Computational Complexity of Natural Morphology Revisited

Noam Chomsky argued in his seminal works (Chomsky, 1956, 1959) that human languages follow a set of computational rules. This naturally leads to the question: What level of computational power is required to process human languages?

This has been a dominant issue in theoretical computational linguistics, and a wide range of arguments have often been published in natural language processing (NLP) journals such as Computational Linguistics (Manaster-Ramer, 1988; Radzinski, 1991; Kuhlmann et al., 2015). The complexity question has two strands of significance.

The central tenet of the Chomskyan school is that human beings have special brain components dedicated to language processing, and thus certain linguistic (meta-)rules are innate in nature (Hauser et al., 2002). The veracity of this claim has been hotly debated; some neuroscientific studies support the idea (Fedorenko and Blank, 2020; Malik-Moraleda et al., 2022), whereas others are skeptical (Tremblay and Dick, 2016). Studies on complexity and the nature of the encoded linguistic properties are directly relevant here (Pullum and Gazdar, 1982; Chesi and Moro, 2014; Fang et al., 2021).

A classical topic is to investigate models and/or algorithms for parsing languages. Context-free problems are parsable using the Earley algorithm in O(n³) time, where n represents the length of a sentence (Earley, 1970). Generally, the strings of most natural morphology can be generated by 1-way finite state transducers (FSTs) (Roark and Sproat, 2007). However, in many cases, more expressive models such as 2-way FSTs are required to capture its strong¹ generative capacity (Dolatian et al., 2021). Recently, there has been growing interest in exploring the patterns and biases learned by neural networks (Weiss et al., 2018; Dolatian and Heinz, 2019; Torres and Futrell, 2023).

Several languages have been examined to assess the weak generative capacity of formal grammars for natural languages. It is important to note that a language is not strongly context-free if it can or cannot be generated by context-free grammars (CFGs); however, if it can be, there exist some parse trees that are not interpretable by linguistic theories. A language is not weakly context-free if it cannot be generated by any CFGs. Therefore, the denial of weak context-freeness is a more stringent condition than that of strong context-freeness.

Extensive research has been conducted on the topic of syntax (see Section 2.1). However, when it comes to morphology, there appears to be a dearth of substantial arguments beyond a paper discussing the vocabulary of Bambara (Culy, 1985). The paper resorted to the fact that a Bambara word construction “wulu(filèla)^h(nyinina)ⁱowulu(filèla)^j(nyinina)^k” (h, i, j, k ≥ 1) forms a valid noun if and only if h = j and i = k, and is otherwise ungrammatical. The Bambara noun roughly means “whoever watches who watches who watches …who watches who searches for who searches for …who searches for dog searchers.”

However, the wordhood of these strings is entirely reliant on tonal behaviors. It was noted that Bambara exhibits two types of tonal rules: one for external sandhi (“the interaction of adjacent lexical items,” that is, syntactic) and another for internal sandhi (“the interaction of components of a compound,” that is, morphological). Culy further states that the Noun o Noun construction shown above has its own special tonal behavior, which is not syntactic. He thus concluded that this is morphological.

Manaster-Ramer (1988, pp. 101–102) questioned this point, arguing:

1. The paper does not provide specific tonal rules but instead refers to a Bambara textbook. The behavior may imply the exact opposite of Culy’s claim: It is a special behavior which is not morphological, and is thus syntactic.

2. Irregular behaviors of external sandhi are commonly seen in the world’s languages, and thus tonal behaviors are not suitable as an absolute criterion for word boundaries.

In reality, given that the boundary between morphology and syntax is not always clear, even in sophisticated linguistic theories, it can be problematic to rely on criteria such as “this phenomenon generally indicates wordhood,” whether phonological or lexical. How can one consciously establish that a construction is morphological rather than syntactic?

Our proposed solution is to create a construction that becomes explicitly ungrammatical when interpreted as a syntactic sequence.

In our research, we focus on Classical Ainu, an endangered language spoken by the Ainu, an indigenous people of the Island of Hokkaido in Japan who originally inhabited islands around the border of what is now Japan and Russia. The 17th edition of Ethnologue labeled Ainu as an 8b or nearly extinct language (Lewis et al., 2013). Constant efforts are being made to revive this language (DeChicchis, 1995; Sato, 2012b; Ōno, 2022). No language has been confirmed to be genetically related to Ainu; thus, it is classified as a language isolate.

Classical Ainu, a language known for its highly polysynthetic nature, has a complex valency-computation system. The language can productively form new verbs by concatenating a variety of nouns (noun incorporation, NI) and affixes. In some cases, a single verb can express information that is almost equivalent to a sentence in other languages. Valency, or the number of arguments a verb can take, is recalculated along with this concatenation process. Several morphemes increase or decrease valency, whereas others constrain it. Using these properties, we devised a construction that was grammatical (but not weakly context-free) at the morphological level and ungrammatical at the syntactic level.

The rest of the paper is organized as follows. In Section 2, we summarize existing studies on the topic. Section 3 provides linguistic data to prepare our proof. Section 4 presents the construction and proof. In Section 5, we discuss the issue from various perspectives. Finally, in Section 6, we offer concluding remarks.

Chomsky first formulated a hierarchy of computational complexity in terms of human languages and investigated whether it exceeded the level of context-free languages (Chomsky, 1956, 1959). Although several arguments have been made, including those on NI in Mohawk (Postal, 1964), almost all have been rejected by Pullum and Gazdar (1982). After nearly 30 years of expeditions, linguists have finally discovered actual cases of non-context-free properties: the cross-serial dependency in the Swiss-German syntax (Huybregts, 1984; Shieber, 1985) and the reduplication of the Bambara morphology (Culy, 1985). Pullum (1986) provided a stimulating history of this period. Moreover, Swedish is not context-free because of the extraction phenomena and obligatory presence of resumptive pronouns (Miller, 1991).

After the discovery of non-context-freeness, the focus shifted to whether human languages were acceptable by mildly context-sensitive grammars (MCSGs) (see the next section). Attempts to prove the weak inadequacy of MCSGs include the following: crossing-dependency in Dutch (Manaster-Ramer, 1987; Groenink, 1997); the number-naming system in Mandarin (Radzinski, 1991); Suffixaufnahme in the Old Georgian syntax (Michaelis and Kracht, 1997), which was counter-argued by Bhatt and Joshi (2004); and recursive copying in the syntax of Yoruba and its kin languages (Kobele, 2006).

In the field of phonology, Heinz (2010) claimed that phonological patterns are subregular (later known as the Subregular Hypothesis).

Among the earliest studies on formal languages, Langendoen’s (1977) work is most relevant to our study, because the technique that he developed has been used in many subsequent proofs (Shieber, 1985; Culy, 1985; Miller, 1991), including ours.

Joshi (1985, p. 225) hypothesized three universal and invariant properties that every natural language must have: 1. limited cross-serial dependencies, 2. constant growth, and 3. polynomial parsing. Joshi named the class of beyond context-free formal grammars that satisfy all these properties as MCSGs. This includes tree-adjoining grammars (TAGs), linear indexed grammars (LIGs), combinatory categorial grammars (CCGs), head grammars (HGs), multiple context-free grammars (MCFGs), and Stabler’s (1996) formalism of the minimalist program. This class is significant because, from a cognitive perspective, it accounts for natural languages well, and from an engineering perspective, it is very efficient for parsing. Vijay-Shanker and Weir (1994) proved the weak equivalence of TAGs, LIGs, CCGs, and HGs. Kuhlmann et al. (2015), however, found modern versions of CCGs are in fact less powerful than TAGs. Furthermore, the work by Kanazawa and Salvati (2012) and Salvati (2015) challenges the coherence of MCSGs.

Whereas the complexity of a linguistic construction was traditionally studied in terms of formal languages, recent studies attempt to examine the complexities of languages from the perspective of statistical metrics. For instance, Cotterell et al. (2019) showed inflectional paradigms could not be both large and highly irregular at the same time.

Park et al. (2021) revealed that difficulty to obtain a good language model (LM) was positively correlated with several statistical complexity measures of the language’s morphology. Morphologically simple languages are better modeled by general tokenizers such as BPE, while moderately complex languages benefit from linguistically motivated ones such as Morfessor, although they are outperformed by simple unigram for polysynthetic languages. Zevallos and Bel (2023) further investigated this nature and demonstrated a method to reduce the amount of language data to obtain good LMs for low-resource languages.

This section provides real-world examples in a natural language that supports the validity of a set of formal rules used in Section 4.

Similar to the case of the defunct language Old Georgian, which Michaelis and Kracht (1997) used in their complexity study, today it is practically impossible to directly assess the grammaticality of Classical Ainu words with native speakers because this language is highly endangered (Sato, 2012b; Lewis et al., 2013). cTherefore, we collected linguistic data to support the construction as far as possible.

Refsing (1986) wrote the first modern grammar book on Ainu, which is also an essential source on the Shizunai dialect. The first part of a book by Shibatani (1990) is by far the most widely referenced Ainu grammar text. An article by Tamura (1988), whose English translation is Tamura (2000), provides more detailed, precise accounts. It also includes a short survey on the linguistic studies of Ainu from the 18th century to the 1980s. Although written in Japanese, a textbook by Sato (2008) provides plenty of Ainu utterances (both colloquial and Classical) obtained directly from native speakers of the Chitose dialect. Each chapter of a handbook edited by Bugaeva (2022b) delineates linguistically interesting phenomena by an expert on the topic.

Among modern dictionaries, Tamura’s (1996) Ainu-Japanese dictionary on the Saru dialect is the most comprehensive one. In addition, Nakagawa’s (1995) Ainu-Japanese dictionary on the Chitose dialect provides insightful accounts of many complex phenomena in the language. Whereas Kayano’s Ainu-Japanese dictionary (Kayano, 2002) on the Saru dialect lacks linguistic information, such as parts-of-speech and the valency of verbs, it is a valuable resource for Ainu, as the author himself was a native speaker. Kirikae’s (2003) dictionary/commentary on Ainu Shin’yoshu, notable kamuyyukar-type poems, serves as a solid corpus for the Classical Ainu language. Bugaeva’s (2011) colloquial Ainu dictionary has detailed English glossings and recorded utterances from a native speaker.

For interlinear glossing, unless otherwise noted, we follow the Leipzig Glossing Rules and Standard Abbreviations, rev. February 2008 (Comrie et al., 2008), with the following redefinitions and additions; int for intensive, itera for iterative, mid for middle voice, mono for monovalent, poly for polyvalent, and poss for the possessed case (not possessive).

Traditionally, the Ainu language had no writing systems. From the early 20th century, two types of systems have been co-developed: one with the modified Japanese katakana and the other with the Latin alphabet.

Linguistically, Ainu is characterized by agglutinativity, polysynthesis, and NI. Although the latter two properties are not so overt in colloquial Ainu, they are fully employed in Classical Ainu, a variant mainly used for epics, tales, poetry, songs, incantations, rituals, and other artistic or religious activities.

In Classical Ainu, with NI, one can rephrase a sentence “I make an inaw (a wooden prayer symbol)” as “I inaw-make” (Shibatani, 1990, p. 63; glossing follows him, with bracketed supplements):

• (1) 

  a.  Inaw a-ke

•      [prayer] 1sg-make

•      ‘I make a wooden prayer symbol.’

•     b.  Inaw-ke-an

•      [prayer-]make-1sg

Notice the difference in the pronominal affixes: the prefix a- is used in (a) and the suffix -an in (b). As mentioned previously, Ainu is sensitive to valency, or the number of arguments that can be taken by verbs. For a first person subject, one must use the suffix -an for monovalents (so-called intransitives) and the prefix a- for polyvalents (so-called transitives).²ke is a bivalent, so a polyvalent prefix a- is used. Conversely, inawke is a monovalent formed by NI; because the incorporated noun filled a direct object slot of the original verb ke, its valency decreased by one, so inawke takes a monovalent suffix -an. Neither *inaw kean³ nor *ainawke is grammatical.

In addition to NI, Ainu has another powerful tool called applicative formation (AF). Each time a verb is prefixed by an applicative, its valency increments by one and a newly formed verb accommodates an additional argument as a direct object. For instance (Shibatani, 1990, p. 65; glossing follows him):

• (2) 

  a. A-kor kotan ta sirepa-an.

•      1sg-have village to arrive-1sg

•      ‘I arrived at my village.’

•    b. A-kor kotan a-e-sirepa.

•            1sg-appl-arrive

Again, notice the differences in the pronominal affixes. The monovalent sirepa (“to arrive”) requires the suffix -an. When it takes the allative applicative prefix e- and forms a new bivalent esirepa (“to arrive at X”), its valency increments by one. Thus, it requires the prefix a-.

Its polysynthetic nature allows various affixes and affixal operations. Table 1 lists major verbal affixes. By combining NI, AF, and the other morphemes, it is possible to construct a complex (but still single) word, such as aeyaykotuymasiramsuypa (Shibatani, 1990, p. 72; glossing follows him):

• (3) 

  a-e-yay-ko-tuyma-si-ram-suy-pa.

•   1sg-appl-refl-appl-far-

•   refl-heart-sway-itera

•   ‘I keep swaying my heart afar and toward

    myself over (something)’ = ‘I wonder about

    (something)’

Readers may question how it is possible to know that such a complex verb is a single word, rather than a phrase with multiple words. In the case of Ainu, the following two criteria are informative:

• Phonological Criterion. 

  Ainu is a pitch-accent language. In modern orthography, a high pitch syllable is marked by an acute on its vowel. The position of a high vowel is distinctive, e.g., nísap (“suddenly”) and nisáp (“shin”) (Tamura, 1996, p. 423). Additionally, the mechanism for determining the pitch position is fairly regular. For almost all words, the second vowel is accented if the first syllable is open (as in kamúy, “god”), whereas the first vowel is accented if the first syllable is closed (as in áynu, “human”). This mechanism is also applied when a new word is formed by affixes. For example, from sapá (“head”), kusápaha (“my head”) is formed (Tamura, 1988, p. 13).⁴ In summary, a hearer can usually determine word boundaries by detecting the positions of high pitches.

• Lexical Criterion. 

  Several affixes, such as the reflexive prefix yay- and the aspectual suffix -kosanu, attach only to a verb. Therefore, the presence of these affixes suggests that the chunks are classified as a verb. Pronominal affixes are also useful markers for this purpose, because they always occur as the last elements of the verb formation process and directly indicates the boundaries of the words.

However, these criteria are not always applicable.

As for the first criterion, we seldom have recordings. In addition, in human language, pitch positions can be changed by factors such as word length and individual habits.

As for the second criterion, it is not always possible to find salient pronominal affixal strings, because the most frequent one is the third-person pronoun, that is, the empty string ∅. Furthermore, several nouns can take pronominal affixes, in which case it is difficult to determine whether they are incorporated nouns or direct object nouns.

However, at least under some constructions, it is possible to unambiguously determine the wordhood, as seen in Section 2 (such as that used in Example 2, devised to satisfy the above lexical criterion).

Ainu has three types of applicative prefixes: e-, o-, and ko-. Most of their functional differences (and similarities) are empirically well known (Tamura, 1996). Bugaeva (2010) derived various statistics on the usage of these prefixes, and uncovered that several semantic roles are shared by more than one applicatives. For example, she showed that both e- and ko- have the same semantic role of Cause/Purpose (6% of e- and 3% of -ko occurred in her corpus).

Shibatani (1990, pp. 76–77) suggested the following ordering of verbal prefixes: 1. subject prefix, 2. object prefix, 3. applicative, 4. generalized object or reflexive or reciprocal, 5. applicative, and 6. verb.

This formulation limits the number of applicatives to two, and is in fact applicable to a large number of verbs. However, several verbs that were collected from informants exceed this limit, indicating that the formulation above is not imperative. For example, the bivalent eyaykewutum-ekosanniyo is formed from the monovalent sanniyo with three applicatives, a NI, and a reflexive. The following instances were taken from Tamura (1996):

• (4) 

  a.  eyaykewtum-ekosanniyo (p. 153)

•     e-yay-kewtum-e-ko-sanniyo

•     appl-refl-heart-appl-appl-know

•     ‘to think in one’s heart about something’

•     ewkoyaykopuntek (p. 146; kopuntek: p. 332)

•     b. e-u-ko-yay-ko-puntek

•     appl-recp-appl-refl-appl-be.pleased

•     ‘(for several people) to be pleased with something together’

•     c. eyaykouwepeker (p. 155; uwepeker: p. 808)

•     e-yay-ko-u-e-peker

•     appl-refl-appl-recp-appl-be.clear

•     ‘to think over one’s troubles’

Now, we examine the possibility of using the same nouns multiple times. As Classical Ainu poetry values periphrasis and verbosity, it is possible to form a tautological statement such as “I am holding a red staff as a staff”, as in Kirikae (2003, p. 140):

• (5) 

  hure kuwa ekuwakor

•    húre kuwa∅-e-kuwa-kor

•    be.red staff 3-appl-staff-have

•    ‘She staff-held a red staff.’

If the object húre kuwa is incorporated, a new verb may be húrekuwaekuwakor. In fact, there is at least one case where the same noun is incorporated twice, as in orupkorupus (Sato, 2012a, p. 15).

• (6) 

  orupkorupus

•    o-rup-ko-rup-us

•    groin-ice-APPL-ice-put

•    ‘(for one’s groin) to be freeze-frozen’

As we have seen, it is semantically tautological but still grammatical to incorporate the same noun more than once as a form of poetical rhetoric.

Ainu has a rich reduplication system. For verbs, Tamura (1988, pp. 65–66) identified three types of reduplication.

1. The full reduplication of stems or roots functions as an iterative and/or intensive aspect. For example, kik “to hit once” and kikkik “to rain punches”.

2. The reduplication of VC in CVC functions as a progressive and incessant aspect. For example, cirir (cir-ir) “(for a stream) to trickle”.

3. The reduplication of the last CV in CVCV functions as a progressive and intensive aspect. For example, sikcupupu (sik-cupu-pu) “to keep on narrowing one’s eyes”.

Though the second and third types are limited to certain cases, Tamura (1988) claimed that the first one is productive. It is also possible to reduplicate a verb formed using NI (Kirikae, 2003, pp. 109, 265, 315)⁵ :

• (7) 

  a.  hokushokus

•       ho-kus ∼ hokus

•       tail-pass.through∼INT

•       (lit.) ‘to pass through one’s buttocks

        severely’ = ‘fall down, overturn’

•    b. chipokonannpe kohokushhokush

•       ∅ -cip-o    kor     ne an pe

•       3-boat-ride and.then it

•         ∅ -ko-hokushokus

•         3 >3-appl-overturn

•       ‘They boarded a boat, and then they rowed it hard as if it overturns.’

Additionally, it is possible to reduplicate a verb formed by AF (Kirikae, 2003, pp. 107, 319):

• (8) 

  a.  kosankosan

•       ko-san ∼ kosan

•       appl.go.downstream∼ int.itera

•    b. chikor wenpuri unkosankosan

•       ci-kor     wen     puri

•       1sg-have    bad     habit

•        un-kosankosan

•         3 > 1sg-go.downstream.int.itera

•         ‘My evil thoughts flowed into me

        impulsively.’

Note that reduplication does not change the valency. Both hokus and hokushokus are monovalent (kohokushokus is bivalent because of AF), and both kosan and kosankosan are bivalent.

Example (8) also shows that its reduplication is morphological and not syntactic. If syntactic, then its form is unkosan unkosan. Because Classical Ainu also favors the syntactic reduplication of a verb, in some cases, it is ambiguous whether the process is morphological or syntactic. However, the presence of some verbal affixes uniquely determines the level of reduplication.

The high polysynthesis of Classical Ainu reaches the point where the language provides various idioms of affixes. The previous examples demonstrated several of them: yay-ko (refl-appl = “by oneself” or “alone”) and u-ko (recp-appl = “together”).

These idioms can form circumfixes. For example, the periphrastic circumfix ci-...-re is the combination of the middle voice prefix ci-⁶ and the causative suffix -re⁷ (Tamura, 1996, p. 48). Semantically, the idiom is meaningless. It is a poetic device used only to embellish verbs in an elegant and graceful tone.

This idiom does not change valency, as the decrement by ci- and increment by -re cancel each other out. The idiom is also unique in that ci- alone affixes to polyvalents but not to monovalents, whereas ci- -re can be affixed to both monovalents and polyvalents. For example, the monovalent cihopunire is derived from the monovalent hopuni (Tamura, 1996, p. 48):

• (9) 

  cihopunire

•   ci-hopuni-re

•   mid-happen-caus

•   ‘(for a serious matter) to happen’

In contrast, *cihopuni is ungrammatical. Such idiomatic circumfixes are useful for identifying word boundaries, because they encircle a verb and prevent it from further concatenating with other morphemes.

To summarize, in Classical Ainu,

• applicatives have overlapping semantic roles (this property was used in Example 1),

• there is no limit on how many times applicatives can be affixed (used in Example 1),

• the same noun can be incorporated iteratively (used in Example 1),

• a verb can be fully reduplicated (used in Definition 3 and Example 2), and

• an affixal idiom may form a circumfix, a useful tool for identifying word boundaries (used in Example 2).

We define ad-hoc lexical models that are intended to capture the properties of the Classical Ainu lexicon. More refined approaches may model the lexical semantics of verb arguments (Comrie et al., 2015; Bugaeva, 2022a, pp. 43–44); however, it is beyond the scope of this study.

There may be several lexemes (entries) for one string. For instance, in English, the verb “give” has three entries (give,1, ∅,head), (give,2, ∅,head), (give,3, ∅,head), which correspond to monovalent (intransitive), bivalent (monotransitive), and trivalent (ditransitive), respectively.

For example, given a bivalent (monotransitive) verb (kar,2, ∅,head) meaning “to make something”, and a noun $(cise,−1,[2,∞),free)$ meaning “house”, a new monovalent (intransitive) verb (cisekar,1, ∅,head) meaning “to build a house (lit. to house-make)”⁸ can be formed using an operation called noun incorporation.

It is also possible to form an avalent verb; e.g., the concatenation of $(sir,−1,[1,∞),prefix)$ “environment” and (sések,1, ∅,head) “to be hot” forms an avalent (sirsesek,0, ∅,head) “it is hot”.⁹ In English, semantically avalent expressions (e.g., “it is hot” and “it rains”) must take dummy subjects to satisfy syntactic constraints, but Ainu does not have such constraints. The maximum valency of Classical Ainu is not known, but Bugaeva (2015, p. 828) reported that there is at least one tetravalent verb (korere,4, ∅,head) “to make someone give something to someone.”

The last verb should mean “to be unable to see because of tears,” which is the same as núpeesiknak. For the reasoning behind this derivation, refer to Sections 3.4.1, 3.4.2, and 3.4.3.

The unambiguously morphological construction xw_{m, n}w_{m, n}y exhibits a symptom of non-context-freeness. Nevertheless, we do not directly prove that this is the case. Even if a subset of a language is not context-free, it is not always true that the entire language is not context-free. To prove the non-context-freeness of the vocabulary of Classical Ainu, we used the fact that context-free languages are closed under intersection with regular languages. This trick is often attributed to Bar-Hillel et al. (1961), but the mathematically rigorous form was developed by Langendoen (1977) as an improvement upon the work of Bar-Hillel et al. It was used in subsequent studies, including Culy (1985).

After Mithun (1984) analyzed NI, the related literature has become “a microcosm of linguistic theory” (Massam, 2009) because NI lies at the interface of morphology and syntax.

Mithun stated that NI is “perhaps the most nearly syntactic of all morphological processes” (Mithun, 1984, p. 847), and went on to say that “Formally, it is a morphological process, not a syntactic one; and it shares all the characteristics unique to such process” (Mithun, 1984, p. 891). The fact that syntax is not context-free suggests that morphological processes with NI are also not context-free. We demonstrated that this is true.

She also reported an interesting story about the mental lexicon. Mohawk speakers remember NI that has actually occurred, and feel excited when hearing a new (and grammatical) NI that they had never heard before, because complex NI is considered an elegant art (Mithun, 1984, pp. 872, 889).

However, as a literary register, the products of Classical Ainu are usually memorized and recited, rather than freely produced. Therefore, it may raise the question whether it is a valid tool to test the linguistic capacity of humans.

Concerning this issue, we point out folklore collected by Nakagawa (2006), which has a unique style that combines uwepeker (prose literature) and kamuyyukar (a type of poetry). The storyteller Nabe Shirasawa explained that she felt sakehe (refrains) used in the first part of the original poem were rustic and boring, and therefore she reframed that part in the prose form.

This suggests that though the Classical register is often used in works whose forms are fixed, native Ainu poets can freely tell tales in Classical Ainu if the situation demands it.

The lexicality of NI is a subject of classical debate (Kroeber, 1910; Sapir, 1911; Kroeber, 1911). Even today, some linguists who advocate transformational types of grammar (such as the government and binding theory) claim that NI is syntactic rather than morphological (Baker, 1988, 1996). This argument is important in the literature, because head-movement, one of the major syntactic operations in their theory, is built upon the syntacticity of NI (Anderson, 2000). Anderson (2000) criticized this analysis as being reasonable for their theory, but not in terms of accounting for NI.

Moreover, their claim comes at the expense of the Lexicalist Hypothesis (Anderson, 2000). This foundational hypothesis, attributed to Chomsky (1970), states that words are the basic and atomic units of syntax. It is also formulated as “The syntax neither manipulates nor has access to the internal form of words” (Anderson, 1992, p. 84). As the Lexicalist Hypothesis is presupposed in various theories, including modern NLP frameworks such as Universal Dependencies (Nivre, 2015, p. 6), interpreting NI as syntactic operations abolishes not only our argument but also many of the existing models in computational linguistics. Thus, it should be reasonable to follow the classic hypothesis and maintain that NI (and the construction we have devised) is morphological.¹⁵

This study demonstrates that Classical Ainu is not weakly context-free unambiguously at the morphological level because of its valency-sensitive structure and valency-changing operators. Thus, we reconfirmed the weak inadequacy of CFGs for natural morphology.

However, the properties of valency, polysynthesis, NI, and AF have not yet been fully explored in computational linguistics. Further examinations may provide more insight into natural languages and aid cognitive and engineering studies.

The authors express their sincere appreciation to the action editor and the three anonymous reviewers for their extensive and insightful feedback. We also extend deep gratitude to Makoto Kanazawa, who provided valuable comments about formal language studies.