What exactly is a word? How can we distinguish morphology (= word structure) from syntax? How do we tell phrases apart from complex words, and clitics from affixes? In my 2011 paper (“The indeterminacy of word segmentation”) I said that we don’t know: The various criteria that we typically use do not necessarily give the same results, and it could be that a notion of “word” plays no role in general grammar. It is best not to presuppose that we know what a word is. That was a negative contribution, but more recently, I have been thinking about a more constructive way of dealing with this situation. One result is my new paper “Defining the word” (published this month in the journal WORD, in open access).
The radical critique of my 2011 paper was widely noticed, but almost nobody has drawn the radical conclusion of avoiding the word “word” and the word “morphology” – even though this should be a logical consequence. (The only other option would be to reject my critique and to defend the universal morphology-syntax distinction – but this has not been done either.)
De facto, while many linguists recognize the failure of earlier attempts at delimiting morphology and syntax, we all continue doing business as usual. We publish papers in journals with names such as “Morphology” (e.g. Haspelmath 2020), we talk about “affix borrowing”, and we talk about “prosodic words”, a notion that presupposes the notion of grammatical word (as is noted by Newell et al. 2017: 2: “Phonology per se lacks a theory of how the ‘word’ comes to be, and phonologists generally look to morphologists or syntacticians to derive this construct.“)
In particular, new generations of linguists grow up, and they are being taught “morphology”, e.g. through textbooks such as Gebhardt (2023), but these textbooks do not tell them what words are. Gebhardt is very honest:
„Full disclosure: no one knows what a word is. There is, so far, no definition that covers all the things that linguists identify as words in all languages.“ (Gebhardt 2023: 59)
But then, what is the point of having separate morphology textbooks? And what would allow us to continue to talk about “morphological complexity” (e.g. Baerman et al. 2017; Arkadiev & Gardani 2020) or “morphological borrowing” (e.g. Gardani 2020)?
These are deep questions which we started asking 150 years ago (when the term “morpheme” was first created, by Baudouin de Courtenay in Kazan, suggesting a unity of syntax and morphology), but which we don’t seem to be getting answers to. The idea that there is no separate morphology and that syntax extends to words (clearly articulated in the Chomskyan tradition by Selkirk 1982) is still very popular (witness the upcoming 3rd workshop in the series “Morphology as syntax”), but the opposite view that morphology is autonomous is also widespread (Spencer 2019 opposes DM to “Mainstream Morphology”, which does not regard morphology as a kind of syntax). Thus, we still don’t know whether morphology exists, and whether words exist (i.e. in general, as natural parts of human languages).
It does not seem that these deep questions will be resolved anytime soon, so in the meantime, I decided to switch to a constructive mode and propose a definition of “word”. Here it is, in simplified form (details can be found in the new paper):
“A word is (i) a free morph, or (ii) a clitic, or (iii) a root with its affixes, (iv) a compound with its affixes.”
The main feature of this definition is that it is disjunctive and thus clearly unnatural. But this is the whole point – in the absence of good reasons to think that words are natural parts of languages, the only way out is to provide an unnatural definition. Words are here defined as comparative concepts, not as natural kinds (or as language-particular descriptive categories).
(Compare this with the definition of “second” – defined with respect to Caesium frequency 9192631770 Hz; as is well-known, unnatural definitions of widely used comparative concepts are common in physics.)
The first two ingredients of the definition, “free” and “morph”, are intuitively natural: A free form is a form that can occur in isolation, and a morph is a minimal form (Haspelmath 2020). It is also quite intuitive to say that a root is a content-denoting morph (i.e. a morph that denotes an action, an object or a property). But clitics are words, too, and they are defined as bound morphs that are neither roots nor affixes – which is a rather strange definition (but something like this is widely assumed by linguists; nobody seems to believe that clitics can be defined positively).
Affixes are sometimes thought to be defined phonologically, but this does not work (as I noted in an earlier blogpost). Instead, they can be defined as nonroot bound morphs that always occur on roots of the same class (I used to call this “non-promiscuity”, but I switched to “word-class selectivity”). In addition, we need a definition of “compound”, which is possible on the basis of the notion of extendability of roots by root modifiers (Haspelmath 2024).
Finally, we need to distinguish between required and non-required affixes, but I will not comment on this here – it’s a novel conceptual distinction which I’m not very sure about. (Comments are welcome!)
My paper about defining the word is thus not a contribution to empirical linguistics, and it does not suggest any hypotheses. It is primarily a new attempt to come to grips with a difficult inheritance from our ancestors: the age-old distinction between “sentence patterns” (or syntax) and “word patterns” (or accidence, or morphology). This distinction may not be more than a stereotype carried on from earlier stages of Western linguistics, but we will not know until we formulate the opposite claim (that “words” are natural parts of languages) in a testable way.
But for the time being, I have at least a definition that I can use in a textbook. In my 2002 textbook, I gave the following definition (in the glossary, p. 276):
“[A word-form is] a ‘text word’ that can be isolated from surrounding elements because it is either prosodically independent (= a free form) or a clitic and not an affix.”
This definition does not work because many non-clitic words are not free forms, e.g. English count nouns such as tree (which need an article or a plural suffix) or English verbs such as put (which minimally need their obligatory complements).
References
Arkadiev, Peter & Gardani, Francesco (eds.). 2020. The complexities of morphology. Oxford: Oxford University Press.
Baerman, Matthew & Brown, Dunstan & Corbett, Greville G. 2017. Morphological complexity. Cambridge: Cambridge University Press.
Gardani, Francesco. 2020. Borrowing matter and pattern in morphology: An overview. Morphology 30(4). 263–282. (doi:10.1007/s11525-020-09371-5)
Gebhardt, Lewis. 2023. The study of words: An introduction. London: Routledge. (doi:10.4324/9781003030188)
Haspelmath, Martin. 2011. The indeterminacy of word segmentation and the nature of morphology and syntax. Folia Linguistica 45(1). 31–80. (doi:10.1515/flin-2017-1005)
Haspelmath, Martin. 2020. The morph as a minimal linguistic form. Morphology 30(2). 117–134. (doi:10.1007/s11525-020-09355-5)
Haspelmath, Martin. 2023. Defining the word. WORD 69(3). 283–297. (doi:10.1080/00437956.2023.2237272)
Haspelmath, Martin. 2024. Compound and incorporation constructions as combinations of unexpandable roots. To appear. (https://zenodo.org/record/8137251)
Newell, Heather & Noonan, Máire & Piggott, Glyne & Travis, Lisa deMena (eds.). 2017. The structure of words at the interfaces. Oxford: Oxford University Press.
Selkirk, Elisabeth O. 1982. The syntax of words. Cambridge, MA: The MIT Press.
Spencer, Andrew. 2019. Manufacturing consent over Distributed Morphology. Word Structure 12(2). 208–259. (doi:10.3366/word.2019.0146)
OpenEdition suggests that you cite this post as follows:
Martin Haspelmath (August 15, 2023). From radical critique to constructive proposal: “Defining the word” (2023). Diversity Linguistics Comment. Retrieved January 17, 2025 from https://doi.org/10.58079/nsx5
Why not define the word as a string of letters between spaces? This is the established meaning outside of linguistics, it conforms to intuition, and it recognizes the concept’s unnaturalness, as it is well known that written language is derivative.
Yes, that’s how we define a word if we have a written text. But languages are primarily systems for spoken or signed interaction, and writing plays no role at this primary level.
Yes, exactly. I believe that the word as a concept is by definition based on writing conventions, so it would be best to avoid it entirely when talking about morphosyntax. Though this seems to be a radical position, as you note.
A definite has to be independent of the exemplifications of what you are defining.
The indeterminacy of words is hard to dispute, but I think the boundary between morphology and syntax is much more determinate (at least if we define them as “what morphologists do” and “what syntacticians do.”) Work on syntax has no equivalent to studying the structure of IE inflectional paradigms or Semitic templates. Work on morphology has no equivalent to studying embedding or Wh-movement.
And then, aren’t those phenomena, and their clear segregation, emergent from some some widespread (if not “universal”) consistent difference between words and non-words?
I don’t think so. Embedding has clear counterparts in morphology, and that “wh-movement” does not occur in morphology is definitional – if we observe this sort of order variability, we call it “syntax”. It seems that the idea that morphological and syntactic patterns are strikingly different is a stereotype inherited from earlier Western linguistics and the practice of writing our languages with spaces.
How does embedding “have clear counterparts in morphology”? There is incorporation and compounding, but there are no open-ended hierarchical structures in [what is usually called] morphology.
Suppose we use “the study of the arrangement and form of mobile morphs” and “the study of the arrangement and form of immobile morphs”; I suggest that they would still, in great part, be dealing with distinct sets of phenomena, whose study would roughly correspond to the traditional terms “syntax” and “morphology”.
Compounding is open-ended in English, and so is some affixation (pre-pre-…). Some languages have iterable causative affixes, and so on. Maybe the two sets of phenomena are “in great part” distinct, but the same applies to, say, verbal and nominal syntax – they also have rather different properties. Or concatenative vs. nonconcatenative moprhology – they are also rather different. There does not seem to be any particular reason to privilege the syntax vs. morphology division.
I found it interesting that you note, towards the end, that you are satisfied if this disjunctive definition captures ~80% of prior uses. In a way, you simply choose a different place for the fuzziness that will always crop up somewhere. Instead of accepting that a term is somewhat vague but working with it (as in the philosophical approach) or trying to explain the vagueness somewhat (as in the prototype approach), you propose to sharpen the term itself, leaving the ~20% of vagueness by the wayside for others to worry about. I think it is a sensible approach by the way, and I appreciate the clear discussion of the five alternatives.
Thanks for this comment! Yes, you may want to say that “I simply choose a different place for the fuzziness that will always crop up somewhere” – but why will it always crop up? I don’t think we understand this, though we often observe it. Fuzziness is not inevitable – for example, all languages have negators and words for ‘head’, with no fuzziness. It seems to me that in p-linguistics, there is often no fuzziness, but when we compare languages, we want to map the p-categories of one language onto categories of other languages, which often “sort of” works. But in such cases, the “fuzziness” is really a consequence of our category-based comparison strategy. If we compare languages using idiosyncratic comparative concepts (concepts that have not been used as categories before), the problem does not arise. So I think that the perceived “fuzziness” may well be the consequence of the (highly problematic) expectation that languages have (more or less) the same categories.