Distributional Structure

Whether each language can be described in terms of a distributional structure, i.e. the occurrence of parts (ultimately sounds) relative to other parts, and whether that description is complete without bringing in history or meaning.

The distribution of an element is the sum of all its environments. An environment of A is an existing array of its co-occurrents, each in a particular position, with which A occurs in an utterance. A’s co-occurrents in a particular position are its selection for that position.

(Non-rigorously) that in terms of that feature we can form an organized system of statements describing the members of a set and their interrelations. Whether language is structured with respect to history, social intercourse, meaning or distribution is a matter for investigation.

That he freely puts together whatever elements have the meanings he intends. In fact he does so only by choosing members of classes that regularly occur together, in the order in which those classes occur.

Classes can: all elements group into classes whose relative occurrence can be stated exactly. But the occurrence of a particular member of one class relative to a particular member of another must be stated in terms of probability, based on frequency in a sample.

No: every dialect studied, including colloquial and “substandard” speech, has regular occurrences, and the relative occurrence of elements can always be stated descriptively, without reference to history.

Direct enumeration (as some engineers proposed, e.g. departures from equiprobability per sound) gives no simple description and doesn’t order the restrictions. Grouping elements into sets at successive levels gives fewer, simpler statements, ordered so the larger restrictions come first and the list can be stopped at any point as an approximation.

Besides discovering the irreducible elements and their relative occurrence empirically, the linguist searches for a simple set of ordered statements expressing those facts.

No: several systems can be equally adequate (e.g. several phonemic solutions), and different ones may be simpler for different purposes. This non-uniqueness is harmless, since each system can be mapped onto the others.

Item (combinatorial) style: form A includes elements e + f, form B includes e + g, so all forms are combinations of elements; more algebraic and parsimonious. Process style: A is derived from B by changing f into g; more like historical statements, useful in morphophonemics. Both rest on relative occurrence, so both are distributional.

(1) In the language: yes, as much as any scientific structure exists in the data it describes, since the relations it states really hold there. (2) In the speakers: a question of fact not settled by the analysis itself, though some behaviour (e.g. “repeating” utterances of one’s own language with one’s own variants) shows perception along distributional lines.

By new utterances not used in building the structural description, both truly productive new formations and held-out data: essentially, testing how well the structure predicts new data.

A new formation can conform to a distributional pattern (cf. periodic/periodicity, simplicity) without showing the habit existed in speakers beforehand: the readiness to combine elements may arise only when needed, by association of partially similar words.

Speakers have heard many utterances, perceive partial similarities among their parts, and produce new combinations along the lines of those heard; new utterances are formed from the distributional relations among parts of previously heard utterances.

Language structure doesn’t match the structure of the physical world or of human response; a person’s meanings grow while their language stays constant; there are ideas one cannot readily express; and words lack a single or even coherent range of meanings because of sound change, homonymy, borrowing, forgotten metaphors.

In connected discourse, especially stylized scientific reports and, above all, formal proofs in mathematics and logic. Even so, form and meaning are two planes whose structures are similar in various respects but not identical.

Meaning makes it hard to decide whether either has one morpheme or two. Distribution decides: consist, resist, pertain, contain, retain pattern with per- and -sist, but nothing parallels person, so persist is two morphemes and person one.

There are scores of person prefixes with no obvious segment meaning “I”. Grouping words whose stems differ but whose sentence environments are the same, he assumed a shared morpheme and worked out its alternants morphophonemically, step by step obtaining morphemes like {z} ‘I’, {h} ‘you (sg.)’, {o} ‘persons excluding you’, {n} ‘another person’, whose combinations need not match any person/number categories we have in mind.

For a test utterance, look at many utterances sharing its first \(n\) phonemes and count how many different phonemes follow. Where this count peaks (greater after \(n\) than after \(n-1\) or \(n+1\)), put a tentative morpheme boundary, then check and correct with patterned relative occurrence. (Later called successor variety.)

(a) Morphemes differ in degree of independence (-hood vs -ness); (b) not every member of one class occurs with every member of another (not every adjective with every noun); (c) regularities hold only in narrow domains, the neighbouring phonemes or the sentence.

That meaning alone decides (hood fits only certain nouns, adjectives make sense with doctor, sentences are strung along by meaning). Harris: choices are made by meaning, but they build a stock of combinations that shapes later choices, for “language is not merely a bag of words but a tool with particular properties which have been fashioned in the course of its use.”

By the number of different morphemes it occurs with and how widely they are spread over classes and subclasses: -hood attaches to very few predecessors, -ness to many. Members of a distributional class share an element of meaning, stronger the more distributional characteristics the class has.

If A and B are more different in meaning than A and C, their distributions will often be more different too: difference of meaning correlates with difference of distribution.

Oculist and eye-doctor occur in almost the same environments, except sentences containing both (“An oculist is just an eye-doctor under a fancier name”). Many environments of oculist (“I’ve had my eyes examined by the same oculist for twenty years”) rarely take lawyer. What matters is the relative frequency of environments, not truth.

No: we can’t list all co-occurrents (rare selections are unobservable) or define meaning fully from them, but we can roughly measure differences in selection between elements and so say something about differences in meaning.

A consecutive discourse of one or more persons: there are regularities between parts of neighbouring sentences (e.g. pronouns), which don’t extend from one discourse to another.

Segment the flow of speech into parts, arbitrarily at first, then group “similar” segments; only similarity groupings have distributional relations. Segmentation may later be revised (including simultaneous components like tone or stress contours) to make distributions more regular.

One native speaker repeats \(x_1z\), \(x_2z\), \(y_1z\)…; if a second speaker can tell \(x_1z\) from \(y_1z\) but not \(x_1z\) from \(x_2z\), then \(x_1\) and \(x_2\) are free variants (one similarity grouping). It tests discrimination of sound, not meaning.

Dependence: if in some environment only AB occurs, never A or B alone, AB is set up as one element; beyond that, elements show degrees of dependence (-dom follows nouns; -ish follows nouns and adjectives). Substitutability: A and B form a substitution set when they share (partially) the same environments, so dependencies can be stated between sets.

The stretches within which dependence and substitutability statements hold (word, phrase, clause, sentence). Co-occurrences within a domain can make the domain itself an element whose co-occurrences are regular within a larger domain, like stems and suffixes within words, words within phrases.

Between neighbouring sentences: if one says A buys B from C, a neighbour with C sells B to … likely continues with A; active and passive sentences invert their nouns. Across a whole discourse, discourse-specific substitution sets recur throughout, though these regularities are weaker than within a sentence.

Speakers’ habits yield regular utterances but they aren’t aware of the distributional details. So we study an actual corpus and derive regularities that would have generated these utterances, hoping they also apply to further utterances.

Techniques that favour the appearance of utterances relevant to a question (e.g. give a speaker a noun + transitive verb and ask for many completions) without pushing them to say what they never would naturally. Rare utterances would be prohibitively laborious to find in an arbitrary corpus.

Under eliciting, the verbs after “The pianist who —” are much the same as after “The pianist —”, and so for every noun: the verb selection of who equals that of the preceding noun, a mark of pronominal elements.

Sentences with a transitive verb (“The kids broke that window last week”) can be elicited with the passive morpheme, the nouns inverted and the rest unchanged (“That window was broken by the kids last week”). Inversion without the passive mostly fails (“The movie saw the kids”), and tense changes keep the nouns but not the rest of the sentence.

If A occurs in environment X, and AB does too, but B does not, then A is the head of AB. Notions like “subject” and “object” are semantic names for distributional positions.