EPISODE TRANSCRIPT
Optimality Theory (René Kager)
Automatically transcribed from the episode audio. Some words and names may be inaccurate.
Listen to the episodeDownload textDownload captions
Welcome to the Deep Dive. Today, we're tackling something really foundational in linguistics, a theory that, well, it really shook things up. We're doing a deep dive into optimality theory, OT for short. That's right. For a long time, you know, linguistics, especially the generative folks, thought language worked like a computer program, a series of ordered steps, these rewrite rules, taking you from some deep form to what you actually say. Like a recipe, step by step. bake the cake this way, exactly this way. Exactly.
And OT just comes along and says, well, what if there's no recipe, no ordered steps? It keeps the big goal finding universal language principles, but totally flips the script on how we get there. Okay, so no more serial derivations. What replaces them? Conflict. At its heart, OT. and we're drawing a lot here from Kager's work, which really lays it out, sees language as a system of competing forces. These are called constraints. Constraints. And these constraints, they define what makes a good word or sentence structurally. Here's the kicker. They're universal, found in all languages, but they can be broken. They're viable. Hold on. Universal, but viable. That sounds like a contradiction. If they apply everywhere, how do languages end up so different? Why don't we all sound the same if the constraints are the same? Ah, that's the genius of it. It's not the constraints themselves that differ, it's their hierarchical ranking. Think of it like a priority list. Every language has the same list of rules, but each language shuffles that list, deciding which rules are more important than others. So language A might say, constraint X is absolutely vital, never break it. While language B says, eh, constraint X is okay, but constraint Y, that's the one we really care about. Precisely. Constraint A dominates B in one language, B dominates A in another. And that simple re -ranking creates radically different patterns, all from the same universal toolkit. Okay, let's get into the big players in this conflict. You mentioned these competing forces. The sources talk about two main families of constraints, right? Yes, the two titans constantly at odds. Markedness, M, and faithfulness, F. Markedness and faithfulness.
Let's break those down. Markedness. Markedness constraints are all about the output. They push for simplicity, for structures that are easy to produce or perceive. Think simple syllables, common sounds, things that are, well, unmarked, less complex. Okay, so markedness is like the efficiency expert, always trying to streamline things. What about faithfulness? Faithfulness is the opposite pull. It's about the input, the underlying form in the lexicon. Faithfulness constraints demand that the output, what you actually say, stays true to the input. It preserves the information, the contrast stored in your mental dictionary. So markedness wants to simplify, maybe even change things from the input to make the output cleaner. But faithfulness says, whoa, hang on. You need to respect the original form.
Don't lose that information. They're fundamentally opposed. Absolutely. You can't usually satisfy both perfectly. If you obey a strong markedness constraint and simplify the output, you're probably violating some faithfulness constraint by changing the input form. And if you stick rigidly to the input to satisfy faithfulness, you might end up with a really complex marked output structure that violates markedness. Exactly. The specific ranking in a given language dictates who wins that fight in any given situation. So our mission today is to really unpack this. We'll go through the architecture, see how this constraint interaction plays out in sounds, syllables, even sentences, basically following Cager's structure chapter by chapter. Let's start right at the beginning, chapter one. the basic architecture, OT is described as an input -output system. Three main parts.
Lexicon, generator, evaluator. What's new here, especially with the lexicon? Well, the lexicon itself contains the morphemes, the basic building blocks, but OT makes this assumption called richness of the base, ROTB. And honestly, this was pretty radical compared to older models. Radical how? It assumes that the input forms, the lexical representations, are completely unconstrained. Anything goes. Your mental dictionary, in theory, could contain any sound combination imaginable, even ones that never show up in the actual spoken language. Wait, so my brain might have an entry for, I don't know, a word, starting with NZ, even though English never allows that? According to ROTB, yes. It gets rid of the need for those old morphine structure constraints, rules that basically filter the dictionary. OT says, don't filter the dictionary. Let the main constraint system do all the work of filtering what actually comes out. It avoids stating restrictions twice.
Okay, so the input is this wild, unrestricted thing. What happens next? It goes to the generator, or GEN. GEN's job is conceptually simple, but vast. For any single input, it turns out a potentially infinite set of output candidates. Infinite? How? Well, it considers everything. Deleting sounds, adding sounds, changing features, moving things around, leaving things exactly as they are. Every possible modification or lack thereof. Freedom of analysis, they call it. That sounds computationally explosive. If GEN throws infinite options onto the table, you need a really decisive judge. That must be the evaluator, Eval. That's Eval. It's the core of the grammar. Eval takes that specific language's ranked list of constraints and uses it to assess all those candidates from GEN. Its job is to find the one optimal candidate.
The winner. The one that actually gets pronounced. Right. The one that best satisfies the hierarchy, meaning it incurs the least serious violations. And the way it judges is absolutely crucial. Strict domination. Strict domination. OK, explain that. Because my first instinct is, why not just count violations? Candidate A breaks five low -ranked rules. Candidate B breaks one high -ranked rule. Maybe A is better overall. Ah, but OT says no. Strict domination means the hierarchy is absolute. If a candidate violates a constraint, even just once, any candidate that doesn't violate that constraint is immediately better, no matter how many lower -ranked constraints the first candidate satisfies perfectly. So one strike on a top -level constraint and you're out, regardless of anything else. You're out of the running against any competitor that satisfies that top constraint.
It's merciless. A higher -ranked constraint's decision is final for comparing any two candidates. That rigidity is what gives each language its specific profile. Adding up violations would just muddy the waters. That makes sense. That absolute priority is how you get different languages. Let's make this concrete with that classic example. Voice sounds at the end of syllables, like in Dutch versus English. Perfect example. It pits faithfulness against markedness. The markedness constraint here is often called something like text. Voiced coda. Basically, don't have voiced obstruents like B, D, G at the end of a syllable.
It prefers voiceless ones like P, T, K there. Because voiceless ones are maybe simpler, less effort. That's the idea behind markedness, yes. Simpler, more common structures are preferred. The competing constraint is faithfulness, specifically text, ident, IO, voiced A. It says, hey, if the input segment was voiced, the output segment should be voiced too. Preserve that feature. Okay, so in English, if I have the word bed in my lexicon. The output is bed. You pronounce the D. English tolerates that voice sound in the coda. Which means English prioritizes faithfulness here. Exactly. The ranking is text, iodine, IO.
Keeping the output true to the input D is more important than avoiding the market voice coda. So the contrast between bet and bet is maintained. Faithfulness wins. But Dutch does the opposite, right? Right. The same input bed comes out as bet in Dutch. The final D becomes a voiceless T. So for Dutch. The ranking is flipped. Avoiding that marked voice code is the top priority. The language is willing to sacrifice faithfulness, change the voicing feature to satisfy that higher ranked markedness constraint. And that leads to positional neutralization. The difference between underlying D and T is lost specifically in that code of position. Precisely. And here's a subtle but important point Kager makes. Even though text edit is ranked lower in Dutch and gets violated, it's still active. Imagine GEN.
also produced a candidate -like pet for input bed. That candidate pet also satisfies the top constraint, text voice to coda. But compared to the actual winner bet, pet is worse because it violates two faithfulness constraints. It changes voicing and it changes the initial B to P. The real output bet only violates one faithfulness constraint, voicing. Ah, so even dominated constraints enforce minimal violation. You break the rules only as much as absolutely necessary to satisfy the higher -ups. You got it. It's not just about winning or losing. It's about losing gracefully, violating the lower constraints as little as possible. And this shows you, the listener, that language differences arise from these fine -grained rankings, not from having totally different constraint sets. Okay, that clarifies the core mechanism.
Let's move into Part 2, Chapters 2 and 3. This idea of parallel evaluation, judging all candidates at once, is key to solving an old problem. Conspiracies. What were those? Conspiracies were a puzzle for the old rule -based systems. You'd see several completely different unrelated rules in a language. Maybe a deletion rule, an insertion rule, a feature -changing rule, all seeming to work together towards a single goal, like making sure every syllable starts with a consonant. The rules themselves look different, but the outcome was suspiciously similar. Exactly. The rule -based approach struggled because the rules were applied sequentially and were formally distinct. There was no way to capture the idea that they were all conspiring towards the same output target. And OT's solution? Parallelism.
Since all candidates are evaluated simultaneously against the entire ranked constraint set, you don't need multiple rules. A single high -ranked markedness constraint? The trigger that dislikes a certain structure can force some kind of repair. And the type of repair chosen depends on? It depends on which faithfulness constraint is ranked the lowest, making it the cheapest one to violate. The language takes the path of least resistance to satisfy the high -ranked trigger. Let's use the example from the book, the text -NN conspiracy. That's the constraint against having a nasal sound followed immediately by a voiceless consonant, right? Like N -T or M. Yes, text -NN marks that sequence as bad, undesirable. So if a language ranks text stand highly, it must do something to fix an input. Like input, what can it do?
Okay, what are the repair options? Well, it could just delete the nasal. That's nasal deletion. But doing that violates text stand N -I -X -I -O, the constraint that says don't delete segments. Or it could stick a vowel in between, like in view put. That's vowel parenthesis. But that violates text stand N -E -D -P -I -O, the constraint against inserting segments. What about changing the consonant? Making the voiceless P into a voiced B like im -but. That's post -nasal voicing. It satisfies texta -n -i -na -tis -na -e -hub -m, not m -bae. But it violates texta -n -a -o, the constraint demanding faithfulness to the original voicing feature of the obstruent. Okay.
Any others? Could you change the nasal itself? You could have denasalization, changing into something else, violating texta -n -i -o -nasal. Or you could have nasal substitution or fusion, where the nasal merges with the following consonant. may be taking its place of articulation. This violates text linearity I .O., which prohibits merging or reordering segments. So multiple ways to fix the text problem, each violating a different faithfulness constraint. Precisely. And OT predicts that the typology of repairs across languages comes down to how they rank these faithfulness constraints relative to the high -ranked text and trigger. Give us an example. Indonesian.
Indonesian often uses fusion or nasal substitution. Kager gives the example. tex manan pili becoming tex pili. The NP sequence is resolved by effectively merging the nasal into the P, which then services as M at the same place of articulation. This tells us that for Indonesian, tex ends ranked high, and crucially, tex linearity IO, the anti -fusion constraint, must be ranked lower than the other faithfulness constraints, like tex WIO or tex Aydin base. Fusion is the cheapest way out. Okay, contrast that with, say, puyu punggok kecoa. In Quechua, the input text complot becomes text comba. It uses boast nasal voicing.
The P becomes B. So for Quechua, text can again be high -ranked. But this time, text ident I -O, the constraint against changing voicing, must be ranked lower than the constraints against deletion, insertion, or fusion. Changing the voicing is the least costly repair strategy for Quechua. It's really elegant how shuffling those faithfulness rankings below the main trigger generates all these different language -specific behaviors. It is. It captures the functional unity fix the bad sequence while allowing for structural diversity in how it's fixed. Moving to chapter three, this same logic applies to basic syllable structure, right? Things like preferring simple syllables. Absolutely. There are universal tendencies captured by markedness constraints. Syllables generally prefer to have an onset. a starting consonant, captured by tex -tum -pult. They prefer not to have complex consonant clusters, penalized by constraints like tex -complex -a -mo -govel or tex -complex -co -de -drew.
And they often prefer simple vowel nuclei. And things like apenthesis inserting a vowel or consonant happen when satisfying one of these structural constraints is more important than avoiding insertion. Exactly. Take axon and cacampa again. It highly ranks textetic. So if an input word would surface without an onset, like a dot -a -a, the language inserts a consonant, often t, to create one, text -a -o. This shows the ranking text epic axonic -e -i -o. Avoiding an onsetless syllable is worth the cost of inserting a segment. And you mentioned earlier the idea of emergence of the unmarked with these inserted segments. Why is the inserted segment usually something simple, like t or a schwa? Because that inserted segment has no input counterpart. It doesn't have an underlying form it needs to be faithful to via text IEO faithfulness. So its features are determined purely by markedness constraints. And markedness universally favors the simplest, least specified segments. So it's unmarked because there's no memory constraint forcing it to be complex. Precisely.
Unless there's some assimilation process driven by context, the epithetic segment will default to the blandest, most basic sound available in the language's inventory. This is... Process economy changes happen only when necessary, and the change itself is minimal. Before leaving this section, we need to touch on generalized alignment from 3 .4. This seems important for connecting different parts of the grammar. It's the key formal tool for the morphology -phenology interface. Alignment constraints demand that the edges of different types of constituents line up. For instance, Alig -NLL word PRWD wants the left edge of a morphological word to coincide with the left edge of a prosodic word. phonological grouping. And these alignment constraints help explain things like where affixes go or how stress domains are determined? Exactly. They are viable rank constraints that mediate between morphological structure and phonological structure, crucial for phenomena like stress, reduplication, and affix placement. Okay, structure established.
Let's dive into part 3, chapters 4, 5, and 6. This is where OT really gets tested against complex phenomena like stress and morphological identity. Starting with Chapter 4, Metrical Structure Stress. Stress systems were a major challenge for rule -based phonology. They often required really complex, multi -step derivations to build feet, rhythmic units, and assign prominence. OT aims to do this in parallel. A fundamental constraint is culminativity, often formalized as text GRWDPRWDT. It basically says every grammatical word needs to correspond to a prosodic word, which implies it must contain at least one foot and thus primary stress. The case study mentioned is Hicks -Gariana rhythmic lengthening. How does OT handle the interaction between needing well -formed feet and changing vowel length? Hicks -Gariana shows how foot structure constraints can force changes.
It has constraints like Tex FTBN, requiring feet to be binary, two syllables or two more as. To satisfy this, the language might lengthen a vowel in a potential foot, even though that violates Tex IOET, don't insert Morris length in vowels. So the need for a well -shaped foot overrides the constraint against lengthening. If text FTBI is ranked higher than text IO, yes. Also relevant is text non -finality, a constraint that discourages stress on the very last syllable of a word. In Higgs -Cariana, the ranking text non -finality cells means it's better to leave the final syllable unstressed, unparsed into a foot, than to force every syllable into a foot if it means stressing the final one. And the key OT advantage here is... Parallelism. The old way needed rules to build feet. then later rules to fix feet by lengthening vols. OT evaluates candidates that already have the lengthening and the correct foot structure simultaneously. The winning candidate satisfies the high -ranked foot constraints from the get -go, accepting the lower -ranked lengthening violation. No intermediate repair step needed.
Got it. Eliminates that abstract middle stage. Now, chapter five, correspondence in reduplication. This is about copying parts of words, right? Like walkie -talkie, how does OT handle the copy part? This requires a whole new set of faithfulness constraints. We still have the standard input out IO faithfulness relating the underlying form to the surface form. But for reduplication, we add base reduplicant, BR faithfulness constraints. So constraints that specifically demand similarity between the original bit, the base, and the copied bit, the reduplicant. Exactly. Constraints like text they had to bear say feature F in the base must be identical to feature F in the reduplicant.
And this dual system, IOFaith and BRFaith, helps explain that emergence of the unmarked thing we saw earlier, but now in reduplicants. Why, copied bits are sometimes simpler than the original. Yes, it provides a really neat account. Imaginal language allows complex codas in general, but in the reduplicant part, those codas simplify. The ranking explanation is often this three -way interaction. First, text IOFaithfulness, like text IOFaithfulness for consonants, is ranked high, ensuring the base cues its complex coda. preserving the input information. Second, this high -ranked text IO faithfulness dominates the markedness constraint against complex coda, say, text complex code. This allows the base to surface with the marked structure. But third, text complex IO gate is ranked above the relevant text DR identity constraint. So the reduplicant is caught in the middle. Right. To satisfy the higher -ranked text complex coda, the reduplicant violates identity with the base and simplifies its coda. The unmarked structure emerges in the copy because faithfulness to the base is sacrificed for structural well -formedness. That's clever. What about those weird cases of over -application and under -application mentioned in 5 .6? What's over -application?
Over -application is when a phonological process applies in the rediblicate in a place where the conditioning environment for that process isn't actually met. The example given is often Malay nasal harmony. How does that work? Let's say nasal harmony spreads nasality onto vowels following a nasal consonant in the base. In over -application, the vowel in the reduplicant might also become nasalized, just to match the base vowel, even if the reduplicant doesn't have the triggering nasal consonant nearby. Why would it do that? It seems illogical. Because the text -ident BR nasal constraint, demanding identity for nasality between base and reduplicant, is ranked very high. It's more important for the copy to sound identical to the original in terms of nasality than it is to obey the normal local rules for when nasalization should happen. Identity trumps phonotactics. Okay, that makes sense under OT logic. Then chapter 6 extends this identity idea even further with output -to -output correspondence.
What's that alluding? Instead of base -to -reduplicant within one word, OO correspondence relates two separate, fully -formed output words that have a morphological relationship. like a full word and its shortened truncated version, or a stem and a word derived from that stem. So we have constraints like BT identity base truncated. Exactly. These demand that the truncated form maintain faithfulness to the full base form it's derived from. This can lead to underapplication, where a process that should apply in the truncated form fails to apply. precisely because applying it would make the truncated form less similar to its base. Or like keeping a vowel sound from the base, even if the truncation changes the environment. Could be. Or it can lead to over -application in derived forms, where some phonetic detail, like allophonic variation, from the stem carries over into an affixed form, even if the conditioning environment is gone, purely to maintain identity between the stem output and the derived output. And the big advantage here compared to older theories. It captures effects that used to be handled by cyclic rule application. The idea that rules applied first to inner structures, then reapplied to outer structures. OO correspondence does this without needing those abstract intermediate stages or the machinery of the cycle. It keeps the analysis closer to the surface forms we actually observe. Okay, we've covered the core phonology and morphology interface. Now, part four feels like zooming out. Chapter 7 and 8, looking at learning and extending to syntax. Let's start with learning, Chapter 7. If rankings define languages, how does a kid figure out the ranking, especially since they only hear the winners, the actual outputs? That's the huge acquisition challenge. The child gets positive evidence, only they hear bet in Dutch, but never the candidate bed labeled as wrong. How do they infer the hierarchy?
The main proposal from Tease and Smolenski involves an algorithm based on constraint demotion. Constraint demotion. How does that work? The learner basically uses the winner to figure out why the loser's lost. When they hear an optimal output, say, old dollar. for some input answer, they implicitly compare it to other potential candidates, like suboptimal dollars. If dollar beat dollars, there must be some constraint that dollar satisfies but dollar violates. And this constraint must outrank any constraint that dollar satisfies but a dollar violates. OK, so the comparison reveals crucial ranking information. Yes. Each such comparison provides evidence to demote constraints violated by the winner below, constraints violated only by the loser. The algorithm starts with all constraints unranked and gradually stratifies them based on these comparisons derived from the input data until the correct hierarchy emerges.
But what about when words change form, like the Dutch bed becoming bet sometimes and be do on other times? This is the alternation problem. How does the learner figure out the single underlying representation, bed with a D, when the surface forms vary? This requires another principle, lexicon optimization. The learner, in essence, tries out different possible URs. The goal is to choose the UR that allows the observed surface forms to be generated with the minimal overall violation cost, given the constraint hierarchy they are learning. So if the child guessed the UR for Dutch was bed, with a T, they'd have trouble explaining the plural B, de en. Big trouble.
They'd need some unusual voicing rule. Hypothesizing the UR as bed, with a D, makes the singular bed easy to explain due to high -ranked Pentext -voiced coda. And the plural bait, also straightforward, voicing is preserved when not in the coda. That UR, bed, leads to a simpler overall analysis across the paradigm. So lexicon optimization pushes the learner towards the UR that requires the least magic or the fewest faithfulness violations overall to generate all the related surface forms. Essentially, yes. The system favors URs that are maximally faithful to the range of outputs.
guided by the developing constraint ranking. There are safeguards, like the anti -loop provision mentioned in Kager, to prevent the learner from getting stuck on incorrect URs when faced with conflicting evidence. It encourages finding the most economical underlying form. Fascinating. Okay, final stretch, chapter eight, extending OT to syntax. Can this constraint ranking system really handle sentence structure, word order, movement? That's the claim, and yes, it applies the same core logic. G -E -N creates candidate syntactic structures, entire sentences or phrases.
Eval uses ranked constraints to pick the best one. The constraints are just different. They refer to syntactic well -formedness. What kinds of constraints operate in syntax? You have constraints demanding certain configurations, like text OB spec, which requires operators, like who or what, to be in a specific structural position, the specifier CP, or text OB HD, obligatory head, which says projections need heads. These function like markedness constraints. And the faithfulness side in syntax. The big one is text stay, which is basically don't move. It penalizes any movement of elements from their base position. Think of it as syntactic inertia. So syntactic movement, like asking a question in English, what did she say, happens when the need to satisfy a structural constraint outweighs the cost of movement. Exactly.
Let's trace what did Mary say. The underlined structure might have what inside the verb phrase, but TextActXFS is high -ranked in English, demanding what move to the front to specifier of CP. Okay, so what moves violating text stay. Right, but that movement leaves the head of CP empty. Now, TextOBHD is violated. Since TextOBHD is also high -ranked in English and higher than text stay, the auxiliary deed has to move into that head position to satisfy it. And that's the inversion we see, but moving did violates text decade.
Correct. The winning candidate violates text twice, but crucially it satisfies the higher rank structural demands of text OB spec and text OB HD. Any candidate that didn't move what or did would fatally violate those top constraints. So the English ranking is something like text OB spec, text OB HD. That's the idea. And this naturally explains syntactic economy. You only move things as much as absolutely necessary to satisfy higher -ranked constraints. Any extra unnecessary movement would just incur extra text violations for no gain, so that candidate would lose. And changing the ranking explains different language types, like languages without Y movement.
Precisely. If a language ranked higher than text type X, then moving the wayward would be too costly. The optimal candidate would leave it in place, Y in situ. as you see in languages like Chinese or Japanese. The factorial typology, all possible rankings, predicts the range of observed syntactic variation. Wow. Okay, let's try to wrap this all up. This deep dive really shows OT as a fundamental shift. Definitely. It threw out the step -by -step derivations and replaced them with this parallel evaluation system driven by universal but viable and ranked constraints.
The core tension is always markedness versus faithfulness structure versus memory. And the key insight is that All the differences between languages boil down to just shuffling the priority list, the ranking of these universal constraints. Strict domination makes that ranking absolute. And we saw how this accounts for so much phonological patterns, conspiracies, stress, reduplication complexities via BR and OO correspondence, even syntactic movement. And it offers a formal model for how kids could learn these rankings via constraint demotion and lexicon optimization. But as the sources acknowledge, especially towards the end, it's not perfect. The standard model, this direct input output mapping, runs into trouble with certain phenomena, particularly something called opacity. Right. Opacity is tricky. It's when the surface form looks like it must have gone through some intermediate stage that isn't transparent anymore. Like a rule seems to have applied, but its effects are somehow masked or unexpected in the final output, making the winner look suboptimal according to the straightforward ranking. So the elegant direct mapping breaks down a bit.
It seems to in these cases. Standard OT struggles because it doesn't have intermediate stages. This has pushed the theory to evolve. Leading to more advanced, maybe more complex versions of OT. Exactly. Things like sympathy theory, where the winner has to bear some resemblance to another sympathetic candidate, even if that candidate isn't the true optimal one. Or... Local conjunction, where violating two specific constraints at the same time is much worse than violating them individually. These are ways to handle opacity without bringing back full serial derivations.
So even this revolutionary theory is constantly being tested and refined by the sheer complexity of language. Always. That's linguistics. It's incredible to think about. So for you listening, the next time you speak or hear language, remember this hidden battle. Countless universal constraints pushing for simplicity, pulling for faithfulness, all evaluated in parallel, with the winner being the single utterance that emerges, having minimally violated the specific priority list that defines your language. A constant, dynamic conflict shaping every sound, word, and sentence. It's quite something.