussure argued
  for studying the synchronic first.
• Syntagmatic / paradigmatic（组合 / 聚合）：the syntagmatic relation = "what goes next to
  what" inside a sentence (in "I eat rice", the relation between "eat" and "rice"); the
  paradigmatic relation = the set of words that could be substituted into the same slot
  ("eat / make / buy" can all go in that slot). The two axes crossing is what the structure of
  language is.
• Phonetics / phonology（语音学 / 音位学）：phonetics studies "what sounds a human can
  produce / hear" (the physical level); phonology studies "which sounds do the work of
  distinguishing meaning in a given language" (the functional level).
• Phone / phoneme（音素 / 音位）：a phone = an actually produced sound (with countless
  subtle variations); a phoneme = the smallest unit that can distinguish meaning in a given
  language. Example： in Chinese "b/p" are two phonemes ("爸/怕" differ), whereas in English
  aspirated and unaspirated p are variants of the same phoneme.
• Minimal pair（最小对立对）：two words that differ in just one sound and mean different
  things.
  Example： Chinese "南 nán / 兰 lán". Finding one proves that these two sounds are different
  phonemes.
• Complementary distribution（互补分布）：two sounds never appear in the same position —
  then they are very likely variants of the same phoneme. This is another way to decide
  what counts as a phoneme.
• International Phonetic Alphabet, IPA（国际音标）：a single set of symbols for writing down
  the sounds of every language in the world uniformly. One sound, one symbol, undisturbed by
  spelling (English spelling and pronunciation do not match up; IPA never has that problem).
• Initial / final / tone（声母 / 韵母 / 声调）：the three components of a Chinese syllable. The
  initial = the opening consonant; the final = the rest of the syllable; the tone =
  the rise and fall of pitch (Chinese "妈麻马骂" are distinguished by it alone).
• Morpheme / word / phrase / sentence（语素 / 词 / 短语 / 句子）：the levels of linguistic
  units, from small to large. A morpheme is the smallest "sound-meaning pairing" (e.g. "们"),
  words are built out of morphemes, and these are assembled upward into phrases and sentences.
• Syntax tree / immediate constituent analysis（句法树 / 直接成分分析）：splitting a
  sentence apart layer by layer and drawing it as a tree, showing "which words join up into a
  block first". An ambiguous sentence is ambiguous because it has two ways of being split.
• Hockett's design features（Hockett 的设计特征）：the string of features Hockett listed
  that distinguish human language from animal communication, such as productivity (being
  able to make brand-new sentences never said before), displacement (being able to talk about
  things not present), arbitrariness, duality of patterning (a sound level plus a meaning
  level, two tiers). The feature that best separates humans from animals is "productivity".

2 Semester 2

2.1 History of Chinese and Foreign Education

• Teaching according to aptitude（因材施教）：adjusting the teaching method to each
  person's characteristics. The source is "闻斯行诸" in the Analects — Confucius told the
  retiring Ran You "go and do it", and told the bold Zilu "not so fast", two opposite answers to
  the same question.
  ⚠️ Often confused with：it is not "giving different content", it is "nudging everyone
  towards the middle path": press down on those who overshoot, give a push to those who fall short.
• Education for all without discrimination（有教无类）：no distinction of rank or status,
  anyone can be taught. Before Confucius, education was an aristocratic privilege.
• The Socratic method / maieutics（产婆术）：Socrates' method of questioning — keep asking
  until you discover the contradictions yourself and reach the conclusion.
  Knowledge is not poured in, it is drawn out (a cross-cultural echo of Confucius' "I do not
  enlighten until the student is struggling to understand").
• The imperial examination system（科举制）：the standardised selection examination that
  began in the Sui-Tang period, lasted about 1300 years, and was abolished in 1905.
  Two-sided: it promoted social mobility, but it also produced teaching-to-the-test (an
  imperial edict itself admitted that "the examinations breed empty verbiage").
• Class-based teaching system（班级授课制）：Comenius' proposal — grouping students into
  classes by age and teaching them at a uniform pace.
  The benefit is that it lets education scale up; the cost is that it turns education into
  "standardised production".
• Five educations in parallel（五育并举）：Cai Yuanpei's 1912 educational programme —
  military-citizen (physical), utilitarian (intellectual), civic-moral (moral), world-view, and
  aesthetic education.
  ⚠️ Often confused with：the five educations in parallel is 1912 (when he was Minister of
  Education); the Peking University reforms are 1917 (when he was president). Do not merge the
  two into one year.
• Herbart's four stages of instruction（赫尔巴特教学四阶段）：clarity → association →
  system → method. It represents the traditional education of teacher-centred,
  textbook-centred teaching.
• Dewey's five steps of thinking（杜威思维五步）：situation → problem → hypothesis →
  reasoning → testing. It represents the progressive education of child-centred, learning by
  doing. ⚠️ It is the same structure as Simon's "problem space" (philosophy came first,
  computation later).
• School systems（学制：壬寅 / 癸卯 / 壬戌）：the modern Chinese school-system reforms. The 壬戌
  system (1922) "6-3-3" is still in use today (6 years of primary, 3 of junior high, 3 of
  senior high).

2.2 Educational Psychology

• Behaviorism（行为主义）：learning = the building of stimulus-response connections,
  looking only at observable behaviour. Reinforcement is its core mechanism. ⚠️ Its
  collapse point is Chomsky: children can produce brand-new sentences they have never heard,
  which shows language is not "a pile of habits" but a rule system.
• Cognitivism（认知主义）：learning = a change in internal cognitive structure (not just
  a change in behaviour). It focuses on memory, comprehension and processing.
• Constructivism（建构主义）：knowledge is not transmitted, it is constructed by the learner
  themselves. The representative figure is Vygotsky (ZPD). The teacher "helps construct", he
  does not "pour in".
• Humanism（人本主义）：a focus on the whole person, self-actualisation and emotional
  needs (Maslow, Rogers). Learning only works if it meets people's emotional needs.
• Meaningful learning（有意义学习）：building a substantive connection between new
  knowledge and existing knowledge (Ausubel).
  ⚠️ Often confused with：its opposite is "rote learning". Memorising a phone number = rote;
  understanding the relation between the Pythagorean theorem and the distance formula = meaningful.
• Advance organizer（先行组织者）：before teaching new content, first give a superordinate,
  general framework to help students "hang" the new material on.
  ⚠️ The effect is in fact small (meta-analysis d≈0.21), so do not treat it as a magic technique.
• Schema（图式）：the knowledge structure you already have in your head, through which new
  information is understood. The difference between experts and novices often lies in the
  "schema", not in "memory power".
• Overjustification effect（过度理由效应）：see "General Psychology".
  In education it means: do not hand out rewards indiscriminately to children who already love
  learning.
• Observational learning / vicarious reinforcement（观察学习 / 替代强化）：watching others
  do something and seeing what consequences they get, and thereby learning it yourself
  (Bandura). Vicarious reinforcement = seeing someone else rewarded for a behaviour makes you
  more inclined to do it too.
• Learning styles（学习风格）：the claim that "people are visual / auditory / kinaesthetic
  learners, and teaching to the style is more effective".
  ⚠️ This is a popular claim that has been falsified: after review, Pashler et al. (2008)
  found no qualifying evidence to support it.
  But mind the wording: it is not "proven useless", it is "no qualifying evidence supports
  it".
• Crossover interaction（交叉设计）：to prove "teaching to learning style works", two things
  must hold at once: visual learners do better when taught visually and auditory learners do
  better when taught auditorily. Most studies only did the first half, so they prove nothing.
• Transfer（迁移）："can what you learnt in one situation be used in another situation".
  The most dangerous case is "only being able to use it in an exam" (→ the street mathematics
  in "Introduction to the Learning Sciences").

2.3 Modern Chinese

• The definition of Putonghua（普通话的定义）：Beijing pronunciation as the standard
  pronunciation, the northern dialects as the base dialects, and exemplary modern vernacular
  writings as the grammatical norm. The three standards govern phonology, vocabulary and
  grammar respectively.
• Tone sandhi（变调）：tones change when syllables are read in sequence. Native speakers do
  it every day, but cannot state the rule.
• Third-tone sandhi（三声变调）：when two third tones are read together, the first becomes a
  second tone ("你好" nǐ hǎo → ní hǎo).
• Sandhi of "一" and "不"（"一"和"不"的变调）：the tone of the following syllable decides
  which tone they are read with.
• Erhua / rhotacisation（儿化）：adding a retroflex "儿" ending — it is not just adding a
  sound, it often also carries a flavour of "small, cute, casual".
• Neutral tone（轻声）：a tone read short and light (like the second "妈" in "妈妈"). It is
  not a "fifth tone", it is the weakening of a syllable.
• Interlanguage（中介语）：Selinker's proposal — a "middle language" that the learner
  develops on their own, neither the mother tongue nor the target language, but with its own
  set of rules (it is not "saying things at random").
• Error analysis（偏误分析）：analysing a learner's mistakes to work backwards to the rules
  in their head. The key point: errors are not random, they are systematic — the same rule
  makes them commit the same kind of error.
• Fossilization（化石化）：a certain error in the interlanguage becomes permanently fixed,
  and no amount of study changes it.
• Mother-tongue transfer (positive / negative)（母语迁移（正迁移 / 负迁移））：the rules of
  the mother tongue get carried over into the target language. Same rule → positive transfer
  (helps); different rule → negative transfer (hurts).
  Example： English marks past time with the past tense, so the learner produces "*我昨天不去学校".
• The 把-construction / the 被-construction（把字句 / 被字句）：two sentence patterns peculiar
  to Chinese. The 把-construction = "张三把花瓶打破了" (emphasising what is done to the
  patient); the 被-construction = "花瓶被张三打破了" (emphasising the patient). They are
  great material for syntax experiments, because the structure is unusual and they carry a
  flavour of native intuition.
• "的 / 地 / 得"：all three are read de, but their grammatical functions differ —
  的 + noun (我的书), 地 + verb (慢慢地走), 得 after a verb introducing a complement
  (跑得快).
• Fundamental frequency (F0)（基频）：the fundamental frequency of a sound determines its
  pitch, and a tone is just the curve of F0 over time. The quantity most often measured in
  speech analysis.
• Formant（共振峰）：the resonance peaks of a sound, which determine the timbre of a
  vowel (the difference between a and i lies exactly in the formants).

2.4 Introduction to Cognitive Science

• Cognitive science（认知科学）：an interdisciplinary field that brings philosophy,
  psychology, linguistics, computer science, neuroscience and anthropology together to study what
  the mind is. Core question: can thought be computed?
• Cognitive revolution（认知革命）：the shift in the 1950s from behaviourism towards
  studying "internal mental processes". Behaviourism studied only observable behaviour; the
  cognitive revolution began studying the invisible processing.
• 1956（1956 年）：the "miracle year" of cognitive science — the Dartmouth conference
  (AI), Chomsky's linguistics, Miller's 7±2, and Newell & Simon's Logic Theorist all
  appeared in this one year.
  ⚠️ "The founding year of cognitive science" is a later construction: those events really did
  happen in 1956, but "founding year" is a retrospective label.
• Physical Symbol System Hypothesis（物理符号系统假设）：proposed by Newell and Simon —
  "a system that can manipulate symbols has the necessary and sufficient conditions for
  producing intelligence."
  This is the core belief of symbolic AI and early cognitive science. It was later
  challenged by connectionism.
• Representation（表征）：the thing in your head that "stands for" something in the outside
  world. For example, the image / concept in your mind when you think of "apple" is a
  representation. One of the core concepts of cognitive science.
• Computation（计算）：transforming symbols according to rules. The basic assumption of
  cognitive science is "thought = a kind of computation" (which is also why it can talk to
  computer science).
• Marr's three levels（Marr 的三层次）：understanding a cognitive system requires asking at
  three levels — ① the computational level (what problem is being solved, what is the goal) →
  ② the algorithmic level (what steps implement it) → ③ the implementation level (what
  hardware it runs on).
  Value: it makes clear that "neural activity" and "cognitive function" are not at the same
  level — do not use one level to explain the other.
• The four research programmes（四大研究纲领）：the four camps within cognitive science on "how
  to study the mind" (symbolism, connectionism, embodied cognition, dynamical systems, etc.).
  They compete with one another, and each explains different problems.
• Embodied cognition（具身认知）：cognition is not just a matter of the brain; the body and
  sensorimotor experience take part in it too.
  Example： the "hot coffee experiment" was once taken to support it — people holding hot coffee
  were more willing to call someone "warm".
  ⚠️ But this experiment later failed to replicate (several times after 2014); it is a good
  case of science correcting itself.
• Chinese Room（中文屋论证）：Searle's thought experiment — you are in a room, and using a
  rule book you match Chinese questions to Chinese answers and pass them out. The people outside
  think you understand Chinese, when in fact you do not understand a single character.
  It is used to refute "being able to process symbols correctly = really understanding".
• Easy / hard problem of consciousness（容易问题 / 困难问题）：Chalmers' distinction — the
  easy problem = how the brain does information processing (in principle explicable); the
  hard problem = why these processes come with "subjective experience" (why pain "feels
  like" pain).
  Nobody yet knows how to study the latter.
• Functionalism（功能主义）：what matters is not "what material implements it" but "what
  function is implemented". So "a mind implemented in silicon also counts as a mind" is
  logically possible — this is the philosophical basis of AI having unlimited minds.

3 Semester 3

3.1 Principles of Artificial Intelligence

• Turing test（图灵测试）：holding a conversation through a screen, if you cannot tell
  whether the other side is a person or a machine, then it counts as intelligent.
  Turing proposed it in 1950, turning the philosophical question "what is intelligence" into an
  operational behavioural question.
  ⚠️ Often confused with：it tests "how much it resembles a human", not "whether it can
  think".
• Agent（智能体）：something that perceives its environment through sensors and acts on that
  environment through actuators.
  The definition is very broad: people, thermostats, chess programs, LLM assistants all qualify.
  The key is not "whether it resembles a human" but "whether it can make decisions from what it
  perceives".
• The five elements of a rational agent（理性 agent 的五要素）：performance measure (what
  counts as good), percepts (what it can see), actions (what it can do), environment
  (what world it is in), utility function (how good each outcome is).
  ⚠️ Rational ≠ omniscient: rationality is "making the best choice given the information you
  have".
• The five elements of a search problem（搜索问题的五要素）：initial state, operators (the
  actions it can take), goal test, path cost, state space. Example： the eight-puzzle.
• BFS / DFS：breadth-first = spreading out layer by layer (guarantees the shortest path,
  but eats memory); depth-first = going down one path to the bitter end (saves memory, but
  does not guarantee optimality).
• A* algorithm（A* 算法）：BFS + "estimate how much further there is to go". It uses a
  heuristic function h(n) to guide the direction of search, and as long as h never
  overestimates the true distance (this is called "admissible"), A* is guaranteed to be optimal.
• Heuristic（启发式函数）：an estimate of "how far there is left to the goal".
  Example： in the eight-puzzle, count the "number of misplaced tiles" or the "Manhattan distance".
• Adversarial search / minimax / alpha-beta pruning（对抗搜索 / 极小极大 / Alpha-Beta 剪枝）：
  search where two players take turns deciding. Minimax = I pick the maximum payoff while
  the opponent picks the move that minimises yours (assuming the opponent plays optimally);
  alpha-beta = cutting off branches that clearly cannot matter, to save time.
• Knowledge representation（知识表示）：writing the world "in a form a machine can reason
  with". For example propositional logic, first-order logic (which can express "all A are B").
• Deduction / induction / abduction（演绎 / 归纳 / 溯因）：the three kinds of reasoning.
  Deduction = from the general to the particular (truth-preserving); induction = from the
  particular to the general (can go wrong); abduction = guessing the cause from the result
  (often used in medical diagnosis).
• Supervised / unsupervised learning（监督 / 无监督学习）：supervised = data with the
  answers attached (labels); unsupervised = only data, no answers (finding structure by
  itself, e.g. clustering).
• Overfitting（过拟合）：the model has "memorised" the training data and falls apart on new
  data. This is the most central pitfall in machine learning. Countermeasures: hold out a
  test set, regularisation, more data.
• Generalization（泛化）：performing well on new data you have never seen.
  What we really want is generalisation, not "a high score on the training set".
• The three AI waves（AI 三次浪潮）：roughly 1956–1974 (symbolism) / 1980–1987 (expert
  systems) / 2011 to now (deep learning). Between them sit two "AI winters". ⚠️ The exact
  year boundaries differ between sources.

3.2 Introduction to the Learning Sciences

• Learning sciences（学习科学）：the interdisciplinary field that studies how people learn
  and improves teaching accordingly. It pulls psychology, education, computer science and brain
  science together.
• Situated（情境化）：knowledge is learnt in a concrete situation, and is thereby "tied" to
  that situation. This is the root cause of "what I learnt cannot be used".
• Transfer（迁移）：using what you learnt in one place somewhere else.
  The street-mathematics example is the most striking: the same children, 98% correct doing
  arithmetic on the street, only 37% correct when it is turned into a written sum —
  they are not unable to calculate, they are unable to "transfer".
• Cognitive load（认知负荷）：working memory can only handle a limited amount of information
  at once; exceed it and nothing goes in. Three kinds: intrinsic (the content itself is
  hard), extraneous (waste caused by bad teaching / presentation), germane (used to build
  schemas — good load).
• Worked-example effect（范例效应）：novices learn better from studying a full solution than
  from jumping straight into solving problems. Because solving problems fills up working memory
  and you never learn the structure. ⚠️ It reverses for experts (see "expertise reversal
  effect").
• Split-attention effect（分散注意效应）：with the diagram and the text in separate places,
  you have to look back and forth, wasting load for nothing. Countermeasure: integrate them.
• Retrieval practice / testing effect（提取练习 / 测试效应）：
  "testing yourself" makes you remember better than "reading it again". Because retrieval
  itself strengthens memory.
• Metacognition（元认知）：the ability to "know whether you know".
  Its typical failure: thinking you have learnt it (it feels fluent on re-reading) when you
  have not.
• Expertise reversal effect（专家逆转效应）：a method that works for novices can actually
  harm experts.
  Example： giving an expert a detailed worked example wastes their mental effort. So the
  question "which method is best" is incomplete; you have to ask "best for whom".

3.3 Language Pedagogy

• Grammar-translation / direct method / audiolingual / communicative (CLT) / task-based (TBLT) / post-method（语法翻译法 / 直接法 / 听说法 / 交际法（CLT）/ 任务型教学（TBLT）/ 后方法）：
  six generations of teaching method evolution, each generation fixing the flaws of the last:
  grammar-translation (can read but cannot speak) → direct method (target language only) →
  audiolingual (pattern drills, but its theory was overturned) → communicative (language is for
  communication) → task-based (organised around tasks) → post-method (there is no universal
  method; it depends on the situation).
• Communicative competence（交际能力）："being good at exercises" is not the same as "being
  able to use it" (proposed by Hymes). It has four components: grammatical competence
  (saying it correctly), sociolinguistic competence (saying it appropriately), discourse
  competence (saying it coherently), strategic competence (how to talk your way around it
  when you cannot say it).
• Explicit / implicit instruction（显性 / 隐性教学）：explicit = stating the rules out
  loud; implicit = not stating the rules, relying on lots of input.
  Meta-analyses find explicit clearly more effective (d≈1.13 vs 0.54), ⚠️ but this may be a
  bias produced by measuring with fill-in-the-blank tests.
• Focus on Form / Focus on Forms（FonF / FonFS）：FonF = focusing on language form
  incidentally inside a communicative task; FonFS = teaching grammar in a separate
  course. Statistically there is no significant difference between the two in effect.
• Recast / prompt（重铸 / 提示）：two ways of correcting errors. A recast = the teacher
  says what you got wrong again, in the correct form (without interrupting); a prompt =
  forcing you to fix it yourself ("what did you just say?").
  Recasts are the most common, but students often do not realise it is a correction; prompts are
  more explicit and usually more effective.
• Effect size d（效应量 d）：a standardised index of "how big the effect is".
  Rough standards: 0.2 small, 0.5 medium, 0.8 large. It is unrelated to sample size, whereas
  the p-value is related — so looking only at p-values lets a large sample fool you.

3.4 Cognitive Neuroscience of Language

• Event-related potential, ERP（事件相关电位）：extracting "the response to a particular
  stimulus" out of the EEG.
  The method: repeat the same stimulus many times and average the EEG — the random background
  noise cancels out and what is left is the response.
• N400：a negative wave appearing about 400 ms after the stimulus, triggered by
  semantic violations ("I drank a cup of table"). It is the classic index of "semantic
  processing".
• P600 / SPS：a positive wave appearing about 600 ms after the stimulus, triggered by
  syntactic violations (Osterhout & Holcomb 1992).
  ⚠️ It was originally thought to be "syntax-specific", but later work found that semantic,
  musical and mathematical violations also elicit it — so do not write "syntax-specific".
• Double dissociation（双分离）：A is broken but B is fine, and in another person B is
  broken but A is fine. This is strong evidence that these two things are handled by
  different systems. N400/P600 is one such example
  (⚠️ but it later turned out not to be that clean).
• Broca's area / Wernicke's area（布洛卡区 / 韦尼克区）：Broca's area = left frontal
  lobe, handles speech production; Wernicke's area = left temporal lobe, handles
  language comprehension. ⚠️ This is a 19th-century simplification; the modern view is a
  distributed network.
• Broca's / Wernicke's / conduction aphasia（布洛卡失语 / 韦尼克失语 / 传导性失语）：
  Broca's aphasia = effortful, telegraphic speech, comprehension relatively preserved;
  Wernicke's aphasia = fluent but meaningless speech (word salad), comprehension severely
  impaired; conduction aphasia = can understand but cannot say it (poor repetition), with
  damage to the arcuate fasciculus.
• Arcuate fasciculus（弓状束）：the white-matter fibre bundle connecting Broca's area and
  Wernicke's area. Cut it → conduction aphasia.
• Lateralization（偏侧化）：which hemisphere language function is mainly in. People often
  say "95% of people have language in the left brain" — ⚠️ but that figure comes from patients
  with brain damage or who have had surgery (the Wada test), not from healthy people; in healthy
  people measured with fTCD the left-right difference is not that extreme. When citing it, say
  where the data comes from.
• Wada test / fTCD（Wada 测验 / fTCD）：Wada = injecting anaesthetic into one carotid
  artery and seeing which side's language stops (used before surgery; this is patient data);
  fTCD = measuring blood flow with transcranial Doppler (usable on healthy people, hence
  the different conclusions).
• Dual-stream model（双流模型）：Hickok & Poeppel (2007) — the ventral stream handles
  "understanding" (broadly bilateral); the dorsal stream handles the mapping from
  "hearing → saying it" (strongly left-lateralised). It parallels the "what pathway /
  where pathway" of the visual system.

4 Semester 4

4.1 Core Experimental Methodology

• The four kinds of validity（效度的四种）：internal validity = does the causal inference
  in this experiment hold; external validity = can it generalise to other situations;
  construct validity = are you really measuring what you intend to measure; statistical
  validity = is the statistical conclusion reliable.
  ⚠️ The classic conflict: experiments with high internal validity (tightly controlled)
  often have low external validity (detached from real situations).
• Statistical power（统计功效）：the probability that you detect an effect when there really
  is one. It is determined by effect size, sample size, α — and of these, only sample
  size is freely under your control. The usual target: 0.80. ⚠️ A psychology experiment with
  20 people often has power as low as 0.2: it cannot detect a real effect, yet easily "happens"
  to detect a fake one.
• Replication crisis（可复现性危机）：a large number of published psychology results cannot
  be reproduced. OSC (2015) replicated 100 studies: 97% of the original studies were
  significant, only 36% of the replications were.
• Publication bias（发表偏倚）：only "significant" results easily get published, so the
  effects in the literature are systematically overestimated.
• p-hacking：trying one analysis after another until p < 0.05.
  It is not fraud, but it likewise lets false positives into the literature.
• HARKing：making up your hypothesis after seeing the results (Hypothesizing After the
  Results are Known). The consequence: dressing "exploration" up as "confirmation".
• Preregistration（预注册）：publicly registering "which hypothesis I will test and how I
  will analyse it" before running the experiment. It solves HARKing and p-hacking.
  ⚠️ What it cannot solve: data fabrication, a broken instrument, too small a sample, a wrong
  theory.
• Within-subjects / between-subjects design（被试内 / 被试间设计）：within-subjects = the
  same people do all the conditions (saves people, high power, but has order effects);
  between-subjects = different people do different conditions (no order effects, but needs
  more people).
• Counterbalancing / Latin square（抵消平衡 / 拉丁方）：devices for removing order
  effects. Counterbalancing = having half the participants do each order; a Latin square = a
  systematic balancing scheme for many conditions.

4.2 Philosophy of Education

• Philosophy of education（教育哲学）：the study of "how education ought to be". ⚠️ It is
  fundamentally different from your other courses — other courses ask "what is" (is), it
  asks "how it ought to be" (ought), and the way you answer is not to find data but to argue
  and rebut.
• Epistemology（认识论）：the study of "where knowledge comes from and what counts as
  knowledge". It determines "what should be taught" — however you believe knowledge comes
  about, that is how you will teach.
• Rationalism / empiricism（理性主义 / 经验主义）：rationalism = knowledge comes from
  rational reasoning (Descartes, Plato); empiricism = knowledge comes from sensory
  experience (Locke, Hume).
• Tabula rasa（白板说）：Locke — the mind begins as a blank slate, and all knowledge is
  written onto it afterwards. Corollary: environment and education decide almost everything.
• Is knowledge "discovered" or "constructed"?（知识是"被发现"还是"被建构"）：the core debate.
  Discovered = knowledge exists objectively, waiting for someone to find it; constructed =
  knowledge is actively made by people. This directly decides whether teaching is "transmission"
  or "guiding construction".
• Views of human nature (good / evil)（人性论（性善 / 性恶））：what you think people are
  like by nature decides how you teach. Rousseau held that human nature is good (so nature must
  be protected); Xunzi held that it is evil (so it must be civilised).
• Veil of ignorance（无知之幕）：Rawls' thought experiment — when designing the rules of
  society, suppose you do not know who you will be born as (rich / poor, clever / dull).
  Only rules designed this way are fair.
• Difference principle（差异原则）：Rawls — inequality is permitted, but only when it is
  "also to the benefit of the worst-off".
• Critical pedagogy / banking education（批判教育学 / 银行储蓄式教育）：Freire — treating
  students as a piggy bank into which the teacher "deposits" knowledge is oppressive education.
  Real education is dialogue; it is making people wake up.
• Pragmatism（实用主义）：Dewey — the value of a theory lies in whether it can solve real
  problems. "Education is life" and "learning by doing" both come from here.

4.3 Core Psycholinguistics

• Psycholinguistics（心理语言学）：the study of "how people process language in real
  time" — comprehension and production on the millisecond scale. Linguistics asks "what is
  the structure"; psycholinguistics asks "how do people run that structure".
• Semantic priming（语义启动）：seeing one word first makes a related word easier to
  recognise.
  Example： after "doctor", "nurse" is recognised faster than "bread". It measures the "spreading
  activation" in the semantic network.
• Spreading activation（激活扩散）：when a concept is activated, it passes along the network
  to related concepts.
• Word-frequency / concreteness / word-length effects（词频效应 / 具体性效应 / 词长效应）：
  the three most robust effects in word recognition — high-frequency words are faster, concrete
  words are faster, long words are slower (but common long words weaken this).
  ⚠️ The exact millisecond difference varies a lot with materials and task; if you cite it, give
  a range and the specific study, not a single-point number.
• Garden path sentence（花园路径句）：the start lures you into the wrong structure, and by
  the end you have to go back and re-parse.
  Example： "Since Jay always jogs a mile seems like a short distance to him."
• Disambiguation region（消歧区）：the spot where you realise "I read that wrong before".
  Here reading time goes up and regressions increase.
• Regression（回视）：the eyes going back. It is a signal that "comprehension has gone
  wrong".
• Syntax-first / interactive models（句法优先 / 交互式模型）：syntax-first = parse by
  syntax first, semantics arrives later (Frazier & Rayner); interactive = semantics and
  context influence syntactic parsing immediately (evidence from the visual world paradigm
  supports this camp).
• Eye-tracking measures (first fixation / gaze duration / total fixation time)（眼动指标（首次注视 / 凝视时间 / 总注视时间））：
  first fixation = how long the very first look rests; gaze duration = the sum of all
  fixations before the first exit; total fixation time = everything including regressions.
  How to read it: first fixation normal but total fixation longer → the word itself is not
  hard, what is hard is "integrating it into the context".
• Statistical learning（统计学习）：infants can "statistically" extract word boundaries from
  continuous speech (Saffran 1996). This shows language learning needs no dedicated mechanism
  — general learning ability is enough — a big point of controversy.
• Surprisal（意外度）：a quantitative index of "how surprising this word is in the current
  context". The more surprising, the slower you read. Language models can also compute
  surprisal, and it predicts human brain responses — ⚠️ but "correlation ≠ the same mechanism".

4.4 Introduction to Second Language Acquisition

• Critical period hypothesis（关键期假说）：there is a window for language learning, after
  which it gets hard. Johnson & Newport (1989): the earlier you arrive in the US, the better
  your English grammar scores, with a clear decline after puberty.
  ⚠️ Criticism: it may not be that "the window closed" but that "input drops off / motivation
  falls".
• Interlanguage / error analysis / fossilization / mother-tongue transfer（中介语 / 偏误分析 / 化石化 / 母语迁移）：
  see "Modern Chinese".
• Krashen's Monitor Model（Krashen 监控模型）：five hypotheses, the most famous being
  ① the acquisition / learning hypothesis ("acquisition" is natural, "learning" is conscious,
  and the two do not convert into each other); ② the input hypothesis i+1 (input must be
  slightly above your current level); ③ the affective filter (nervousness blocks input).
• Output Hypothesis（输出假说）：Swain — input alone is not enough; "forcing yourself to
  speak / write" is what pushes grammar forward. Because when you produce output you notice
  that you cannot say it, and this "getting stuck" is the engine of learning.
• Interaction Hypothesis（互动假说）：Long — "negotiation of meaning" is the key to
  acquisition. That is: the process of asking back and forth, restating, confirming when you
  do not understand has the strongest facilitative effect.
• Negotiation of meaning（意义协商）：the process in which, when communication hits an
  obstacle, both sides adjust what they say until they are understood.
• Universal Grammar / emergentism（普遍语法 / 浮现论）：fundamentally opposed. Universal
  grammar = people are born with a language acquisition device; emergentism = linguistic
  structure "emerges" from general cognitive abilities + massive usage, and no dedicated
  device is needed.
• Frequency effects（频次效应）：the more often you see it, the faster you process it. This
  is important evidence for emergentism.

4.5 Core Cognitive Psychology

• Subtractive method（减法法）：Donders (1868) — subtracting the reaction times of two
  tasks to isolate the time of "purely one mental process".
  Example： measure a, and measure b+c; the subtraction gives you the time for c. ⚠️ The key
  assumption is "pure insertion".
• Pure insertion（纯插入假设）：assuming that adding one process does not change the other
  processes. ⚠️ This is exactly what Wundt criticised — in reality it often fails to hold,
  and this is the biggest soft spot of the subtractive method.
• Additive factors method（加因素法）：Sternberg (1969) — instead of subtracting, look at
  which factors "add" and which "interact". Adding → they are serial; interacting → they are in
  the same stage.
• Exhaustive search（穷尽扫描）：Sternberg's finding — when searching short-term memory, the
  slope of reaction time is the same no matter which position the target is in, showing that
  you scan everything before answering.
• The components of working memory（工作记忆的成分）：the central executive (allocating
  resources), the phonological loop (handling sound / language), the visuospatial sketchpad
  (visual space). Evidence: word-length effect, articulatory suppression, dual-task interference.
• DRM paradigm（DRM 范式）：give a list of highly related words (all to do with "sleep"),
  but "sleep" itself never appears. Participants nonetheless often "remember" hearing it — the
  false recall rate reaches 40% (experiment 1) to 55% (experiment 2).
  It shows that memory "auto-completes".
• Conjunction fallacy（合取谬误）：thinking "both conditions hold" is more likely than "one
  of them holds".
  The Linda problem: most people think "Linda is a bank teller and a feminist" is more likely than
  "Linda is a bank teller" — which is logically impossible (a conjunction cannot be more likely
  than a single conjunct).
• Availability heuristic（可得性启发式）："if it comes to mind easily, I assume it happens
  more often."
  Example： lots of plane-crash news → thinking flying is more dangerous than driving (the opposite
  is true).

4.6 Core Cognitive Neuroscience

• fMRI：using "changes in blood-oxygen level" to see brain activity indirectly. High
  spatial resolution (millimetre scale), poor temporal resolution (second scale).
• BOLD signal（BOLD 信号）：what fMRI measures — the blood-oxygen-level-dependent signal.
  ⚠️ It is not "measuring neurons firing": neural activity → oxygen consumption → an
  over-compensating rush of blood → more oxygenated blood locally.
• Hemodynamic response function, HRF（血流动力学反应）：the BOLD curve caused by a brief
  stimulus. Three phases: initial dip (1–2 s, not reliable) → peak / overshoot (canonical
  value 5 s, commonly 4–6 s) → post-stimulus undershoot (dipping below baseline, lasting
  ten-odd seconds). The whole curve is about 30 s.
• ERP naming convention（ERP 命名规则）：N/P + latency. N = negative (negative wave),
  P = positive (positive wave), and the number is roughly how many milliseconds to peak.
  Example： N400 = a negative wave at around 400 ms.
• The case of H.M.（H.M. 病例）：to treat epilepsy, his bilateral medial temporal lobe
  (including the hippocampus) was removed.
  Result: short-term memory normal, old memories preserved, but he could never form new
  long-term memories again (anterograde amnesia), yet he could still learn new skills (though
  he did not remember practising).
  It proves three things: memory is not "one thing"; the hippocampus is necessary for new
  memories; "remembering" and "being able to do" are two systems.
• Population coding / grandmother cell（群体编码 / 祖母细胞）：is a concept represented by a
  large population of neurons together (population coding), or is there a single "grandmother
  cell" dedicated to it? Population coding is the mainstream, but cells that respond very
  strongly to specific concepts do exist.
  ⚠️ "Grandmother cell" is an exaggerated nickname.
• Reverse inference（反向推理）："this brain region lit up → so it handles this function".
  ⚠️ This is the commonest misuse of fMRI — a region being involved ≠ that region is dedicated
  to that function (it may do many things).

4.7 Introduction to Cognitive Linguistics

• Cognitive linguistics（认知语言学）：holds that language is not an independent module; it
  is just part of general cognitive ability.
  ⚠️ This is the root of its opposition to Chomsky: Chomsky asks "why is language special",
  cognitive linguistics asks "why is language not that special".
• Conceptual metaphor theory（概念隐喻理论）：Lakoff & Johnson — metaphor is not a rhetorical
  device, it is the very way we think. We understand an abstract domain through a concrete
  domain ("argument is war", "time is money").
  It is not an individual expression but something that systematically permeates the whole
  language.
• Source / target domain（源域 / 目标域）：the source domain = the concrete concept you
  borrow (money); the target domain = the abstract concept you want to talk about (time).
• Image schema（意象图式）：a pre-linguistic basic structure abstracted from repeated
  interaction between the body and the world.
  Such as container (the body has an inside and an outside), path (movement), up-down
  (gravity). It is the core argument of "embodied cognition".
• Cognitive grammar (Langacker)（认知语法）：grammar is not "rules", it is the symbolisation
  of "imagery". Core concepts: figure-ground, specificity (animal → dog → golden
  retriever), perspective.
  In one sentence: the same scene can be "construed" in different ways, and whoever you
  foreground becomes the subject.
• Construction Grammar（构式语法）：Goldberg — a "construction" itself carries meaning; it
  is not just a container you put words into. Example： in "我烤了他一个蛋糕" there is no word
  "给", yet the meaning "give" is there — that meaning comes from the "double-object
  construction" itself.
• Prototype theory（原型理论）：Rosch — category boundaries are fuzzy and members have
  degrees of "typicality".
  A robin is more "bird-like" than a penguin. It overturned the classical theory that
  categories are decided by necessary and sufficient conditions.
• Typicality effect（典型性效应）：the more typical a member, the faster you judge whether it
  belongs to the category.
• Family resemblance（家族相似性）：proposed by Wittgenstein and supported experimentally by
  Rosch — members of the same category need not share one common feature; rather, "there is
  partial overlap between any pair" (like the way members of a family look alike).
• Iconicity（象似性）：linguistic form "resembles" its meaning (an important revision of
  Saussure's "arbitrariness").
  Three kinds: sequential iconicity (word order = order of events), distance iconicity
  (closer relation → closer words), quantity iconicity (longer form → larger quantity).

5 Semesters 4–5

5.1 Educational Assessment (4 credits — the basic course with the most credits for you)

• Measurement / assessment / test（测量 / 评价 / 测验）：measurement = objective
  assignment of values ("scored 85"); assessment = making a value judgement ("85 counts as
  excellent"); test = the instrument of measurement.
  ⚠️ Measurement ≠ assessment: measurement is objective, assessment involves standards and
  values.
• The four levels of measurement (Stevens)（四种测量水平）：nominal (mere classification,
  e.g. sex) → ordinal (ordered but unequal intervals, e.g. class rank) → interval (equal
  intervals, no absolute zero, e.g. Celsius) → ratio (has an absolute zero, e.g. reaction time,
  height).
  It decides what statistical operations you may perform: strictly speaking you cannot take an
  average of an ordinal scale.
• Reliability（信度）：how consistent repeated measurements are (how steady it is).
  Four kinds: test-retest, parallel forms, internal consistency, inter-rater.
• Cronbach's α：the most commonly used index of internal consistency — how much the items
  "agree with each other".
  ⚠️ α too low = the test is unstable; α too high (>0.95) is not necessarily good either —
  the items may be asking the same thing over and over.
  The main reading conventions (George & Mallery 2003): >0.9 excellent / >0.8 good / >0.7
  acceptable / <0.6 poor.
• Validity（效度）：did you measure it correctly. Content validity (does it cover the
  content), criterion validity (does it correlate with an external standard), construct validity
  (did you really measure that theoretical construct).
• The relation between reliability and validity（信度与效度的关系）：low reliability → validity
  must be low (if it is unstable, how could it be accurate); high reliability → validity is not
  necessarily high (very steady, but it might be steadily wrong). High validity → reliability
  must be high.
  Analogy: a ruler that is always 2 cm short — very high reliability, very low validity.
• Formative / summative assessment (proposed by Scriven)（形成性 / 终结性评价（Scriven 提出））：
  formative = done during the process, in order to improve; summative = done at the
  end, in order to judge.
  ⚠️ But in reality the weight of the summative (the exam) often overwhelms everything else —
  this is the fundamental tension in educational assessment.
• Norm-referenced / criterion-referenced（常模参照 / 标准参照）：norm-referenced =
  comparing with others (the gaokao, IQ) → the goal is to spread people out;
  criterion-referenced = comparing with a standard (driving licence, passing HSK) → the goal
  is to reach the standard.
  The philosophies behind the two reference frameworks are completely different (the former
  assumes somebody will always be eliminated).
• Classical Test Theory (CTT) / true-score model（经典测验理论（CTT）/ 真分数模型）：
  observed score = true score + error (X = T + E). It assumes the error is random with a mean
  of 0. Reliability = true-score variance / observed-score variance.
  ⚠️ The limitation of CTT: item difficulty and examinee ability cannot be estimated
  separately (this is what IRT was made to solve).
• Item difficulty P / discrimination D（难度 P / 区分度 D）：difficulty P = the proportion
  who answer correctly; discrimination D = the correct rate of the high group − that of the low
  group.
  The ideal P is 0.3–0.7 (too hard or too easy and it cannot separate people); the larger D
  the better, D<0.2 should be revised or dropped, and a negative D means the item is broken.
  ⚠️ These thresholds are textbook rules of thumb, not statistical conclusions.
• Item Response Theory (IRT)（项目反应理论）：using a mathematical model to estimate "item
  parameters" and "ability parameters" at the same time, with the two separated.
  The three parameters (3PL): a discrimination, b difficulty, c guessing.
  Advantages: ability estimates do not depend on the specific items, and equating is
  possible (converting different test forms onto the same scale).
  GRE and PISA both use it.
• Item characteristic curve (ICC)（项目特征曲线）：in IRT, the S-shaped curve of "the
  stronger the ability, the higher the probability of answering correctly".
• Large-scale assessments (PISA / TIMSS / PIRLS)（大规模测评（PISA / TIMSS / PIRLS））：
  PISA (OECD) tests 15-year-olds' reading / mathematical / scientific literacy; TIMSS
  (IEA) tests mathematics and science achievement; PIRLS (IEA) tests reading literacy.
  ⚠️ PISA's definition of "literacy" is worth noting: it does not test "what you have learnt"
  but "what you can do with knowledge" — which is transfer.

6 Semester 5

6.1 Educational Statistics

• Descriptive / inferential statistics（描述统计 / 推断统计）：descriptive statistics =
  stating clearly the data you have in hand (mean, standard deviation); inferential statistics
  = inferring the population from a sample (this is the real power of statistics, and also the
  source of error).
• Mean / median / mode（均值 / 中位数 / 众数）：the mean is sensitive to extreme values;
  the median resists extreme values (income and house prices usually use it); the mode =
  the most frequent value.
• Standard deviation（标准差）：on average, how far the data are from the mean (dispersion).
  ⚠️ Often confused with：be clear about whether the denominator is n or n−1 — the sample
  standard deviation uses n−1 (Bessel's correction), the population uses n. Get this wrong in
  your code and the result is biased.
• Standard error（标准误）：how inaccurate the estimate "the sample mean" itself is.
  ⚠️ Often confused with (extremely common): the standard deviation describes "how spread
  out the data are"; the standard error describes "how accurate your estimate of the mean is".
  The two differ by a factor of √n. When reporting, say which one it is.
• Hypothesis testing（假设检验）：first assume "there is no effect" (the null hypothesis),
  then see how rare the data are under that assumption.
• Null hypothesis (H₀)（零假设）：the hypothesis of "no effect / no difference". It is what
  we test, not what we claim.
• p-value（p 值）：the probability of getting data "as extreme as this, or more extreme" if
  the null hypothesis were true.
  ⚠️ The most misunderstood quantity of all! The p-value is not "the probability that your
  hypothesis is true", nor "the probability that the effect is false".
  It only answers "how rare the data are under the assumption of no effect".
• α (significance level)（α（显著性水平））：the bar you set in advance for "how rare counts
  as rare" (commonly 0.05). p < α is called "statistically significant" — note that
  "significant" only says "probably not a coincidence", and is not the same as "the effect is
  large" or "important".
• Type I / Type II error（第一类错误 / 第二类错误）：Type I = reporting an effect when
  there is none (false positive; α controls it); Type II = failing to detect an effect that is
  there (a miss; 1−β is the power).
• Effect size（效应量）：how big the effect is. Commonly Cohen's d (difference of two
  group means / standard deviation). Rough standards: 0.2 small, 0.5 medium, 0.8 large.
  ⚠️ It is unrelated to sample size — looking only at the p-value ignores "whether the effect
  is even worth caring about".
• Correlation (r) / regression（相关（r）/ 回归）：correlation = the direction and
  strength with which two variables vary together (−1 to +1); regression = using one variable
  to predict another.
  ⚠️ Correlation is not causation (it may be caused by a third variable, or the direction may
  be reversed).
• Coefficient of determination (R²)（决定系数）："what proportion of the variance is
  explained". When r=0.3, R²=0.09 — only 9% is explained; a correlation coefficient that does
  not look small actually has very weak explanatory power.
• The multiple comparisons problem（多重比较问题）：run many tests and a few will "happen" to
  be significant.
  Example： run 20 tests at α=0.05, and the probability of at least one false positive is about
  64%. Countermeasures: Bonferroni correction (divide α by the number of tests), FDR
  control.
• Parametric / non-parametric tests（参数 / 非参数检验）：parametric tests (t-test, ANOVA)
  require the data to roughly meet distributional assumptions; non-parametric tests (rank sum,
  sign test) do not, but have slightly lower power. If the data are badly skewed, use
  non-parametric.

6.2 Principles and Design of Agents

• Agent（智能体）：see "Principles of Artificial Intelligence". This course is about building
  one with your own hands.
• The five architectures（五种架构）：simple reflex (look-up table) → model-based
  (maintaining internal state) → goal-based (considering what must be achieved) → utility-
  based (considering "how good") → learning (improving from experience).
  The "cleverer" the agent, the richer its internal world model.
• The six dimensions of environments（环境的六个维度）：fully / partially observable,
  deterministic / stochastic, episodic / sequential, static / dynamic, discrete / continuous,
  single- / multi-agent.
  Partially observable + stochastic + dynamic + multi-agent = the hardest (the real world, the
  stock market, multiplayer games).
• Q-learning：the most basic reinforcement learning algorithm — learning "how much is it
  worth to do a certain action in a certain state" (the Q value).
  The update rule: correct the old estimate using "the reward actually received + the best value
  of the next state".
• Q value（Q 值）：the estimate of "in the long run, how many points is it worth doing this
  action in this state".
• Learning rate α / discount factor γ（学习率 α / 折扣因子 γ）：α = how much new information
  overwrites the old estimate; γ = the weight on future rewards (0.95 ≈ caring quite a lot
  about the future, 0 ≈ only caring about right now).
• ε-greedy（ε-贪婪）：with probability ε try something at random (explore), otherwise pick
  the current best (exploit).
  Gradually shrinking ε is a common practice: explore a lot early, exploit a lot later.
• Exploration vs exploitation（探索 vs 利用）："try something new" or "use the best known" —
  the most central hard problem in reinforcement learning. Not exploring and you may miss
  something better; not exploiting and you have learnt for nothing.
• Reward shaping（奖励塑造）：the reward function you write is the agent's "goal".
  ⚠️ It only optimises the function you wrote, not the thing you wanted.
• Reward hacking / specification gaming（奖励黑客（reward hacking）/ 规范博弈）：the agent
  finds a way to get "high reward but not what you wanted". Classic case: the CoastRunners boat
  AI does not race; it circles in a lagoon bumping into targets over and over to farm points —
  scoring 20% higher than humans while never finishing the race.
• Goodhart's law（Goodhart 定律）："when a measure becomes a target, it ceases to be a good
  measure."
  Reward hacking and "teaching to the test" are the same structure.

6.3 Language Teaching Assessment and Application

• Language ability (construct)（语言能力（构念））："what language ability actually is" is a
  question of defining a construct — ⚠️ what you are actually measuring decides the validity of
  this test. This is the most fundamental difficulty in language testing.
• CEFR：the Common European Framework of Reference for Languages, dividing language
  ability into six levels A1–C2, described by "can do" statements ("can order in a
  restaurant" rather than "masters 2000 words").
  This is the shift from a "knowledge orientation" to an "ability orientation". A common
  benchmark: IELTS 6.5 ≈ B2 (upper boundary).
• Washback（反拨效应）：the test influences teaching in turn.
  ⚠️ And it goes both ways: a good test guides teaching towards improvement, a bad test leads to
  "teaching whatever is tested".
• Inter-rater reliability（评分者一致性）：do different raters give consistent scores.
  Subjective items (essays, speaking) must report this, otherwise the scores are not credible.
• ICC (intraclass correlation coefficient)（ICC（组内相关系数））：the core index for
  measuring rater agreement.
  ICC(2,1) is the most commonly used form (two raters, random-effects model).
  There is no single authoritative standard for interpreting it; when citing, say which framework
  you are using (Koo & Li / Landis & Koch).
• Validity argument (Messick's framework)（效度论证（Messick 框架））：validity is not "a
  kind of validity" but "an overall argument" — all the evidence together, does it support or
  not support this interpretation and use.
  Messick in particular proposed "consequential validity": the social consequences the
  assessment brings are also part of validity.
• CLA model (Bachman & Palmer)（CLA 模型（Bachman & Palmer））：a framework that splits
  "language use" into "language ability + strategic competence + task characteristics", used to
  guide item design and validity argumentation.

7 Semester 6

7.1 Curriculum and Instruction

• Curriculum（课程）："curriculum" has several definitions — it can be "school subjects",
  "learning experiences", "teaching objectives" or "a plan". Different definitions lead to
  different teaching practices.
• Hidden curriculum（隐性课程）：what the school does not state explicitly but students
  actually learn — such as "hand your homework in on time", "obey authority", "what ranking
  means".
  ⚠️ Its influence is often greater than the explicit curriculum.
• Bloom's Taxonomy（Bloom 分类学）：layering educational objectives by cognitive complexity.
  The original (1956): knowledge → comprehension → application → analysis → synthesis →
  evaluation.
  The revised (2001): remember → understand → apply → analyse → evaluate → create.
  ⚠️ It is not just a change of nouns — "create" was moved to the top (the original had
  "evaluation" on top).
  Three domains: cognitive, affective, psychomotor.
• The three curriculum orientations（课程取向（三大））：subject-centred (organised around
  the structure of subject knowledge), child-centred (organised around students' interests and
  experience), society-centred (organised around social problems). This is a choice of
  stance, each with its own costs.
• ADDIE：an instructional design process — Analysis → Design → Development →
  Implementation → Evaluation. It is a general-purpose process framework.
• Understanding by Design / backward design（逆向设计）：Wiggins & McTighe —
  first think "how will I prove the student has learnt it", then design the teaching.
  That is, "begin with the end in mind": objectives → assessment evidence → teaching activities.

8 Appendix: concepts in the reading guide to Simon's Cognition

• Bounded rationality（有限理性）：Simon — people are not "omniscient, optimal" decision
  makers, but "good enough" ones under limited time, information and computing power
  (satisficing). This is a revision of the "rational man" assumption in economics.
• Problem space（问题空间）：seeing problem solving as "searching in a state space" —
  initial state → apply operators → intermediate states → goal state. The core of chapter 3 of
  Simon's Cognition.
• Operator（操作符）：an action that can change the state (e.g. "move the blank left").
• Heuristic search（启发式搜索）：not blind exhaustive enumeration, but using "rules of
  thumb" to choose where to go (e.g. means-ends analysis).
• Means-ends analysis（手段-目的分析）：comparing "where I am now" with "where the goal is"
  and finding a way to shrink that gap — the strategy humans use most in solving problems.
• Protocol analysis（口语记录分析）：having participants say out loud what they are thinking
  while solving a problem, then analysing those verbal records. It is a direct method for
  studying "thought processes" (and Simon's signature method).
• EPAM：an early model of learning and memory from Simon's group — simulating how people
  memorise nonsense syllables and how the "serial position effect" arises. It showed that
  "learning" itself can also be simulated by a program.
• Mental representation（心理表征）：the "way it is stored in the head" differs between
  experts and novices for the same thing.
  Chase & Simon's chess experiment: experts remember real game positions far better than novices,
  but with random positions there is no difference — showing that the expert's advantage lies in
  a "chunked representation", not in "memory power".

9 One last thing: how to study with this dictionary

Do not read this dictionary from cover to cover. That is the least efficient way to use it.

Recommended:
1. Look up any word you do not understand while reading your notes (search with Cmd+F),
read the one-sentence explanation, and go straight back to reading.
2. Each time you finish a course, go through that section's words, cover the explanations
and self-test.
3. If you find "I have seen this word three times and still do not get it" → what you are
missing is not this one sentence of explanation but the block of background knowledge it
sits in — go back and read the corresponding set of notes.

⚠️ The boundary of this dictionary: it only explains "what a word means";
it does not replace "why this theory holds and where it was refuted" — for that, go back to
the notes. Everything I marked with ⚠️ is where mistakes and misuse are most likely, and is
worth a second look.

10 Related

• Major-course study roadmap — the semester arrangement of these courses (the dictionary is ordered after it)
• Major-course textbook index — 24 sets of notes (the source of the explanations in this dictionary)