Ergativity for Novices

[cedh]

Also, it's worth pointing out the ergativity in English. (i.e. My reading vs Troy's destruction by the Greeks). In English, action nominals seems to have a split intransitive alignment. Compare above with "My singging the Marseillaise"

But you can just as well say "The Greeks' destruction of Troy", with an agentive interpretation. The only thing that's different is the preposition used with the oblique argument, and the semantics of that preposition cause the subject argument to have a different default role. That's not really any significant level of ergativity, just two distinct nominalizing constructions, of which the one with "by" happens to have a semantically patientive argument structure.

[Pabappa]

You could also say just "Troy's destruction" though . Or "the girl's murder" etc if it is known from context the girl is not the perpetrator.

[bradrn]

Also, it's worth pointing out the ergativity in English. (i.e. My reading vs Troy's destruction by the Greeks). In English, action nominals seems to have a split intransitive alignment. Compare above with "My singging the Marseillaise"

This could certainly be interpreted as an ergative construction, since here S and O appear to be treated similarly with regards to word order. However, I would concur with cedh and Pabappa that this construction is not actually ergative (and for that matter, it isn’t accusative either). My reasoning here is that that nominalization can be restated in any of the following ways, only one of which equates S and O:

1. Troy’s destruction by the Greeks (S=O)
2. The Greeks’ destruction of Troy (S=A)
3. The destruction of Troy by the Greeks (S≠A, S≠O)

For me, there is only a minor semantic difference between these: (1) focuses more on the resulting state, whereas (2) focuses more on the action, while (3) focuses on neither. But I would say that this difference is more akin to a difference in voice than to a difference in alignment. (In fact, by your reasoning, the passive voice would also be considered to be ergative: [I]_S read vs [Troy]_O was destroyed by [the Greeks]_A.)

EDIT: The WALS chapter on nominalizations calls construction (1) an ‘ergative-possessive’ construction, and calls construction (2) a ‘possessive-accusative’ construction. So there is at least one major source which agrees with you in saying that (1) can be seen as ergative in some way. Also, I feel I should justify why the passive is not in fact an ergative construction: it is because a passivised verb becomes intransitive, so has only an S argument rather than O and A arguments.

[chris_notts]

The distribution of ergativity

Although not particularly relevant for conlanging, it can sometimes be useful to know where in the world ergative languages occur:
In Europe, almost all languages are nominative-accusative. The only exceptions are Basque, in western Europe, and the three Caucasian families (Northwest Caucasian, Northeastern Caucasian/Nakh-Dagestanian, South Caucasian/Kartvelian) of northeastern Europe.

Minor nitpick: surely that should be south-east?

[chris_notts]

EDIT: The WALS chapter on nominalizations calls construction (1) an ‘ergative-possessive’ construction, and calls construction (2) a ‘possessive-accusative’ construction. So there is at least one major source which agrees with you in saying that (1) can be seen as ergative in some way. Also, I feel I should justify why the passive is not in fact an ergative construction: it is because a passivised verb becomes intransitive, so has only an S argument rather than O and A arguments.

Or to say it another way, ergativity is a pattern across "equivalent", normally basic intransitive and transitive constructions. Comparing the passive a normal English intransitive clause isn't the correct comparison.

This does raise the obvious question of how to treat languages with more than one transitive construction. There are a few obvious examples:

1. The typical Algonquian direct - inverse scheme

verbal agreement driven mainly but a person hierarchy but also 3rd person relative topicality, with voicing marking for subject > object vs object > subject on said hierarchy

2. Austronesian type alignment

Verbal morphology to mark one of many semantic roles (e.g. actor, patient, location, instrument, recipient, ...) as the trigger. Often the actor retains certain properties that suggest it is not an oblique even in the other voices, meaning that in some analyses there may be many transitive constructions.

3. The A- and O-constructions of some Arawan languages like Jarawara

There are two transitive constructions: selection is driven by discourse prominence and inter-clausal reference. Interestingly, an O-construction ("inverse") is only possible if the O is 3rd person or in certain morphologically limited conditions, which is not what would be expected if the purpose is to promote SAPs. An A-construction clause allows its A to be freely omitted, especially if coreferential with the main argument of a preceding coordinated clause, but the O must normally be explicit. The opposite is the case for an O-construction clause.

An example from Jarawara:

Jara ati na-re-ka
Branco(m) speak AUX-IPem-DECm
The branco spoke (intransitive clause)

ati otaa haa.haa ka-ne-hiba-no-ho
language 1.exc laugh APPLIC-AUX-FUTm-IPnm-DEP
And we laughed at his words (= the branco's words, O-construction used because the previous S carries over into O function, even though the A is 1st person)

Do these count as split ergative, or should all but one of the transitive constructions be considered less basic? If so, on what basis? Is it based on which transitive construction shows the greatest morpho-syntactic similarity to the intransitive construction, or on which shows the greatest pragmatic similarity (highest discourse frequency, most pragmatically neutral, ...)? Of course, the traditional answer has been to say that these are hierarchical systems and therefore neither accusative or ergative, but it's not obvious why the same argument doesn't apply to split-ergative case marking, since the functional motivation seems to be similar. The main difference is that languages that go down the voice route tend to be head marking, and those that go down the case route tend to be dependent marking.

[chris_notts]

You could also say just "Troy's destruction" though . Or "the girl's murder" etc if it is known from context the girl is not the perpetrator.

To me, using "X's murder" to mean the murder X committed is by far the more pragmatically marked option, and needs contextual reinforcement to work. If I hear "Jack the Ripper's murders", I know we must be talking about the murders he committed, but if I heard "John's murder", where John's some random guy I don't know well, I'd 100% assume John was dead, not a killer.

Same thing with "X's destruction". I think the rule is:

1. If the action is some form of telic state change (murder, destruction, boiling, ...), and there's only one possessor, then it's a patient by default
2. If the action is atelic with no permanent effect, then it's probably an agent by default (e.g. John's singing)
3. If there's both a pre- and post-nominal possessor, then the pre-verbal one can be an an agent

[bradrn]

The distribution of ergativity

Although not particularly relevant for conlanging, it can sometimes be useful to know where in the world ergative languages occur:
In Europe, almost all languages are nominative-accusative. The only exceptions are Basque, in western Europe, and the three Caucasian families (Northwest Caucasian, Northeastern Caucasian/Nakh-Dagestanian, South Caucasian/Kartvelian) of northeastern Europe.

Minor nitpick: surely that should be south-east?

No, I don’t think so: the Caucasus looks pretty north-easterly to me, when I look on a map. (South-east would be Bulgaria or Greece, surely.)

EDIT: The WALS chapter on nominalizations calls construction (1) an ‘ergative-possessive’ construction, and calls construction (2) a ‘possessive-accusative’ construction. So there is at least one major source which agrees with you in saying that (1) can be seen as ergative in some way. Also, I feel I should justify why the passive is not in fact an ergative construction: it is because a passivised verb becomes intransitive, so has only an S argument rather than O and A arguments.

Or to say it another way, ergativity is a pattern across "equivalent", normally basic intransitive and transitive constructions. Comparing the passive a normal English intransitive clause isn't the correct comparison.

That’s certainly another way to say it, although I must confess it took me a couple of minutes to figure out what you were saying. (I think you’re saying that we should only compare ‘underived’ transitive and intransitive clauses, so e.g. no comparison of active to passive clauses.)

This does raise the obvious question of how to treat languages with more than one transitive construction. … Do these count as split ergative, or should all but one of the transitive constructions be considered less basic? If so, on what basis? Is it based on which transitive construction shows the greatest morpho-syntactic similarity to the intransitive construction, or on which shows the greatest pragmatic similarity (highest discourse frequency, most pragmatically neutral, ...)? Of course, the traditional answer has been to say that these are hierarchical systems and therefore neither accusative or ergative, but it's not obvious why the same argument doesn't apply to split-ergative case marking, since the functional motivation seems to be similar. The main difference is that languages that go down the voice route tend to be head marking, and those that go down the case route tend to be dependent marking.

I would say that if a language has more than one ‘underived’ transitive construction (to use my term), then that constitutes a completely separate system of alignment. We happen to classify some of these alignment systems together as ‘split ergative’, due to the way they have many characteristics in common (most importantly the fact that they have an ergative subsystem, but I will cover others in the forthcoming section on split ergativity), but there certainly exist other alignment systems which have multiple transitive constructions (you mention direct/inverse and Austronesian alignments).

1. The typical Algonquian direct - inverse scheme … with voicing marking for subject > object vs object > subject on said hierarchy

I thought direct/inverse was a different, mostly unrelated phenomenon to voice. (Especially since voicing is a way of converting a transitive to an intransitive clause, but the inverse construction remains transitive.)

3. The A- and O-constructions of some Arawan languages like Jarawara

Jarawara is an interesting case. It certainly looks a bit like a separate alignment, but I think I’m inclined to classify it as another variety of split-ergativity. (It reminds me a bit of the system of Lhasa Tibetan, actually, although I’d need to review that system to be sure. You can find a description of Lhasa Tibetan in that Blue Bird book linked earlier.)

You could also say just "Troy's destruction" though . Or "the girl's murder" etc if it is known from context the girl is not the perpetrator.

To me, using "X's murder" to mean the murder X committed is by far the more pragmatically marked option, and needs contextual reinforcement to work. If I hear "Jack the Ripper's murders", I know we must be talking about the murders he committed, but if I heard "John's murder", where John's some random guy I don't know well, I'd 100% assume John was dead, not a killer.

Same thing with "X's destruction". I think the rule is:

1. If the action is some form of telic state change (murder, destruction, boiling, ...), and there's only one possessor, then it's a patient by default
2. If the action is atelic with no permanent effect, then it's probably an agent by default (e.g. John's singing)
3. If there's both a pre- and post-nominal possessor, then the pre-verbal one can be an an agent

I think all of this is the same for me as well.

[chris_notts]

No, I don’t think so: the Caucasus looks pretty north-easterly to me, when I look on a map. (South-east would be Bulgaria or Greece, surely.)

The place is level with Bulgaria and Romania! Only slightly north of Greece.

https://www.google.com/maps/place/Cauca ... 51814!4d45

I thought direct/inverse was a different, mostly unrelated phenomenon to voice. (Especially since voicing is a way of converting a transitive to an intransitive clause, but the inverse construction remains transitive.)

Voice doesn't just change transitivity, or at least many things that people call voice don't. For example:

1. There are a number of transitive "middle" verbs in languages like Ancient Greek, Fula, ...
2. Applicatives are normally considered to be voice, and when applied to transitive roots they may not affect valency at all (in some languages they just replace one object with a different object)

Depending on who's writing, voice seems to cover some mixture of:

1. Valency adjustment, both increasing (causative) and decreasing (passive, anti-passive)
2. Argument rearrangement (regardless of valency)
3. Verbal marking of argument roles and affectedness (middle voice, version ..., which may not change valency at all but do mark subject affectedness, e.g. "he built himself a house" vs "he built a house", or similarly with cutting hair, washing parts of one's body, ...)
4. Verbal marking of pragmatics and promotion to privileged roles for coreference and control purposes (again, may not involve detransitivisation)

Jarawara is an interesting case. It certainly looks a bit like a separate alignment, but I think I’m inclined to classify it as another variety of split-ergativity. (It reminds me a bit of the system of Lhasa Tibetan, actually, although I’d need to review that system to be sure. You can find a description of Lhasa Tibetan in that Blue Bird book linked earlier.)

But it's pretty much the same as a direct-inverse system, surely? The drivers for construction choice are a bit different (person vs pragmatics and control) and the morphology is a bit messier than a prototypical direct-inverse system, but I'm not sure why one is split ergative but not the other.

[Pabappa]

i would group Bulgaria & Romania as part of the Balkan sprachbund and thus definitely southeastern Europe. Linguistically, the Caucasus is pretty much in a world of its own and not SAE at all. Culturally, they are on the very fringe of Europe, and really can be considered as much part of the Middle East as of Europe.

[bradrn]

No, I don’t think so: the Caucasus looks pretty north-easterly to me, when I look on a map. (South-east would be Bulgaria or Greece, surely.)

The place is level with Bulgaria and Romania! Only slightly north of Greece.

https://www.google.com/maps/place/Cauca ... 51814!4d45

i would group Bulgaria & Romania as part of the Balkan sprachbund and thus definitely southeastern Europe. Linguistically, the Caucasus is pretty much in a world of its own and not SAE at all. Culturally, they are on the very fringe of Europe, and really can be considered as much part of the Middle East as of Europe.

Alright, then. How about I reword it to:

In Europe, almost all languages are nominative-accusative. The only exception is Basque, in western Europe. The three Caucasian families — namely Northwest Caucasian, Northeastern Caucasian/Nakh-Dagestanian and South Caucasian/Kartvelian — are also highly ergative, although this area may or may not be considered to be part of Europe.

Is everyone happy with this?

Jarawara is an interesting case. It certainly looks a bit like a separate alignment, but I think I’m inclined to classify it as another variety of split-ergativity. (It reminds me a bit of the system of Lhasa Tibetan, actually, although I’d need to review that system to be sure. You can find a description of Lhasa Tibetan in that Blue Bird book linked earlier.)

But it's pretty much the same as a direct-inverse system, surely? The drivers for construction choice are a bit different (person vs pragmatics and control) and the morphology is a bit messier than a prototypical direct-inverse system, but I'm not sure why one is split ergative but not the other.

The whole point of a split ergative system is that it’s ergative: there must be some construction which treats S and O the same, but A differently. Direct-inverse systems contain no such construction: they have a marker to indicate which argument is A and which is O, but they don’t have any sort of marking which equates S and O. By contrast, Jarawara verbal suffixes agree with S in intransitive clauses and agree with O in O-construction clauses, so S and O can be treated the same with regards to agreement.

EDIT: I’ve just realised there was a whole extra part of your post which I forgot about when I submitted this earlier — I’ll respond to it now.

I thought direct/inverse was a different, mostly unrelated phenomenon to voice. (Especially since voicing is a way of converting a transitive to an intransitive clause, but the inverse construction remains transitive.)

Voice doesn't just change transitivity, or at least many things that people call voice don't. For example:

1. There are a number of transitive "middle" verbs in languages like Ancient Greek, Fula, ...
2. Applicatives are normally considered to be voice, and when applied to transitive roots they may not affect valency at all (in some languages they just replace one object with a different object)

Depending on who's writing, voice seems to cover some mixture of:

1. Valency adjustment, both increasing (causative) and decreasing (passive, anti-passive)
2. Argument rearrangement (regardless of valency)
3. Verbal marking of argument roles and affectedness (middle voice, version ..., which may not change valency at all but do mark subject affectedness, e.g. "he built himself a house" vs "he built a house", or similarly with cutting hair, washing parts of one's body, ...)
4. Verbal marking of pragmatics and promotion to privileged roles for coreference and control purposes (again, may not involve detransitivisation)

I’d say that voice encompasses (3) and (4), but not (1) and (2), and I don’t consider applicatives to be voice. But, as you’ve noticed, others disagree. Personally, I think that ‘voice’ is one of those older terms from linguistics which really has no place outside Indo-European and should be thrown out (as with ‘infinitive’, ‘participle’ or ‘adverb’).

[Pabappa]

well, i dont know if it's really necessary to be that specific. we all know where the Caucasus is ... whether we call it eastern Europe, northeastern, southeastern, or not part of Europe at all .... i only posted because i was originally going to respond to the "girl's murder" thing too but realized you had already made my point better than I could have and I had just misread at first.

[bradrn]

Split ergativity — Part 1

(This section ended up taking more time than I expected, so I’ll publish it in two or more parts. This part focuses on the foundations of split ergativity, as well as active-stative systems; the next part(s) will focus on the other different types of split ergativity.)

Introduction to split ergativity

In previous chapters, I briefly mentioned split ergativity. This term refers to a situation in which an ergative alignment is only used in some situations — there is a ‘split’ between ergative-absolutive alignment and some other alignment. The other alignment will usually be nominative-accusative, although more complex splits can include tripartite and direct alignments. (If you don’t remember what these terms mean, refer to the earlier post What is ergativity?.) For instance, although Dyirbal is an ergative language (refer to earlier posts for examples), it uses nominative-accusative alignment in its pronouns:

ŋana-∅

1p-NOM

banaga-nʸu

return-NONFUT

We returned

ŋana-∅

1p-NOM

nʸurra-na

2p-ACC

bura-n

see-NONFUT

We saw you all

nʸurra-∅

2p-NOM

banaga-nʸu

return-NONFUT

You all returned

What is odd about split-ergativity is not that it exists (there are many other weird and wonderful alignment systems out there), but just how common it is: almost every known ‘ergative’ language has some sort of split in alignment. This is a notable asymmetry between nominative-accusative and ergative-absolutive systems — there are very many languages with an entirely nominative-accusative system (although these often have the occasional ergative construction as an exception), but almost all languages with an ergative-absolutive alignment have some sort of obvious split. I have no idea why this is, although I would very much like to know.

There are several different varieties of split found across split-ergative languages:

• Splits by the nature of the verb (split-S and fluid-S systems, collectively known as active-stative)
• Splits by the nature of the argument (animacy hierarchy-based systems)
• Splits by tense, aspect or mood
• Splits by clause type (main or subordinate)
• Combinations of the above

Each of these will be covered in the following sections.

The basis of split ergativity

Before understanding the details of how each variety of split ergativity works, it is often very useful to understand the underlying motivations behind why split ergative systems exist. There exist several different phenomena which can underlie split ergativity, and these will be dealt with in this section.

Recall that in the first chapter (What is ergativity?), I said the following about A and O:

The A relation represents the argument which prototypically controls the event, while the O relation represents the argument which prototypically does not control the event.

(To clarify, here I use the phrase ‘controls the event’ to mean ‘most relevant to the success of the activity’, or ‘could initiate or control the activity’ (Dixon); other synonyms are ‘volition’ and ‘agentivity’ (the opposite being ‘patientivity’).)

Now, recall that the unmarked case in a nominative-accusative system is the nominative (ignoring marked nominative systems), which covers A, whereas the unmarked case in an ergative-absolutive system is the absolutive, which covers P. But, as the quote above states, A is the argument which is prototypically in control, whereas P is the argument which prototypically is not in control. Combining these two observations leads to the following fact, which is highly relevant for a split ergative system:

In a nominative-accusative system, the agent or controller is unmarked, whereas in an ergative-absolutive system, the patient or undergoer is unmarked.

This fact is the basis of many split ergative systems: accusative alignment will be used for arguments which are more likely to be agentive, whereas ergative alignment will be used for arguments which are more likely to be patientive.

Another factor which can motivate a split is the status of S. Based on the above statement about the volition of A and O, we might attempt to determine whether S should most naturally be grouped with A or with O on semantic grounds. But it becomes clear very quickly that this fails: with some verbs, such as ‘speak’, the S argument is in control of the activity, whereas with other verbs, such as ‘break’, the S argument is not in control of the activity. Even worse, there are verbs such as ‘laugh’, in which S is sometimes in control and other times not (you can make a forced, voluntary laugh or you can laugh involuntarily). And there are still other verbs, such as ‘move’, with which some choices for S are interpreted as being in control (e.g. ‘I move’) while other arguments are interpreted as being out of control (e.g. ‘it moves’). Clearly S cannot be identified conclusively as being either the controlling agent or the controlled patient. But as we saw earlier, nominative-accusative alignment groups S with agentive A, whereas ergative-absolutive alignment groups S with patientive P. Thus, in some split systems, accusative alignment will be used when S is more agentive, but ergative alignment will be used when S is more patientive. Note that, as with the previous conclusion, nominative-accusative alignment ends up associated with agentivity, whereas ergative-absolutive alignment ends up associated with patientivity.

A third factor which can motivate a split is the structure of discourse. In some discourse structures, the S argument and A argument naturally go together; in others, the S argument and O argument naturally go together. This can lead to a split between nominative-accusative alignment for the former types of discourse, and ergative-absolutive alignment for the latter types. Dixon argues that this is the motivation behind splits on tense/aspect/mood, although I personally find his reasoning doubtful.

Splits by the nature of the verb

Probably the simplest variety of split to understand is a split by the nature of the verb. As discussed above, an intransitive verb may have a more agentive S or a more patientive S. Thus it is impossible to group S consistently with either A or O. So-called active-stative systems solve this problem in an obvious way: grouping S with A when S is more agentive, but grouping S with O when S is more patientive. There are two main systems used to accomplish this: Split-S systems and Fluid-S systems.

Split-S systems

Split-S systems divide intransitive verbs into two groups: some intransitive verbs mark S the same as A, and others mark S the same as O. As outlined above, the former marking is generally used for verbs where S is typically in control of the action, whereas the latter marking is used for verbs in which S is typically not in control of the action or which represent a state, although of course there are differences between languages in exactly which verbs go into which group. I will call the former group ‘S=A verbs’, and the latter group ‘S=O verbs’, and their arguments ‘S=A NPs’ and ‘S=O NPs’ (giving four different types of verbal argument, namely A, O, S=A, S=O), although terminology differs: for instance, Dixon calls the verbs and arguments S_a and S_o instead. (I would use this notation as well, but BBCode makes it particularly inconvenient to type, so I will stick with S=A and S=O.) Other terms include ‘active’ for S=A and ‘stative’ for S=O.

It may be useful to elaborate on exactly how such a system works. As we will see below, the majority of languages using a split-S system use verbal agreement affixes rather than case markers. A typical split-S system uses two sets of agreement affixes: one set of affixes is used for both S and A, while the other is used for both S and O. Using the names of the corresponding case markers, we might term the first set ‘nominative’, and the second set ‘absolutive’ (although curiously enough I actually don’t think I’ve seen those names used anywhere else). For a normal transitive verb, A is signalled using the nominative set and O is signalled using the absolutive set; for instance, ‘I hit you’ would be expressed as some variant of ‘hit-1s.NOM-2s.ABS’ (although of course the affixes could be in different places depending on the language). However, for an intransitive verb, the argument could be marked with either set, since they overlap for the S argument; S=A verbs are those which use the nominative set for agreement, whereas S=O verbs are those which use the absolutive set for agreement. For instance, if ‘enter’ is an S=A verb, then ‘I enter’ would be expressed as ‘enter-1s.NOM’; if ‘weep’ is an S=O verb, then ‘I weep’ would be expressed as ‘weep-1s.ABS’.

Now going back to the verbs themselves, note that the exact semantic basis for S=A vs S=O verbs can differ from language to language, although it is always some variant of the volitionality distinction described above. For instance, Mithun (1991) (cited in Dixon) describes the semantic basis of S=A vs S=O in Lakota, Caddo and Mohawk by saying that prototypical S=A arguments ‘perform, effect, instigate and control events’, whereas prototypical S=O arguments are ‘affected; things happen or have happened to them’. (Mithun also investigated the changes in this distinction over time, although Dixon doesn’t go into any detail on these changes.) McGregor (1997) has proposed that verbs appear in a hierarchy, starting from verbs of cognition and speech, and extending through violence, causation, states, and ending at existentials; he suggests that verbs to the left of the hierarchy usually show accusative marking, whereas verbs to the right usually show ergative marking. Cultural differences can also play a role in assigning intransitive verbs to groups, with Dixon giving the example of the word ‘vomit’: some cultures see this as an involuntary activity, while in others it may be induced volitionally as a cultural activity. And, as with all linguistic distinctions, there are always a few exceptions to the general rule, often through borrowing or phonological change: for instance, Guaraní borrows all Spanish intransitive verbs as S=A, leading to unexpected S=A verbs like avuři ‘to be bored’ (although the native Guaraní verb kaigwá ‘to be/become bored’ is S=O, as expected). The sizes of the two classes can also vary greatly: according to Merlan (1985) (cited in Dixon), Arikara has a small closed S=O class and a large open S=A class, whereas Dakota has a small closed S=A class and a large open S=O class, while Guaraní treats both classes as open.

To give some concrete examples, Dixon lists some representative verbs for three languages:

• For the Siouan language Mandan, S=A verbs include ‘break camp’, ‘enter’, ‘arrive’, ‘think it over’, ‘go’, while S=O verbs include ‘fall’, ‘be lost’, ‘lose balance’; S=O also includes adjectival verbs such as ‘be alive’, ‘be brave’, ‘be strong’.
• For the Tupí-Guaraní language Guaraní, S=A verbs include ‘go’, ‘remain’, ‘continue’, ‘follow’, ‘fall’ (which was S=O in Mandan), while S=O verbs include those with adjectival meanings, but also ‘remember’, ‘forget’, ‘tell a lie’, ‘weep’. Interestingly, although the personal affixes for S=A verbs are exactly the same as the A affixes for transitive verbs, the personal affixes for S=O verbs are only similar — but not identical — to the O affixes for transitive verbs.
• For the Siouan language Hidatsa, the two classes are not quite as distinct as in the two previous examples: S=A includes ‘talk’, ‘follow’, ‘run’, ‘bathe’, ‘sing’, but also ‘have hiccups’, ‘forget’ and ‘die’, whereas S=O includes ‘yawn’, ‘err’, ‘cry’, ‘fall down’ but also ‘stand up’, ‘roll over’, ‘dress up’.

In some languages, the distinction between S=A and S=O verbs goes further than just marking. For instance, in the Athabaskan language Slave, ‘causatives can be based on S_o, but not on S_a; passive on S_a but not on S_o; noun incorporation can involve O or S_o, but not S_a; and so on’ (Rice 1991, cited in Dixon). (We might justify this by saying e.g. passives are mainly used to background an agent, so must be used on verbs which have an agent, and noun incorporation is used to background a patient, so can only be used on verbs which have a patient, and so on and so forth, although this feels suspiciously like a post hoc rationalisation.) As another example, Guaraní ‘imperative inflection’ can occur with transitive verbs and S=A verbs, but not S=O verbs. Conversely, some languages have constructions in which S=O and S=A arguments are combined as a unitary S class: for instance, in subordinate clauses, many Tupí-Guaraní languages use O agreement affixes for all S arguments. As another example, the Caddoan language Wichita normally has a distinction between S=A and S=O intransitive arguments, but it has several constructions which group a unitary S argument with either A or O: for instance, A and S behave the same with respect to word order, and O and S behave the same with respect to noun incorporation.

There is a very strong tendency for split-S languages to mark their alignment using verbal agreement rather than noun case: all of the languages described above use verbal agreement. This isn’t too surprising: since the split is dependent on the nature of the verb rather than the nature of the noun, it makes sense for the split to be marked on the verb rather than the noun. Typically, there are two sets of affixes — one for A, and one for O — and intransitive verbs use affixes either from the A set or the O set depending on the verb being used. However, there are a few more unusual systems in use:

• The Papuan language Yawa has a set of verbal agreement prefixes, but only for O and S=O; A arguments are marked with a postposition, which is placed after the argument and inflects for person, number and (in third person) gender. For S=A verbs, this postposition has developed into a verbal prefix, which is used in addition to the postposition. Thus transitive sentences contain A (marked with a postposition), O and a verb (agreeing with O), intransitive sentences with S=O contain S=O and a verb (agreeing with S=O using the O agreement markers), and intransitive sentences with S=A contain S=A (marked with a postposition) and a verb (agreeing with S=A using an agreement marker derived from the postposition). (Note that I don’t know the exact word order; I’m simply listing the constituents in each type of sentence, but not necessarily in order.)
• The Austronesian language Tolai (mentioned in an earlier post) has an unusual system, working solely through word order. The word order in transitive clauses is AVO; in intransitive sentences, the word order is SV for S=A verbs, and VS for S=O verbs. This makes Tolai a rare example of an ergative language where the ergativity is signalled only via word order.
• The Arawak language Waurá has a similar system to Tolai, but the verb also contains an agreement marker: for transitive verbs and S=A verbs, the verb agrees with A or S=A, while this marker is absent with S=O verbs.

I would also like to briefly mention the system of Tongan. According to McGregor, Tongan has an unusual alignment system in which different types of verbs use different case-marking techniques:

A different type of split in case-marking is found in languages such as Tongan (Austronesian, Tonga), where bivalent verbs show different patterns in case-marking. For instance, verbs of violence have an ergative marked Agent and an unmarked Undergoer; verbs of feeling have an unmarked experiencer and a second argument in the locative or dative; the ‘have’ verb takes two unmarked NPs; various other patterns exist (Tsunoda 1981: 405–406). Thus in Tongan, ergative patterning in case-marking is restricted to higher transitivity event types, events that are more effective (Tsunoda
1981).

This sort of pattern could be thought of as an unusual variety of split-S marking, albeit one which varies on bivalent words rather than univalent ones.

Fluid-S systems

Split-S systems are certainly an improvement on having one unitary S class which is not distinguished for agentivity. But they still have a problem: the verb itself may not be sufficient to predict the agentivity of the S argument. Above, the example of ‘vomit’ was given as a word whose agentivity is dependent on culture, but there are less contrived examples too: for instance, a sentence like ‘the box moved’ could either be agentive (maybe it’s slipping down a slope) or patientive (maybe I’m pushing it).

So-called fluid-S systems avoid these problems. These fascinating systems are a bit like a more extreme version of split-S systems: in split-S systems, the choice of verb consistently determines whether S is marked like A or O, but in fluid-S systems, S is marked like A or O depending on the agentivity of S in that particular sentence. Fluid-S systems could be considered to be a combination of ‘syntactic-based marking’ (for A and O) and ‘semantically-based marking’ (for S), as the terms were defined in the introductory chapter.

Perhaps an example might help illustrate how this system works. The most commonly-referenced fluid-S language is probably Tsova-Tush (also known as Bats or Batsbi), from the Northeastern Caucasian family. In Tsova-Tush, the S argument may be marked with either the nominative case (like A) or the absolutive case (like O), depending on the agentivity of S:

∅-txo

ABS-we

naizdrax

to.ground

kxitra

fell

We fell to the ground (and this was unintentional on our part)

a-txo

NOM-we

naizdrax

to.ground

kxitra

fell

We fell to the ground (and we caused this somehow)

(From McGregor; glosses and translations have been altered slightly, but should still be correct.)

Note that this alternation only applies to intransitive verbs; fluid-S languages are the same as split-S languages (and indeed the vast majority of languages with verbal agreement) with regards to transitive verbs, in that transitive verbs always have A and O marked distinctly and in a consistent manner. (Although this by no means precludes such a contrast being expressed in some other way: for instance, the fluid-S language Acehnese has a special ‘uncontrolled’ verbal prefix teu- used with transitive verbs to denote an action which is normally controlled but in this instance happens by chance.)

As with split-S languages, fluid-S languages can certainly contain intransitive verbs which are only ever used with one type of marking. However, unlike in split-S languages, this is not due to any inherent property of those words, but rather due to semantic, pragmatic and cultural considerations, as this quote (from Dixon) about Tsova-Tush illustrates:

The results Holisky obtained were determined partly by speakers’ world-view and by other pragmatic factors. She mentions ‘when I constructed the first person form for the verb “get poor” in Tsova-Tush using S_a [i.e. S=A, agentive] marking, my consultant did not say categorically that it wasn't possible. She said it isn't possible because you would never want to be poor’ (Holisky 1987: 115).

As the Tsova-Tush examples above illustrate, fluid-S languages are not nearly as restricted as split-S languages are in terms of marking. Tsova-Tush uses case markers, as does the Eastern Pomo language (from northern California), but the Tabassaran and Crow languages use verbal agreement instead. By contrast, most split-S languages use verbal agreement to identify S with A or O, and few (if any) use case-marking.

One intriguing consequence of fluid-S marking is that a lesser number of verbs may be needed, due to the existence of pairs of verbs which only differ in volition. For instance, English has two words ‘slip’ and ‘slide’, which have the same meaning, but differ in volitionality (I am in control for ‘slide’, but not for ‘slip’). However, languages like Tsova-Tush and Eastern Pomo need only one verb encompassing those two meanings: a sentence like I-NOM slip/slide would mean ‘I slide’, whereas I-ABS slip/slide would mean ‘I slip’.

Finally, I would like to note that, although split-S and fluid-S constitute two separate systems, it is also possible to combine them. For instance, it is not uncommon for a split-S system to contain a set of verbs which are fluid between S=A and S=O. For instance, in Guaraní, che-karu (with S=O) means ‘I am a big eater’, whereas a-karu (with S=A) means ‘I am eating’ (although Dixon notes that this contrast ‘is not productive’, i.e. it can only be applied to a limited set of words).

[akam chinjir]

The other alignment will usually be nominative-accusative

Do you have numbers on this? I'd have thought constructions in which neither argument is marked would be relatively common, at least with TAM splits. (More on that in a bit.)

Incidentally, if you're right about this, why speak of split ergativity rather than split accusativity?

almost every known ‘ergative’ language has some sort of split in alignment

And do you have numbers on this? I feel like sometimes people confuse this issue with the separate fact that no syntactically ergative language is syntactically ergative in every imaginable way.

there are very many languages with an entirely nominative-accusative system (although these often have the occasional ergative construction as an exception), but almost all languages with an ergative-absolutive alignment have some sort of obvious split. I have no idea why this is, although I would very much like to know

Given how you're setting things up, I expect you should be counting differential object marking in accusative languages as a kind of split accusativity, and that's reasonably common.

To fill out both those points, I'm going to anticipate your next post a bit, and say something about TAM splits.

Take Basque. In perfective clauses, the subject of a transitive verb takes ergative case. In a progressive clause, you get something like this:

emakuma-a          ogi-a              ja-te-n      ari   d-a      
woman-ART.SG(ABS)  bread-ART.SG(ABS)  eat-NML-LOC  PROG  3.ABS-AUX
"The woman is eating the bread"

(I got the example from Coon and Preminger, "Split ergativity is not about ergativity," which is in the Oxford Handbook of Ergativity, something you should probably browse; the Baker and Bobaljik paper I recommended earlier is also in there. Drafts of both articles are easily googlable. Coon's book Aspects of Split Ergativity goes into a lot more detail, unsurprisingly.)

The thing to notice is that what you'd think of as the main verb is occurring here in a nominalised form, with locative case-marking, and is the complement of an auxiliary verb. That's to say, it looks something like "The woman is at eating bread." Now, that doesn't look like a transitive construction, and the fact that the verb doesn't get ergative case shouldn't be a surprise, and doesn't justify any talk of a split. And as I understanding (mostly from reading Coon), this is very often how things look in languages that supposedly have TAM splits.

Two final points.

First of all, it's important to this analysis that the supposed split in Basque is between an erg/abs system and an abs/abs (or unmarked/unmarked) system; the analysis wouldn't work if "bread" were marked accusative. Which is a big part of why I asked about the basis of your claim that with ergativity splits, the other alignment is usually nom/acc.

Second, in a language with nom/acc case-marking, you could have exactly the same thing going on in imperfective or progressive aspects, but it wouldn't make a difference to how the subject got marked. But that's a pretty weak reason to think that ergative languages are more prone to splits than accusative ones.

(Though maybe you'd expect there to be accusative languages in which objects in imperfective clauses don't get accusative case. Actually, there's another pattern that Coon and Preminger talk about---in Samoan, for example---where the object in imperfective clauses takes an oblique case, making the clause intransitive, and the subject doesn't end up ergative. No obvious reason why you couldn't get the same thing in a language with mostly accusative case-marking.)

[bradrn]

The other alignment will usually be nominative-accusative

Do you have numbers on this? I'd have thought constructions in which neither argument is marked would be relatively common, at least with TAM splits. (More on that in a bit.)

No numbers, sorry. But a split of ergative vs. no marking definitely isn’t terribly rare, although it’s certainly much less common than ergative vs. accusative. Aside from TAM splits, it also tends to happen in more complex animacy hierarchy-based distinctions which have a split between three or more alignments. (I’ll cover this in detail in Part 2).

Incidentally, if you're right about this, why speak of split ergativity rather than split accusativity?

Because split ergative systems don’t always occur with accusative alignment as one of the splits. There exist examples of split systems containing ergative alignment but not accusative alignment (e.g. Umpithamu, which is split between ergativity and optional ergativity but not accusativity), whereas there are no examples (at least as far as I’m aware) of split systems containing accusative alignment but not ergative alignment. Also, syntactic ergativity never occurs with morphological accusativity, but it does often occur with split systems (e.g. Dyirbal), so on this grounds it is reasonable to group split ergativity as being more related to ergativity rather than accusativity; this allows you to make the elegant generalisation that ‘syntactic ergativity only occurs with some sort of morphological ergativity (either ‘normal’ or split)’.

(This is actually one of my main complaints about that Blue Bird article: there are so many interrelationships and similarities between the various types of ergativity that I find it hard to doubt that there’s some sort of unified category behind all this.)

almost every known ‘ergative’ language has some sort of split in alignment

And do you have numbers on this? I feel like sometimes people confuse this issue with the separate fact that no syntactically ergative language is syntactically ergative in every imaginable way.

Again, unfortunately I have no numbers. But the notion that ‘not all parts of a language need to be ergative’ is important, and I will cover it later in the section on syntactic ergativity.

there are very many languages with an entirely nominative-accusative system (although these often have the occasional ergative construction as an exception), but almost all languages with an ergative-absolutive alignment have some sort of obvious split. I have no idea why this is, although I would very much like to know

Given how you're setting things up, I expect you should be counting differential object marking in accusative languages as a kind of split accusativity, and that's reasonably common.

What do you mean by ‘differential object marking’? My best guess is that you’re talking about direct/inverse alignment — and as I mentioned earlier in the thread, I count that as being a completely different sort of alignment to either split ergativity or ‘split accusativity’:

The whole point of a split ergative system is that it’s ergative: there must be some construction which treats S and O the same, but A differently. Direct-inverse systems contain no such construction: they have a marker to indicate which argument is A and which is O, but they don’t have any sort of marking which equates S and O. [And there’s no such grouping of S and A either.]

To fill out both those points, I'm going to anticipate your next post a bit, and say something about TAM splits.

Take Basque. In perfective clauses, the subject of a transitive verb takes ergative case. In a progressive clause, you get something like this:

emakuma-a ogi-a ja-te-n ari d-a
woman-ART.SG(ABS) bread-ART.SG(ABS) eat-NML-LOC PROG 3.ABS-AUX
"The woman is eating the bread"

(I got the example from Coon and Preminger, "Split ergativity is not about ergativity," which is in the Oxford Handbook of Ergativity, something you should probably browse; the Baker and Bobaljik paper I recommended earlier is also in there. Drafts of both articles are easily googlable. Coon's book Aspects of Split Ergativity goes into a lot more detail, unsurprisingly.)

The thing to notice is that what you'd think of as the main verb is occurring here in a nominalised form, with locative case-marking, and is the complement of an auxiliary verb. That's to say, it looks something like "The woman is at eating bread." Now, that doesn't look like a transitive construction, and the fact that the verb doesn't get ergative case shouldn't be a surprise, and doesn't justify any talk of a split. And as I understanding (mostly from reading Coon), this is very often how things look in languages that supposedly have TAM splits.

Thanks for the resource recommendations — I’ll have to track those down and read them! And thanks for telling me about this phenomenon in TAM splits; now that I know about it, I’ll definitely include this.

I think the problem here is a common one: people are far too eager to call something ‘ergative’ or ‘split ergative’ when it isn’t really. (Going back to Blue Bird, this is a place where I thoroughly agree with that article.) In this case, the problem is that Basque uses a progressive construction which happens to contain two intransitive clauses ([The woman is] at [eating bread], treating is as syntactically intransitive here). Of course it’s no surprise in this situation that Basque should use two absolutive arguments, but if you don’t look at this carefully, it’s easy to misinterpret this as an aspectually conditioned split.

First of all, it's important to this analysis that the supposed split in Basque is between an erg/abs system and an abs/abs (or unmarked/unmarked) system; the analysis wouldn't work if "bread" were marked accusative. Which is a big part of why I asked about the basis of your claim that with ergativity splits, the other alignment is usually nom/acc.

This is true: if Basque had nom/acc alignment in the progressive, that would be evidence of a split. (Although possibly this would be a split between main vs. subordinate clauses rather than perfective vs. progressive aspect.)

Second, in a language with nom/acc case-marking, you could have exactly the same thing going on in imperfective or progressive aspects, but it wouldn't make a difference to how the subject got marked. But that's a pretty weak reason to think that ergative languages are more prone to splits than accusative ones.

I don’t follow: surely a nom/acc language which uses a Basque-like construction for the progressive would have a split between ‘woman-NOM bread-ACC eat-PERF’ for perfective and ‘woman-NOM bread-NOM eat-NML-LOC PROG is’ for progressive? To me, this looks like exactly as big of a difference as in the ergative version.

(Though maybe you'd expect there to be accusative languages in which objects in imperfective clauses don't get accusative case. Actually, there's another pattern that Coon and Preminger talk about---in Samoan, for example---where the object in imperfective clauses takes an oblique case, making the clause intransitive, and the subject doesn't end up ergative. No obvious reason why you couldn't get the same thing in a language with mostly accusative case-marking.)

According to McGregor, this is pretty much what happens in Kurmanji Kurdish, albeit with a split on tense rather than aspect:

min

1s.ERG

hon

2p.ABS

dit-in

saw-2p

I saw you all

ez

1s.ABS

we

2p.OBL

di-bin-im

PRS-see-1s

I see you all

[akam chinjir]

Differential object marking occurs when different objects are case-marked differently depending on their properties. These could be more or less invariant lexical properties, such as animacy, or more variable pragmatic properties such as definiteness or specificity. (Spanish has DOM of the first sort, Turkish DOM of the second sort, for example.)

An interesting wrinkle: in languages with an ergative case you can get differential subject marking whereby the subject is ergative when the object is sufficiently definite, animate, or whatever; but it's absolutive otherwise. (Actually I don't know if this ever depends on the animacy of the object, in the cases I remember it involved specificity or definiteness.) The Samoan and Kurmanji Kurdish examples are of this sort, I think.

Second, in a language with nom/acc case-marking, you could have exactly the same thing going on in imperfective or progressive aspects, but it wouldn't make a difference to how the subject got marked. But that's a pretty weak reason to think that ergative languages are more prone to splits than accusative ones.

I don’t follow: surely a nom/acc language which uses a Basque-like construction for the progressive would have a split between ‘woman-NOM bread-ACC eat-PERF’ for perfective and ‘woman-NOM bread-NOM eat-NML-LOC PROG is’ for progressive? To me, this looks like exactly as big of a difference as in the ergative version.

I was imagining a case where "bread" is accusative in both cases, though my next paragraph raised the possibility of an accusativity split of the sort you mention.

(It's actually a tricky question, and the answer seems to be different from language to language, whether and when a subclause 'counts' for the purposes of settling agreement, case marking, and such in the matrix clause.)

[bradrn]

Differential object marking occurs when different objects are case-marked differently depending on their properties. These could be more or less invariant lexical properties, such as animacy, or more variable pragmatic properties such as definiteness or specificity. (Spanish has DOM of the first sort, Turkish DOM of the second sort, for example.)

An interesting wrinkle: in languages with an ergative case you can get differential subject marking whereby the subject is ergative when the object is sufficiently definite, animate, or whatever; but it's absolutive otherwise. (Actually I don't know if this ever depends on the animacy of the object, in the cases I remember it involved specificity or definiteness.) The Samoan and Kurmanji Kurdish examples are of this sort, I think.

Thanks for explaining! I ended up looking at the Wikipedia article on DOM as well, and that also helped a lot.

Second, in a language with nom/acc case-marking, you could have exactly the same thing going on in imperfective or progressive aspects, but it wouldn't make a difference to how the subject got marked. But that's a pretty weak reason to think that ergative languages are more prone to splits than accusative ones.

I don’t follow: surely a nom/acc language which uses a Basque-like construction for the progressive would have a split between ‘woman-NOM bread-ACC eat-PERF’ for perfective and ‘woman-NOM bread-NOM eat-NML-LOC PROG is’ for progressive? To me, this looks like exactly as big of a difference as in the ergative version.

I was imagining a case where "bread" is accusative in both cases, though my next paragraph raised the possibility of an accusativity split of the sort you mention.

I’m still not quite sure about how this would work. Why would ‘bread’ be accusative in both cases?

(It's actually a tricky question, and the answer seems to be different from language to language, whether and when a subclause 'counts' for the purposes of settling agreement, case marking, and such in the matrix clause.)

This sounds interesting. Do you have any idea about where I could find some resources about this topic?

[Xwtek]

Interestingly, an O-construction ("inverse") is only possible if the O is 3rd person or in certain morphologically limited conditions, which is not what would be expected if the purpose is to promote SAPs.

I once had a conlang that works this way. I thought I misunderstand the direct-inverse system. In that conlang, the verb agrees with both the subject and the direct object. The verb marks which one is the pivot (subject or object). The verb then agree the nonpivot argument. The pivot can only be 3rd person. If both are SAP, then a special agreement affix is used that agrees with both.

I decided to use a more standard direct-inverse system, where SAP outranks the pivot. But still, in my conlang, there are two relative constructions. One is the non-reduction strategies, where you can relativize anything at a cost of occasional ambiguity. The second is a gap strategy, which is much less ambiguous, but you can only relativize the pivot argument. (i.e. arguments marked as oblique, cannot be relativized, except with applicatives, but there are no locative applicatives nor genitive applicatives, only benefactive, and instrumental ones.) Is the last one realistic?

[bradrn]

Interestingly, an O-construction ("inverse") is only possible if the O is 3rd person or in certain morphologically limited conditions, which is not what would be expected if the purpose is to promote SAPs.

I once had a conlang that works this way. I thought I misunderstand the direct-inverse system. In that conlang, the verb agrees with both the subject and the direct object. The verb marks which one is the pivot (subject or object). The verb then agree the nonpivot argument. The pivot can only be 3rd person. If both are SAP, then a special agreement affix is used that agrees with both.

I decided to use a more standard direct-inverse system, where SAP outranks the pivot. But still, in my conlang, there are two relative constructions. One is the non-reduction strategies, where you can relativize anything at a cost of occasional ambiguity. The second is a gap strategy, which is much less ambiguous, but you can only relativize the pivot argument. (i.e. arguments marked as oblique, cannot be relativized, except with applicatives, but there are no locative applicatives nor genitive applicatives, only benefactive, and instrumental ones.) Is the last one realistic?

If I understand correctly, your system selects a pivot for transitive sentences, and then marks whether the other argument ranks above or below the pivot on the animacy hierarchy — is that correct? I’ve never heard of such a system, and it does sound a bit odd, but it certainly isn’t unrealistic. The two relativisation strategies both look realistic as well.

[akam chinjir]

I’m still not quite sure about how this would work. Why would ‘bread’ be accusative in both cases?

I was ignoring all sorts of potential complications and just thinking that it's still the object of a verb, so would still get accusative case. (I think probably the main alternative would be for it to get genitive or oblique case, given that this is a nominalisation.)

(It's actually a tricky question, and the answer seems to be different from language to language, whether and when a subclause counts for the purposes of settling agreement, case marking, and such in the matrix clause.)

This sounds interesting. Do you have any idea about where I could find some resources about this topic?

I actually don't, and I'm afraid in retrospect I think I was speaking unreasonably vaguely. I now think it would've been better to say it can be a tricky question when superficially similar constructions do and do not involve a subclause and how that might affect case-marking and agreement and such. Like, in "I am eating bread," is "eating bread" a subclause? If it is a subclause, does it count as an object of the matrix verb? Can the embedded object count as an object of the matrix verb?

(Aside. I think I can get two interpretations of "I am not eating bread": Not (I am eating bread), and I am (not (eating bread)). And I think "I am not not eating bread" is interpretable, as Not (I am (not (eating bread))). So the common negation test seems to imply that "eating bread" is actually a separate clause here, fwiw.)

I'm afraid I don't know of anything I can suggest that's at all general about this, the things I can think of offhand are both pretty narrow and pretty technical.

One pattern that I think is fairly common that might be interesting to look into goes by the name of "restructuring."

Suppose we have "I want to eat bread." Again, is this one clause or two? Is "bread" an object of "want"? Is "I" the subject of a transitive verb?

In some languages, there'll be verbs, likely including a want verb, that you might expect to take a clausal complement, but the resulting structure behaves as if it's monoclausal. Like, a translation of "I want to eat bread" might have an ergative subject; or the want verb might be inflected to agree with the embedded object. (And for some reason this is often called restructuring.)

You can even get alternations where a verb will sometimes take a truly clausal complement, and behave as intransitive, but other times will sometimes restructure, and behave as transitive when the embedded verb is transitive. So you can get something like "I want to eat bread" where sometimes the subject is ergative and sometimes it's not. (Presumably there'd be other differences too.)

(I happen to know that a case like that gets brought up in Baker and Camargo Souza, Agree without Agreement: Switch reference and reflexive voice in two Panoan languages, but that's a technical article about an unrelated topic. Admittedly it's an interesting topic: a switch reference system where there's a marker used when the object of the embedded verb corefers with the matrix subject.)

(Am I right to think that French allows both je le veux manger and je veux le manger? If so, that looks like the right sort of alternation, though not one involving ergative case, of course.)

Er, I think I mostly haven't managed to answer your question, but I need to get on with the day. Hope what I came up with is useful!

[bradrn]

I’m still not quite sure about how this would work. Why would ‘bread’ be accusative in both cases?

I was ignoring all sorts of potential complications and just thinking that it's still the object of a verb, so would still get accusative case. (I think probably the main alternative would be for it to get genitive or oblique case, given that this is a nominalisation.)

Sorry, I think I misinterpreted what you were saying — I think I understand now!

(It's actually a tricky question, and the answer seems to be different from language to language, whether and when a subclause counts for the purposes of settling agreement, case marking, and such in the matrix clause.)

This sounds interesting. Do you have any idea about where I could find some resources about this topic?

I actually don't, and I'm afraid in retrospect I think I was speaking unreasonably vaguely. I now think it would've been better to say it can be a tricky question when superficially similar constructions do and do not involve a subclause and how that might affect case-marking and agreement and such. Like, in "I am eating bread," is "eating bread" a subclause? If it is a subclause, does it count as an object of the matrix verb? Can the embedded object count as an object of the matrix verb?

…

I agree with you on this — it can be very difficult to determine what is and isn’t a subclause, especially in English. (I think the Basque case is a bit easier because of the case-marking.)

Er, I think I mostly haven't managed to answer your question, but I need to get on with the day. Hope what I came up with is useful!

Not at all, you’ve actually been very informative! What you said is definitely very useful — I’ll try to remember this information, since I’m sure I’ll end up using it at some point.