Ways of Knowing

Ways of Knowing LING209 Language and Human Mind, Fall 2025

Announcements & Readings

  • No class on Monday due to Labor Day.

  • Required Readings:

    • LU Chapter 2.

    • Article by SEP section on Behaviorism.

    • Article by Marler (2005/1991).

Agenda

  1. Empiricism and its discontents.

  2. Rationalism: The slate isn't blank.

  3. The linguistic turn.

Understanding Language Models and Interpretation

  • Example: Spatial Relations (Page 5)

    • Request: "a red box on a pyramid under a sphere."

    • Model's Interpretation: Vertical stack of objects, with the red box at the bottom, supporting the pyramid, which supports the sphere (red box \rightarrow pyramid \rightarrow sphere).

    • Reasoning: Models (and humans) often interpret sequential spatial relations like "X on Y under Z" as a vertical dependency, where X supports Y, and Y supports Z.

    • This highlights the complexity of natural language interpretation and potential ambiguities.

  • Example: Object Permanence (Page 6)

    • Question: "Lisa cut an apple into two pieces. How many apples are there?"

    • Model's Answer: "There are still two apples."

    • This demonstrates a model's understanding of conceptual integrity, where cutting an object results in more pieces but doesn't necessarily reduce the count of the original type of object (if referring to the pieces as individual entities).

    • Model Parameters Shown: text-davinci-003, Temperature 0.7, Maximum length 256, Top P 1, Frequency penalty 0.

  • Mechanism: Guess Next Word (Pages 7-10)

    • Language models fill in blanks (e.g., "I would hardly call myself an expert on…") through an iterative process:

      1. See Context: Analyze the preceding words.

      2. Guess Next Word: Predict the most probable word based on the context and training data (e.g., "the future").

      3. Accuracy/Adjustment: Evaluate the guess and adjust internal settings or probabilities for future predictions. The example showed the eventual completion as "genuine human emotion."

  • Fundamental Learning Loop (Page 11)

    • A continuous cycle of learning and refinement:

      • Experience \rightarrow Guess next \rightarrow Check & Adjust \rightarrow Experience… (repeated 10,000,000,000,000 times or more).

    • This describes how both AI models and potentially human cognitive processes learn through vast amounts of data, feedback, and adaptation.

Empiricism: Knowledge from Senses

  • Core Idea: Knowledge is derived solely from sensory experiences.

  • John Locke's View (1690, "An Essay Concerning Human Understanding") (Page 12)

    • "Our senses, conversant about particular sensible object, do convey into the mind several distinct perceptions of things according to those various ways wherein those objects affect them."

    • Senses transmit information from external objects, producing perceptions (ideas) in the mind.

    • Examples: Ideas of yellow, white, heat, cold, soft, hard, bitter, sweet (sensible qualities).

  • Historical Associations (Page 13)

    • Prominent in Western Enlightenment philosophy: John Locke, George Berkeley, David Hume.

    • Earlier philosophical echoes in systems like the Kanada Sutra (6th-2nd c BCE), Greek Stoic philosophy (3rd c BCE), and 10th-century Arabic philosophy (Al Farabi, Ibn Sina).

    • Forms the foundation for many modern Artificial Intelligence (AI) systems, especially those processing language based on vast datasets.

  • David Hume's View (1740, "A Treatise of Human Nature") (Page 14)

    • "All our simple ideas in their first appearance are deriv’d from simple impressions, which are correspondent to them, and which they exactly represent."

    • Reinforces that all basic ideas originate from direct sensory impressions.

Empiricism's Discontents: Problems with the Senses

  • Two Primary Issues (Pages 15, 23):

    1. The senses can mislead.

    2. Some truths/knowledge do not seem to be available to the senses.

  • The Senses Can Mislead (Pages 17, 27)

    • Descartes's Skepticism: Illustrated by a comic, where Descartes challenges certainty in sensory input, questioning if one's senses could be lying or manipulated by an "evil demon" (a thought experiment).

    • He suggests grounding knowledge in "pure certainty" (e.g., mathematics) rather than fallible sensory perceptions.

    • Visual Illusions: Provide direct evidence that the senses are not always reliable and can be tricked (\rightarrow Poll result: (C) The senses aren't always reliable).

  • Knowledge Not Readily Available to Senses (Pages 28)

    • Mathematical Laws: The relationship between a circle's circumference and its diameter (Pi, or \pi) is an abstract mathematical concept (C = \pi d).

    • This knowledge is not directly observable or discoverable solely through sensory experience (\rightarrow Poll result: (B) is not readily available to the senses).

Rationalism: The Slate Isn't Blank

  • Core Idea: Sensory input is structured by the mind, meaning certain ideas are innate or "built-in" (a priori) rather than purely learned from experience.

  • Origin of Mental Structures (Page 30)

    • Mental structures that organize sensory input are proposed to come from evolution.

    • Evolution provides a mechanism for useful knowledge or predispositions to be present in the brains of newborns across species (e.g., people, birds, frogs) as biological properties that enhance viability.

  • The (Rationalist) Frog Analogy (Page 31)

    • Mentalism: Abstract categories (like "the frog") are mental representations that emerge from the interaction of innate mental structures with sensory experience.

    • These mental structures are products of evolution.

  • Lingering Problems with Rationalism (Page 31)

    • Innate ideas don't guarantee that beliefs derived from them are true.

    • Posing abstract concepts (like mathematics) as solely resident in people's minds raises the question of whether math is only "real" because humans have minds.

Avoiding Binary Thinking (Page 32)

  • It's important to recognize that a strict "rationalist" or "empiricist" label may not apply universally.

  • Different kinds of knowledge might lean more towards one or the other:

    • Mathematical knowledge vs. scientific knowledge.

    • Physical principles vs. social principles.

  • The question then becomes: What kind of thing is knowledge of language?

  • Poll (Page 33): Domains like laws of motion, social norms, language, and beauty can be rated on an Empiricist-Rationalist spectrum, encouraging nuanced application of these philosophies.

The Linguistic Turn: Language Knowledge

  • The Debate: Empiricist vs. Rationalist for Language (Page 35)

    • Rationalist View: Humans, like other species, evolved with built-in (instinctual) behaviors, and language is one such innate capacity.

    • Empiricist View: Humans possess superior general-purpose reasoning skills (compared to other animals) that allow them to construct linguistic knowledge based on experience.

    • Non-binary approach: Suggests that perhaps general-purpose learning mechanisms are built-in, rather than language itself being pre-programmed.

Adger's Argument Against Empiricism for Language

  • Challenge to the "Mental Dictionary" (Page 36)

    • While words (e.g., English "Cat," Maori "Ngeru," Tamil "Pūṉai") are indeed learned from experience, the complexity arises for infants.

    • How do babies, from the chaotic sounds and signs they encounter, manage to isolate and associate specific ideas purely from experience?

  • Challenge to "Mental Grammar": Can Rules Be Learned from Experience? (Page 37)

    • Consider English question formation (e.g., "Who did you think would win?").

    • Recipe for 'wh'-questions:

      1. Identify the phrase to be questioned.

      2. Move it to the front of the sentence.

      3. Replace it with an appropriate 'wh'-word ("who," "what," "where," etc.).

      • Example: "You think SOMEONE would win." \rightarrow "Who did you think would win?"

      • Other Examples:

        • "Moira went to the store AT SOME TIME." \rightarrow "When did Moira go to the store?"

        • "I get this darn thing to turn on BY DOING SOMETHING." \rightarrow "How do I get this darn thing to turn on?"

        • "David put those purple socks SOMEWHERE." \rightarrow "Where did David put those purple socks?"

    • Rule for 'that' Optionality (Page 39)

      • In sentences with verbs like "thought," "believe," "heard," "said," etc., the complementary conjunction "that" is often optional.

      • Example: "Anita thought David would win." \leftrightarrow "Anita thought that David would win."

      • Example: "Alexis said (that) David is wearing SOMETHING." \rightarrow "What did Alexis say (that) David is wearing?"

  • The Problem with Combining Rules (Pages 40-43)

    • Combining the question formation rule and the 'that' optionality rule leads to ungrammatical sentences in English, which children never explicitly learn not to say.

    • Statement: "You thought Anita would win."

    • Applying Q-rule: "Who did you think would win?" (Acceptable)

    • Applying 'that' optionality: "You thought that Anita would win." (Acceptable)

    • Combining both: "Who did you think that would win?" (Unacceptable - indicated by a poll where most speakers find sentence 2 "Who did you think would win?" more acceptable than sentence 1).
      \bullet

      The that is only optionally deletable in certain contexts, not universally: "I think (that) Anita will win." vs. "Who do you think \times (that) will win?"

    • An empiricist perspective would demand that children hear sufficient examples, both correct and incorrect, to deduce these complex rules.

  • Empirical Evidence from Corpus Analysis (Pearl & Sprouse 2013, Pages 44-45)

    • Research Question: Do children receive enough linguistic input to learn these specific grammatical distinctions purely from experience?

    • Methodology: Analysis of a corpus of speech directed at 25 children (approximately 1 million words).

    • Findings:

      • Over 30,000 utterances with question-like structures were found.

      • However, only a very small percentage (<$ \$1\% or 2\%$$) of these were "embedded" questions (involving verbs like "said," "believe").

      • Crucially, utterances corresponding to the acceptable combined pattern ("Who do you think will win?") and the unacceptable combined pattern ("Who do you think that will win?") were found zero times in the child-directed speech (Table 5.1). This means children are not hearing the data necessary to distinguish between them empirically.

    • Conclusion: The linguistic input children receive is insufficient to differentiate between grammatical rules for acceptable and unacceptable English questions. This strongly supports the idea of built-in learning structures, thus aligning with a rationalist view of language acquisition.

Overall Take-Aways (Page 46)

  • There is a significant debate regarding the origins of knowledge: entirely from sensory experience (Empiricism) versus involvement of some built-in mental structures (Rationalism).

  • Knowledge of language serves as a critical point of contention in this debate: Can children acquire the intricate grammar of their native language without some innate, pre-existing mental framework for language?

  • Next Topic: Investigating whether animals are empiricists in their ways of knowing.