{
  "version": "https://jsonfeed.org/version/1",
  "title": "natural-language on Context may be all you need",
  "icon": "https://www.gravatar.com/avatar/8d4e83e124462d944a8d448a6f213e9f?s=96&d=https%3A%2F%2Fmicro.blog%2Fimages%2Fblank_avatar.png",
  "home_page_url": "https://matt.thompson.gr/",
  "feed_url": "https://matt.thompson.gr/feed.json",
  "items": [
      {
        "id": "http://thompsonson.micro.blog/2026/08/13/131513.html",
        "title": "[IA Series 12/n] The Grammar of Natural Language — A Term Sheet",
        "content_html": "<p><em>This is a term sheet for the grammar of natural language, in the same spirit as the <a href=\"https://matt.thompson.gr/2025/05/16/ia-series-n-intelligent-agents.html\">Intelligent Agents Term Sheet</a> and the <a href=\"https://matt.thompson.gr/2026/08/13/125930.html\">Grammar of Logic</a>. The core terms come from the standard linguistics literature and Russell and Norvig&rsquo;s &lsquo;Artificial Intelligence: A Modern Approach&rsquo; (Chapter 23).</em></p>\n<h2 id=\"introduction\">Introduction</h2>\n<p>My interest in meaning goes back to reading the Thesaurus as a kid — grammar didn&rsquo;t interest me much, but meaning did. The agents in this series are natural-language interfaces: they read prompts and write answers in this language. The grammar of logic (Series 11) is how an agent reasons; the grammar of natural language (Series 12) is how it communicates. The LLM agents I build sit exactly on that bridge, so the language they process deserves the same term-sheet treatment as the logic they reason in. This post is deliberately lean — a shared vocabulary, not a textbook.</p>\n<h2 id=\"the-grammar--core-terminology\">The Grammar — Core Terminology</h2>\n<h3 id=\"morphology--the-shape-of-words\">Morphology — the shape of words</h3>\n<ul>\n<li><strong>Morpheme</strong>: The smallest unit of meaning. A word is built from morphemes.\n<ul>\n<li><strong>Free morpheme</strong>: Stands alone (<code>run</code>, <code>book</code>).</li>\n<li><strong>Bound morpheme</strong>: Attaches to another morpheme (<code>-s</code>, <code>-ed</code>, <code>un-</code>).</li>\n</ul>\n</li>\n<li><strong>Inflection</strong>: Modifying a word to fit grammar without changing its core meaning — <code>walk</code> → <code>walked</code> (tense), <code>cat</code> → <code>cats</code> (number). It does not create a new word.</li>\n<li><strong>Derivation</strong>: Building a <em>new</em> word — <code>run</code> → <code>runner</code>, <code>happy</code> → <code>unhappy</code>. It can change meaning and category.</li>\n</ul>\n<h3 id=\"lexical-categories-parts-of-speech\">Lexical categories (parts of speech)</h3>\n<ul>\n<li><strong>Noun (N)</strong>: Names an entity — <code>agent</code>, <code>pipeline</code>.</li>\n<li><strong>Verb (V)</strong>: Names an action or state — <code>query</code>, <code>deploy</code>.</li>\n<li><strong>Adjective (Adj)</strong>: Modifies a noun — <code>observable</code>, <code>stochastic</code>.</li>\n<li><strong>Adverb (Adv)</strong>: Modifies a verb, adjective, or sentence — <code>quickly</code>, <code>partially</code>.</li>\n<li><strong>Determiner (Det)</strong>: Marks a noun phrase — <code>the</code>, <code>a</code>, <code>this</code>.</li>\n<li><strong>Preposition (P)</strong>: Relates a noun phrase to the rest of the sentence — <code>in</code>, <code>through</code>, <code>at</code>.</li>\n<li><strong>Pronoun</strong>: Stands in for a noun phrase — <code>it</code>, <code>they</code>.</li>\n<li><strong>Conjunction</strong>: Joins constituents — <code>and</code>, <code>or</code>, <code>if</code>.</li>\n</ul>\n<h3 id=\"phrases-and-sentences\">Phrases and sentences</h3>\n<ul>\n<li><strong>Phrase</strong>: A group of words functioning as a unit. Named by its head — <strong>noun phrase (NP)</strong>, <strong>verb phrase (VP)</strong>, <strong>prepositional phrase (PP)</strong>.</li>\n<li><strong>Constituent</strong>: A word or group of words that behaves as a single unit in the grammar.</li>\n<li><strong>Clause</strong>: A phrase built around a verb. <strong>Independent</strong> clauses stand alone as sentences; <strong>dependent</strong> clauses do not.</li>\n<li><strong>Sentence (S)</strong>: An independent clause — a complete unit of the grammar.</li>\n<li><strong>Parsing</strong>: Recovering a sentence&rsquo;s grammatical structure from its word string.</li>\n</ul>\n<h3 id=\"grammatical-functions\">Grammatical functions</h3>\n<ul>\n<li><strong>Subject</strong>: The noun phrase a sentence is about; typically the agent of the action.</li>\n<li><strong>Object</strong>: The noun phrase the action is directed at.</li>\n<li><strong>Predicate</strong>: What the sentence says about the subject — the verb and its dependents.</li>\n<li><strong>Agreement</strong>: Grammatical matching between constituents — subject and verb agree in number (<code>the gate is</code> vs <code>the gates are</code>).</li>\n<li><strong>Case</strong>: The grammatical role of a noun phrase — <code>I</code> (subject) vs <code>me</code> (object).</li>\n<li><strong>Tense</strong>: When the event occurs — past, present, future.</li>\n<li><strong>Aspect</strong>: How the event unfolds in time — progressive, perfective.</li>\n<li><strong>Mood</strong>: The speaker&rsquo;s attitude to the event — indicative, imperative, subjunctive.</li>\n<li><strong>Voice</strong>: How the action relates to its participants — active (<code>the agent queries</code>), passive (<code>is queried</code>).</li>\n<li><strong>Word order</strong>: The ordering of subject, verb, and object — English is SVO.</li>\n</ul>\n<h2 id=\"the-grammar-stated\">The Grammar, stated</h2>\n<p>The standard first approximation is a <strong>context-free grammar</strong> of phrase-structure rules:</p>\n<pre tabindex=\"0\"><code>S  → NP VP\nNP → Det N | Det Adj N | Pronoun | ProperNoun\nVP → V NP | V NP PP | V AdvP\nPP → P NP\n</code></pre><p>A caveat worth naming: natural language is <em>not</em> fully context-free — agreement, case, and dependencies leak across the boundaries this grammar draws. The CFG is the workable approximation, not the whole truth.</p>\n<h2 id=\"semantics--meaning\">Semantics — Meaning</h2>\n<ul>\n<li><strong>Reference</strong>: What an expression points at in the world — <code>the merge gate</code> refers to a particular object.</li>\n<li><strong>Sense</strong>: The meaning of an expression independent of what it points at — <code>the morning star</code> and <code>the evening star</code> differ in sense but share a reference.</li>\n<li><strong>Compositionality</strong>: The meaning of a whole is a function of the meanings of its parts (Frege&rsquo;s principle).</li>\n<li><strong>Semantic roles</strong>: The parts participants play in an event — <strong>agent</strong> (does it), <strong>theme</strong> (undergoes it), <strong>instrument</strong> (does it with), <strong>recipient</strong> (receives it).</li>\n<li><strong>Lexical semantics</strong>: Word meaning, organized into relations — <strong>synonymy</strong> (same meaning), <strong>antonymy</strong> (opposite), <strong>hyponymy</strong> (kind-of: <em>deployment</em> is a hyponym of <em>action</em>), <strong>meronymy</strong> (part-of), <strong>polysemy</strong> (one word, related senses: <em>gate</em> = the physical object or the decision), <strong>homonymy</strong> (one word, unrelated senses: <em>bank</em>).</li>\n<li><strong>Distributional semantics</strong>: The view that meaning is <em>determined by distribution</em> — a word&rsquo;s meaning is fixed by the contexts in which it occurs. The <a href=\"https://en.wikipedia.org/wiki/Distributional_semantics#Distributional_hypothesis\">distributional hypothesis</a> is due to <a href=\"https://doi.org/10.1080/00437956.1954.11659520\">Harris (1954)</a>; <a href=\"https://en.wikipedia.org/wiki/Distributional_semantics\">Firth (1957)</a> gave its slogan: <em>&ldquo;a word is characterized by the company it keeps.&quot;</em> Meaning is quantified as vectors over co-occurrence (<a href=\"https://en.wikipedia.org/wiki/Latent_semantic_analysis\">LSA</a>, <a href=\"https://en.wikipedia.org/wiki/Word2vec\">word2vec</a>) — the basis of modern LLM semantics, and the natural-language counterpart of the logic sheet&rsquo;s Determination: where logic&rsquo;s truth is <em>attributed</em> by a human-chosen model, natural-language meaning is <em>determined</em> by the distribution of the corpus — learned from usage, not attributed.</li>\n<li><strong>Compositional distributional semantics</strong>: Merging the grammar&rsquo;s compositionality with distributional vectors, so the meaning of a phrase is computed from the meanings of its parts (<a href=\"http://www.cs.ox.ac.uk/people/stephen.clark/papers/qai08.pdf\">Clark, Coecke &amp; Sadrzadeh 2008</a>). The same family as <a href=\"http://ftp.icsi.berkeley.edu/ftp/pub/techreports/1991/tr-91-012.pdf\">distributed representations (Rieger 1991)</a>, where meaning lives spread across many dimensions rather than in one place.</li>\n<li><strong>Ambiguity</strong>: A phrase with more than one meaning.\n<ul>\n<li><strong>Lexical</strong>: a word has multiple senses.</li>\n<li><strong>Structural</strong>: the grammar permits two structures — <em>&ldquo;the agent sees the man with the telescope&rdquo;</em> attaches the PP to the object or the verb.</li>\n</ul>\n</li>\n<li><strong>Truth conditions</strong>: The conditions under which a sentence is true. This is where natural-language meaning meets the logic term sheet: a sentence&rsquo;s truth conditions are <em>determined</em> by the grammar and <em>attributed</em> by the reader, exactly as the logic sheet&rsquo;s Determination defines.</li>\n<li><strong>Anaphora</strong>: A word that refers back to an earlier one — <em>&ldquo;the agent finished its run; it then merged.&quot;</em></li>\n</ul>\n<h2 id=\"pragmatics--using-the-grammar\">Pragmatics — Using the Grammar</h2>\n<ul>\n<li><strong>Pragmatics</strong>: What a speaker <em>means</em> by an utterance, beyond what the sentence literally says.</li>\n<li><strong>Utterance</strong>: A particular use of a sentence in a context.</li>\n<li><strong>Context</strong>: The situation of the utterance — who speaks, to whom, when, and against what shared background.</li>\n<li><strong>Speech acts</strong>: What an utterance <em>does</em> — <strong>asserting</strong> (<code>the pipeline is green</code>), <strong>asking</strong> (<code>is it green?</code>), <strong>requesting</strong> (<code>make it green</code>), <strong>promising</strong>.</li>\n<li><strong>Implicature</strong>: What is communicated beyond what is said — <em>&ldquo;the gate is open&rdquo;</em> can imply &ldquo;you may go,&rdquo; without stating it.</li>\n<li><strong>Presupposition</strong>: What an utterance takes for granted — <em>&ldquo;the merge failed&rdquo;</em> presupposes a merge existed.</li>\n<li><strong>Deixis</strong>: Meaning that depends on the context of utterance — <em>I</em>, <em>you</em>, <em>here</em>, <em>now</em>, <em>this</em>.</li>\n<li><strong>Discourse</strong>: A sequence of utterances — the level above the sentence, where anaphora, coherence, and topic live.</li>\n</ul>\n<h2 id=\"closing\">Closing</h2>\n<p>Logic is how the agent reasons; natural language is how it communicates; the ubiquitous language of the ontology work is the shared vocabulary expressed in this grammar. The LLM sits on that bridge — parsing the natural language into meaning, and generating natural language that carries it back. This term sheet names the pieces of that interface so the series can describe it precisely.</p>\n",
        "date_published": "2026-08-13T13:15:13+02:00",
        "url": "https://matt.thompson.gr/2026/08/13/131513.html",
        "tags": ["ia-series","natural-language"]
      }
  ]
}