NOTE-073: Haspelmath 2011, the indeterminacy of word segmentation

Contribution

The paper takes the ten morphosyntactic criteria that linguists actually use to separate affixes, clitics, words and phrases. It shows, criterion by criterion and with cross-linguistic counterexamples, that none of them singles out orthographic words. It also shows that the usual practice of combining criteria into “test batteries” is not rigorous, because different authors pick different criteria (§4, Table 1). Before this paper, the difficulty of defining “word” was well known but was usually treated as a nuisance. The paper turns it into an explicit negative thesis. With no cross-linguistic word, there is no cross-linguistic morphology–syntax division, and general claims that rest on one (lexical integrity, creole morphology, complexity, the suffixing preference, WALS-style affixal-vs-periphrastic typology, (43), p. 66) are not currently testable.

Key insight

Morphology and syntax are defined in terms of the word, and every attempt to define the word runs through criteria that do not converge. The distinction therefore has no foundation that is independent of the Western habit of writing spaces. The alternatives Haspelmath offers are as follows:

• “word in language X” as a descriptive category, which is legitimate but has nothing to say about comparison;
• “word” as a fuzzy category, which is theoretically meaningful only if grammatical units cluster along a boundness scale, and that is an empirical question nobody has answered;
• cross-linguistically definable comparative concepts (formative, morph, root, construct, bound/free construct, (46), p. 70).

Assumptions

• The target is orthographic practice. A criterion counts as a success only if it picks out “words as identified by written spaces in the normal spelling used by speakers and/or by linguists” (p. 38). If a systematic level unlike spacing were found, “we would not want to call it ‘word level’” (p. 38). This choice is deliberate, and it is what makes “no criterion works” a sharp result. It also means the result concerns the traditional word concept, not every possible intermediate unit.
• Only morphosyntactic criteria. Four other bases are set aside first (§2):
  • speaker intuitions: literate speakers’ intuitions are not independent of spelling rules, and Peterson’s (2008) six Kharia speakers disagreed on word counts, e.g. 1 vs 4 words for one sentence (fn. 5);
  • semantic non-compositionality: idioms are non-compositional phrases, and many complex words are compositional;
  • orthography: it is arbitrary and inconsistent, e.g. German zu gehen vs wegzugehen;
  • phonology: phonological words are acknowledged not to coincide with morphosyntactic words, and do not even converge within one language (Turkish harmony vs stress; Schiering et al. 2010).
• Comparative vs pre-established categories. The argument is meant to apply equally to “word” as an innate UG category and to “word” as a comparative concept (fn. 3).
• One general criterion for the mobility test. It assumes that “same construction, different order” can be told apart from “different construction”. The paper argues there is no criterion for this (fn. 18). That is part of the result, not an external assumption, but the mobility and free-occurrence arguments depend on it.

Key results

• §3 — none of the ten criteria is necessary and sufficient. The list is (2), p. 38.

  1. Potential pauses (§3.1). Clitics are words but cannot be preceded by a pause. Dalabon allows pauses inside a morphosyntactic word: ka-h-…rak-…m-iyan, (3).

  2. Free occurrence (§3.2). Bloomfield’s “minimum free form” is too strict, since compounds such as firewater would come out as phrases. It is also too loose: a flower, to Lagos and take the money would come out as words, because a, to and English transitive verbs cannot stand alone. It gives only a one-way diagnostic: something that can stand alone is not an affix (p. 40).

  3. Mobility / internal fixedness (§3.3). Few words are freely mobile, and the test depends on deciding “same construction vs different construction”:

     • Latin cum;
     • Finnish negation e-t mene, (4);
     • the Basque auxiliary, (5)–(6);
     • Dalabon prepound extraction, (7);
     • the Lithuanian reflexive -si, (8).

     Affix order is not always fixed either (Chintang free prefix order, Bickel et al. 2007; Romance object markers as prefixes and suffixes, and clitic climbing, (9)–(10)).

  4. Uninterruptibility (§3.4). Interruptibility by a free form is sufficient for two-word status. As a definition (“maximal uninterruptible string of morphs”) it is too loose (both+my parents, even+Kim, very+good, (11)). It is also violated by incorporation (Pawnee tā-tu-rīks-kut, (12)).

  5. Non-selectivity (§3.5). It is an important criterion in practice (Lezgian -na vs -ni, (13)). But promiscuous elements that cannot be words exist: edge inflection, the Òko locative high tone (14), and English possessive ‘s with haplology (15). Non-selectivity therefore cannot be sufficient.

  6. Non-coordinatability (§3.6). Affixes can be ellipted in coordination (Turkish suspended affixation, (17); English pro-choice and -gun control, German trink- und essbar, Hungarian, (18); Japanese tense -ta, (19)). Affixes can also be coordinated (pro- and anti-war, (20)).

  7. Anaphoric islandhood (§3.7). Outbound anaphora is blocked by non-referentiality, not wordhood. Non-word phrases block it too ((22): weapons of mass destruction, Italian giacca a vento, German Fahrrad fahren), and names allow it (McCarthyites, (23)). Inbound anaphora is possible in Georgian imis-ian-i ‘him-ite’ (24; Harris 2006).

  8. Nonextractability (§3.8). Word parts cannot be extracted, but neither can individual words or many phrases (What kind (of) did you buy a _ cake?, (27)). Extraction targets whole referential phrases, so nonextractability says nothing about wordhood (p. 52).

  9. Morphophonological idiosyncrasy (§3.9). Idiosyncrasy occurs in clitic-like combinations:

     • the Pitjantjatjara ergative -tu/-‘u/-tju/-ngku at the NP edge, (28);
     • Russian s/so (so mnoj vs s mneniem);
     • Polish mobile person markers that condition stem alternation, (29).

  10. Deviations from biuniqueness (§3.10). All four deviation types occur in syntax:

      • zero: Ø Tomek’s house (31), Bulgarian renarrated (32);
      • multiple exponence: have eaten, avons mangé (33), Russian s + instrumental (34);
      • cumulative exponence: French du (35), Tagalog kita (36);
      • morphomic: the Slovene l-participle (37), auxiliary selection.

      The realizational/paradigm approach to morphology and the constructional approach to syntax differ “mostly … from different traditions, not from any substantive differences” (p. 59).

• §4, Table 1 (p. 60) — nine test-battery studies, and the criteria used vary. The nine studies are Zwicky & Pullum 1983, Kanerva 1987, Bresnan & Mchombo 1995, Ackerman & LeSourd 1997, Monachesi 1999, Harris 2000, Milićević 2005, Lieber & Scalise 2006 and Bickel et al. 2007. From the extracted rows, the number of studies using each criterion is:

  criterion	studies
  free occurrence	2
  mobility / fixedness	4
  uninterruptibility	2
  non-selectivity	6
  non-coordinatability	6
  anaphoric islandhood	2
  nonextractability	2
  morphophonological idiosyncrasy	5
  biuniqueness	1

  No criterion is used by all nine. The method “is not rigorous, because the criteria can be selected opportunistically by the author” (p. 59).

• §5 — the language-specific word is legitimate but beside the point. In most languages one could posit many intermediate units (stems, clitic groups, tight and loose compounds, and so on) with no reason to single one out as “word” (pp. 61–62). What can be asserted “with great confidence” is that all languages have degrees of tightness. No language is both thoroughly isolating and thoroughly non-configurational, and this may be “an innate design feature” (p. 62).

• §6, Figure 1 (p. 64) — clustering vs random distribution. A fuzzy word concept is theoretically significant only if units cluster on the boundness scale. “We do not have the empirical evidence yet” (p. 64). Fn. 24 says why a test is hard: the population is open-ended, it is unclear how to quantify boundness, and the continuum is multidimensional.

• §7, (41)–(42) (pp. 65–66). (41) is the negative conclusion. (42) gives three consequences:

  • (a) believers should find a consistent identification procedure;
  • (b) the uncommitted should stop worrying about the division;
  • (c) everyone should be careful with claims that presuppose the word.

  (43) lists seven research areas that presuppose it.

• §8 — lexical integrity is partly circular. Formulations such as Lapointe (1980), Selkirk (1982) and Anderson (1992), (45), “come dangerously close to saying that linguistic sign combinations that cannot be manipulated by syntactic rules cannot be manipulated by syntactic rules” (p. 68). The non-circular version (“different principles”) is “very weak” (p. 68).

• §9 — practice. For everyday description, keep calling a spaced unit a “word”, as long as no theoretical difference is implied (p. 69). For comparison, use the comparative concepts in (46).

Claims

id	claim	strength	support
C1	None of the ten criteria is both necessary and sufficient for orthographic wordhood	strong	counterexamples for each criterion from cited descriptions, §3.1–3.10. One clear counterexample per direction is enough to defeat “necessary and sufficient”, and each criterion gets at least one
C2	No combination of criteria gives a definition that matches orthographic practice	moderate	informal argument. §4 shows that test batteries differ across nine studies (Table 1) and that the criteria conflict (e.g. Lezgian -ni, Polish -(e)m). It does not search combinations exhaustively. The abstract states this more strongly than §4 shows
C3	Hence there is no good basis for a general morphology–syntax distinction	moderate	follows from C1–C2 together with the premise that morphology and syntax are defined via the word (p. 32; fn. 2 says he has “never seen” a word-independent definition). The premise is asserted, not surveyed
C4	Deviations from biuniqueness are not specific to morphology; all four types occur in syntax	strong	examples (31)–(37), largely following Spencer 2001
C5	Realizational morphology and constructional syntax differ by tradition, not substance	weak	assertion, “as far as I have been able to determine” (p. 59)
C6	A fuzzy word concept is significant only if units cluster on a boundness scale, and there is no evidence yet that they do	moderate (logic) / open (empirics)	conceptual argument with Fig. 1. No data are presented, and fn. 24 concedes a test is hard to design
C7	Classic lexical-integrity formulations are partly circular	moderate	informal argument, §8, citing Sproat 1988 and Julien 2002
C8	All languages show graded tightness of sign combination, and isolating languages tend to be configurational	weak	assertion (p. 62). Presented as a generalization with no data in the paper
C9	The idea of universal words may come from the bias of writing with spaces	weak	offered as a possibility “that should be taken seriously” (p. 33), not argued
C10	Speakers’ intuitions cannot settle segmentation	weak–moderate	the absence of a word for ‘word’ in many languages (Dixon & Aikhenvald 2002), the etymology of ‘word’ terms, and one small study (six Kharia speakers, fn. 5)

Concepts

• morphosyntactic word: the paper’s term for the grammatical, morphological or syntactic word, or word-form (fn. 9). It is the unit whose existence is in question. It is distinct from the phonological word, which is set aside.
• test battery: a set of wordhood criteria chosen by an author and applied to a case. It is persuasive when the criteria converge, and not rigorous because the choice of criteria is free (§4).
• comparative concept: a category defined for cross-linguistic comparison. It makes no claim that languages have identical descriptive categories (fn. 3; Haspelmath 2010).
• formative (46a): “a minimal coherent set of phonological features that plays a role in the language system (= a minimal sign)”. A morph is a formative that biuniquely expresses a meaning.
• construct (46d): a set of formatives that together play a role in the system. A bound construct cannot be a complete utterance. A free construct can. All words and phrases are constructs, and all clitics and affixes are bound constructs (p. 70).
• non-selectivity / promiscuity: the ability to combine with hosts of many categories. It is a practical word criterion that edge inflection defeats (§3.5).
• clustering vs random distribution (Fig. 1): the two ways grammatical units might be spread along a boundness scale. Only clustering would make “affix / clitic / word” theoretically meaningful.

Connections

The three works share one thesis, that the morphology–syntax split is not a principled division of grammar. They reach it by different routes, and their relationships are asymmetric.

Who cites whom.

• Haspelmath (2011) cites neither of the others; both are later.
• Bruening (LIT-071) cites Haspelmath 2011 twice. Footnote 2 (p. 1) says the lexicalist hypothesis needs a definable word, that “this assumption is false”, and “See Haspelmath 2011”, and then puts the issue aside. He also cites “Haspelmath 2011:52” (§5.1, p. 23) for the point that extraction targets phrases, not heads. That is §3.8 here.
• Krauska & Lau (LIT-031) cite the 2017 reprint of this paper (“Folia Linguist. 51, 31–80”, doi 10.1515/flin-2017-1005; same page range) twice: in a list of anti-lexicalist works (§1) and for “there are no good criteria to empirically define wordhood” (pp. 2, 15). They also cite Bruening 2018 once, in the same list.

Where they agree.

• Haspelmath and Bruening agree on the negative verdict about lexical integrity, and even on particular diagnostics. Haspelmath’s §3.6 (word-part coordination and ellipsis, Chaves 2008), §3.7 (Georgian inbound anaphora, Harris 2006) and §3.8 (extraction) match Bruening’s §3.1, §5.3 and §5.1. Both find the circularity or vacuity of lexical integrity: Haspelmath’s §8, and Bruening’s “superfluous”.
• Krauska & Lau adopt the conclusion shared by both and apply it to production models.
• Haspelmath’s orthography-bias suggestion (p. 33) is echoed by Krauska & Lau’s conclusion (p. 16) that intuitive wordhood may be grounded in white-space orthography. They cite Hoosain 1992 for this, not Haspelmath.

Where they part.

• Method and commitment. Haspelmath is a typologist and framework-neutral. He argues that the word cannot be defined across languages, and he proposes comparative concepts, not a theory of grammar. Bruening argues inside generative syntax, grants for the sake of argument that words are identifiable, and replaces “word” with the X0/XP distinction of a single syntactic component. From Haspelmath’s standpoint, X0/XP is itself a theory-internal category that would need its own cross-linguistic identification procedure. The paper does not say this. It is an inference from §§3–5 and (46).
• Morphology. Haspelmath’s constructive hint is to merge realizational morphology with constructional syntax (§3.10.5). Bruening’s single component is syntagmatic, close to Distributed Morphology, and he concedes in fn. 6 that a paradigmatic (Becker-style) alternative is untested.
• The phonological word. Haspelmath excludes it from the question (§2.4). Bruening suggests that “word” probably has no meaning other than the prosodic one. Krauska & Lau deny that intuitive words are phonological units either (dyawanna).

Outside the trio. Haspelmath (2023), “Defining the word”, WORD 69(3):283–297, is the author’s own constructive sequel. From its abstract (not read here), it proposes a stipulative cross-linguistic definition (roughly: a free morph, a clitic, or a root or compound plus affixes) built on comparative concepts of the §9 kind.

Bearing on the record

• Nucleation. The caller says nucleation holds no other linguistics work. I checked, and theory.d/ contains only its template. So nothing is supported or contradicted. If a THEORY were written from this trio, the defensible content is C1 and C4 of this note. The claim “morphosyntax is one domain” would be Proposed at most, since the paper itself hedges it and §6 is open.
• Tokenization, the paper’s own claims only. The paper makes no claim about tokenization, subword segmentation or any computational unit. What it does establish:
  • there is no cross-linguistically valid criterion that picks out orthographic words, so whitespace-delimited “words” are not a linguistically privileged unit across languages;
  • within one language, “word” can often be defined consistently (§5, §10);
  • sub-word units are no more privileged, since morphs fail biuniqueness in morphology as much as in syntax (§3.10);
  • for practical work, the orthographic word is an acceptable convention with no theoretical status (§9, which explicitly includes computational linguists).
• Tokenization, extrapolation (mine, not the paper’s). The paper removes one possible rationale for word-level tokenization, namely that words are the natural atoms of language. It supplies no rationale for any particular alternative, whether BPE (ANTH-SOTA-007), bytes, or learned chunk boundaries (ANTH-SOTA-418, from ANTH-LIT-679, H-Net). Those choices are settled empirically in the anthology, and nothing here bears on those results. The seed’s “supports subword and byte-level approaches” should not be filed.
  • The one place where the paper and ML practice could meet is §6. Whether units cluster on a boundness scale is a statistical question, and learned-segmentation models produce boundaries that could in principle be examined for clustering across languages. The paper proposes no such test, and fn. 24 argues that quantifying boundness is the hard part.
  • ANTH-THEORY-021 (a model’s latent units are not its tokenizer’s) is consistent in spirit, but this paper is not evidence for it.
• Anthology. This carries no instruction for ML practice, so nothing belongs in the Anthology of the SOTA on this reading.

Limitations

• The target is orthographic practice by design (p. 38). “No criterion reproduces spaced words” is partly a result about spelling conventions. A critic can accept every counterexample and still hold that some non-orthographic intermediate unit is universal. Haspelmath answers only that such a unit would not deserve the name “word” (p. 38) and that there are too many candidates to privilege one (§5).
• The evidence is counterexample-based and relies on cited analyses. Several counterexamples depend on contested analyses: the Lezgian -ni classification, Japanese (19) as coordination (fn. 13), Romance object markers as affixes, and Turkish suspended affixation. Each is enough to defeat “necessary and sufficient”, but the paper does not quantify how often each criterion fails.
• Combinations are not tested systematically. C2 is argued from disagreement among nine studies, not from a search over weightings. §6 concedes that a fuzzy or weighted notion is possible, and that it is untested.
• The clustering question is posed, not answered. There is no sample and no measurement.
• Several broad claims are asserted. The innate graded tightness (p. 62), the equivalence of realizational and constructional approaches (p. 59), and the writing-bias origin of the word concept (p. 33) are not argued with data.
• The consequences in (43) are flagged, not assessed. Only Sampson’s complexity claim and McWhorter’s creole claim get a paragraph (pp. 66–67).

Open questions

• Do grammatical units cluster on a boundness scale when the criteria of §3 are coded across a diverse sample? This needs a sampling frame for an open class and a principled weighting (fn. 24). A typological database coding these properties for many languages would settle §6.
• Can the claims listed in (43) (the suffixing preference, analytic vs synthetic complexity, inflectional future and case counts) be restated in terms of comparative concepts, and do they survive? The paper calls this “an important task for future research” (p. 72).
• Does the author’s 2023 stipulative definition meet the 2011 criteria for a good comparative concept? It is not read here, and a reading of it would say whether it supersedes §9’s proposal or only fills it in.

Corrections to the seeded skim

• Tokenization is not in the paper. The seed says the paper “bears directly on tokenization” and “supports subword and byte-level approaches (H-Net and similar)”. The paper says nothing about tokenization, subword units, statistical segmentation or processing units. Computational linguists appear once, in a list of practitioners asking how to proceed (§9, p. 69), and the answer to them is to keep using the orthographic word as a convention. Sub-word units get no privilege either: morphs fail biuniqueness in both morphology and syntax (§3.10). Any tokenization claim is extrapolation (see Bearing on the record).
• The paper does not say “morphosyntax should be treated as one domain” as a finding. The conclusion is hedged: morphosyntax is “for theoretical purposes … currently best viewed as a unitary domain” (p. 72). The paper also explicitly leaves open that a cross-linguistic word could still be found (p. 33; (42a), p. 66).
• Language-specific words are fine. The seed’s framing (“whitespace segmentation is an orthographic convention, not a linguistic unit”) goes further than the text. §5 and §10 say “word in language X” can often be defined, and defined so that it predicts the spelling (p. 71). The claim is only that no definition is cross-linguistically valid.
• The other two works do not build on it in the way the seed says. The seed calls it “the typological foundation that c62 and c63 build on”. Bruening (LIT-071) cites it in footnote 2 and then sets the word-definition problem aside: his argument grants the lexicalist that words can be identified. Krauska & Lau (LIT-031) cite the 2017 reprint twice, for the conclusion only. See Connections.
• §4 is about test batteries in general. It is not specific to lexicalism. The history of lexicalism is in §8, where Haspelmath says lexical integrity re-emerged in generative grammar “since the 1970s” and was “not prompted by any progress in defining the word” (p. 67). §§7–8 are consequences, not the conclusion. The conclusion is §10.
• A small internal inconsistency in the paper. Table 1 lists “Ackema & LeSourd 1997”, but the reference list gives Ackerman, Farrell & LeSourd, Philip (1997).