1 00:00:00,000 --> 00:00:02,919 Welcome to the Deep Dive. Today, we're tackling 2 00:00:02,919 --> 00:00:05,139 something really foundational in linguistics, 3 00:00:05,299 --> 00:00:08,039 a theory that, well, it really shook things up. 4 00:00:08,080 --> 00:00:10,539 We're doing a deep dive into optimality theory, 5 00:00:10,839 --> 00:00:14,259 OT for short. That's right. For a long time, 6 00:00:14,339 --> 00:00:16,179 you know, linguistics, especially the generative 7 00:00:16,179 --> 00:00:18,660 folks, thought language worked like a computer 8 00:00:18,660 --> 00:00:22,379 program, a series of ordered steps, these rewrite 9 00:00:22,379 --> 00:00:25,300 rules, taking you from some deep form to what 10 00:00:25,300 --> 00:00:27,620 you actually say. Like a recipe, step by step. 11 00:00:27,839 --> 00:00:30,480 bake the cake this way, exactly this way. Exactly. 12 00:00:30,679 --> 00:00:32,320 And OT just comes along and says, well, what 13 00:00:32,320 --> 00:00:34,719 if there's no recipe, no ordered steps? It keeps 14 00:00:34,719 --> 00:00:36,920 the big goal finding universal language principles, 15 00:00:37,100 --> 00:00:39,619 but totally flips the script on how we get there. 16 00:00:39,799 --> 00:00:42,640 Okay, so no more serial derivations. What replaces 17 00:00:42,640 --> 00:00:46,460 them? Conflict. At its heart, OT. and we're drawing 18 00:00:46,460 --> 00:00:49,259 a lot here from Kager's work, which really lays 19 00:00:49,259 --> 00:00:51,460 it out, sees language as a system of competing 20 00:00:51,460 --> 00:00:53,719 forces. These are called constraints. Constraints. 21 00:00:53,820 --> 00:00:55,679 And these constraints, they define what makes 22 00:00:55,679 --> 00:00:58,159 a good word or sentence structurally. Here's 23 00:00:58,159 --> 00:01:01,700 the kicker. They're universal, found in all languages, 24 00:01:01,899 --> 00:01:06,299 but they can be broken. They're viable. Hold 25 00:01:06,299 --> 00:01:10,200 on. Universal, but viable. That sounds like a 26 00:01:10,200 --> 00:01:12,579 contradiction. If they apply everywhere, how 27 00:01:12,579 --> 00:01:14,319 do languages end up so different? Why don't we 28 00:01:14,319 --> 00:01:16,120 all sound the same if the constraints are the 29 00:01:16,120 --> 00:01:18,920 same? Ah, that's the genius of it. It's not the 30 00:01:18,920 --> 00:01:20,620 constraints themselves that differ, it's their 31 00:01:20,620 --> 00:01:23,379 hierarchical ranking. Think of it like a priority 32 00:01:23,379 --> 00:01:26,140 list. Every language has the same list of rules, 33 00:01:26,319 --> 00:01:29,040 but each language shuffles that list, deciding 34 00:01:29,040 --> 00:01:31,359 which rules are more important than others. So 35 00:01:31,359 --> 00:01:33,900 language A might say, constraint X is absolutely 36 00:01:33,900 --> 00:01:36,859 vital, never break it. While language B says, 37 00:01:36,980 --> 00:01:40,640 eh, constraint X is okay, but constraint Y, that's 38 00:01:40,640 --> 00:01:43,099 the one we really care about. Precisely. Constraint 39 00:01:43,099 --> 00:01:45,359 A dominates B in one language, B dominates A 40 00:01:45,359 --> 00:01:47,659 in another. And that simple re -ranking creates 41 00:01:47,659 --> 00:01:50,040 radically different patterns, all from the same 42 00:01:50,040 --> 00:01:52,159 universal toolkit. Okay, let's get into the big 43 00:01:52,159 --> 00:01:54,040 players in this conflict. You mentioned these 44 00:01:54,040 --> 00:01:56,340 competing forces. The sources talk about two 45 00:01:56,340 --> 00:01:59,120 main families of constraints, right? Yes, the 46 00:01:59,120 --> 00:02:02,599 two titans constantly at odds. Markedness, M, 47 00:02:02,799 --> 00:02:06,049 and faithfulness, F. Markedness and faithfulness. 48 00:02:06,109 --> 00:02:08,789 Let's break those down. Markedness. Markedness 49 00:02:08,789 --> 00:02:10,949 constraints are all about the output. They push 50 00:02:10,949 --> 00:02:13,490 for simplicity, for structures that are easy 51 00:02:13,490 --> 00:02:16,650 to produce or perceive. Think simple syllables, 52 00:02:16,969 --> 00:02:20,949 common sounds, things that are, well, unmarked, 53 00:02:20,949 --> 00:02:24,310 less complex. Okay, so markedness is like the 54 00:02:24,310 --> 00:02:26,289 efficiency expert, always trying to streamline 55 00:02:26,289 --> 00:02:29,229 things. What about faithfulness? Faithfulness 56 00:02:29,229 --> 00:02:31,550 is the opposite pull. It's about the input, the 57 00:02:31,550 --> 00:02:34,460 underlying form in the lexicon. Faithfulness 58 00:02:34,460 --> 00:02:36,879 constraints demand that the output, what you 59 00:02:36,879 --> 00:02:39,979 actually say, stays true to the input. It preserves 60 00:02:39,979 --> 00:02:42,039 the information, the contrast stored in your 61 00:02:42,039 --> 00:02:45,250 mental dictionary. So markedness wants to simplify, 62 00:02:45,449 --> 00:02:48,169 maybe even change things from the input to make 63 00:02:48,169 --> 00:02:50,509 the output cleaner. But faithfulness says, whoa, 64 00:02:50,610 --> 00:02:52,430 hang on. You need to respect the original form. 65 00:02:52,530 --> 00:02:54,629 Don't lose that information. They're fundamentally 66 00:02:54,629 --> 00:02:57,349 opposed. Absolutely. You can't usually satisfy 67 00:02:57,349 --> 00:03:00,849 both perfectly. If you obey a strong markedness 68 00:03:00,849 --> 00:03:03,210 constraint and simplify the output, you're probably 69 00:03:03,210 --> 00:03:05,270 violating some faithfulness constraint by changing 70 00:03:05,270 --> 00:03:07,729 the input form. And if you stick rigidly to the 71 00:03:07,729 --> 00:03:10,289 input to satisfy faithfulness, you might end 72 00:03:10,289 --> 00:03:12,669 up with a really complex marked output structure 73 00:03:12,669 --> 00:03:15,889 that violates markedness. Exactly. The specific 74 00:03:15,889 --> 00:03:19,250 ranking in a given language dictates who wins 75 00:03:19,250 --> 00:03:21,909 that fight in any given situation. So our mission 76 00:03:21,909 --> 00:03:24,270 today is to really unpack this. We'll go through 77 00:03:24,270 --> 00:03:26,270 the architecture, see how this constraint interaction 78 00:03:26,270 --> 00:03:29,270 plays out in sounds, syllables, even sentences, 79 00:03:29,389 --> 00:03:31,930 basically following Cager's structure chapter 80 00:03:31,930 --> 00:03:33,949 by chapter. Let's start right at the beginning, 81 00:03:34,129 --> 00:03:37,539 chapter one. the basic architecture, OT is described 82 00:03:37,539 --> 00:03:40,259 as an input -output system. Three main parts. 83 00:03:40,840 --> 00:03:44,939 Lexicon, generator, evaluator. What's new here, 84 00:03:45,020 --> 00:03:47,000 especially with the lexicon? Well, the lexicon 85 00:03:47,000 --> 00:03:49,219 itself contains the morphemes, the basic building 86 00:03:49,219 --> 00:03:51,939 blocks, but OT makes this assumption called richness 87 00:03:51,939 --> 00:03:55,500 of the base, ROTB. And honestly, this was pretty 88 00:03:55,500 --> 00:03:57,879 radical compared to older models. Radical how? 89 00:03:58,219 --> 00:04:00,180 It assumes that the input forms, the lexical 90 00:04:00,180 --> 00:04:02,740 representations, are completely unconstrained. 91 00:04:02,800 --> 00:04:05,539 Anything goes. Your mental dictionary, in theory, 92 00:04:05,699 --> 00:04:07,919 could contain any sound combination imaginable, 93 00:04:07,979 --> 00:04:10,259 even ones that never show up in the actual spoken 94 00:04:10,259 --> 00:04:12,759 language. Wait, so my brain might have an entry 95 00:04:12,759 --> 00:04:14,479 for, I don't know, a word, starting with NZ, 96 00:04:14,680 --> 00:04:16,660 even though English never allows that? According 97 00:04:16,660 --> 00:04:19,939 to ROTB, yes. It gets rid of the need for those 98 00:04:19,939 --> 00:04:22,660 old morphine structure constraints, rules that 99 00:04:22,660 --> 00:04:25,839 basically filter the dictionary. OT says, don't 100 00:04:25,839 --> 00:04:27,720 filter the dictionary. Let the main constraint 101 00:04:27,720 --> 00:04:29,899 system do all the work of filtering what actually 102 00:04:29,899 --> 00:04:32,560 comes out. It avoids stating restrictions twice. 103 00:04:32,920 --> 00:04:35,519 Okay, so the input is this wild, unrestricted 104 00:04:35,519 --> 00:04:38,160 thing. What happens next? It goes to the generator, 105 00:04:38,360 --> 00:04:42,839 or GEN. GEN's job is conceptually simple, but 106 00:04:42,839 --> 00:04:46,639 vast. For any single input, it turns out a potentially 107 00:04:46,639 --> 00:04:49,279 infinite set of output candidates. Infinite? 108 00:04:49,459 --> 00:04:51,889 How? Well, it considers everything. Deleting 109 00:04:51,889 --> 00:04:54,490 sounds, adding sounds, changing features, moving 110 00:04:54,490 --> 00:04:56,730 things around, leaving things exactly as they 111 00:04:56,730 --> 00:04:59,910 are. Every possible modification or lack thereof. 112 00:05:00,189 --> 00:05:02,470 Freedom of analysis, they call it. That sounds 113 00:05:02,470 --> 00:05:05,389 computationally explosive. If GEN throws infinite 114 00:05:05,389 --> 00:05:07,870 options onto the table, you need a really decisive 115 00:05:07,870 --> 00:05:10,810 judge. That must be the evaluator, Eval. That's 116 00:05:10,810 --> 00:05:13,709 Eval. It's the core of the grammar. Eval takes 117 00:05:13,709 --> 00:05:16,230 that specific language's ranked list of constraints 118 00:05:16,230 --> 00:05:18,689 and uses it to assess all those candidates from 119 00:05:18,689 --> 00:05:21,529 GEN. Its job is to find the one optimal candidate. 120 00:05:21,670 --> 00:05:23,730 The winner. The one that actually gets pronounced. 121 00:05:24,110 --> 00:05:26,509 Right. The one that best satisfies the hierarchy, 122 00:05:26,750 --> 00:05:29,430 meaning it incurs the least serious violations. 123 00:05:29,850 --> 00:05:32,310 And the way it judges is absolutely crucial. 124 00:05:33,480 --> 00:05:36,060 Strict domination. Strict domination. OK, explain 125 00:05:36,060 --> 00:05:39,100 that. Because my first instinct is, why not just 126 00:05:39,100 --> 00:05:41,899 count violations? Candidate A breaks five low 127 00:05:41,899 --> 00:05:44,920 -ranked rules. Candidate B breaks one high -ranked 128 00:05:44,920 --> 00:05:48,319 rule. Maybe A is better overall. Ah, but OT says 129 00:05:48,319 --> 00:05:51,480 no. Strict domination means the hierarchy is 130 00:05:51,480 --> 00:05:54,079 absolute. If a candidate violates a constraint, 131 00:05:54,279 --> 00:05:57,240 even just once, any candidate that doesn't violate 132 00:05:57,240 --> 00:05:59,839 that constraint is immediately better, no matter 133 00:05:59,839 --> 00:06:01,860 how many lower -ranked constraints the first 134 00:06:01,860 --> 00:06:04,060 candidate satisfies perfectly. So one strike 135 00:06:04,060 --> 00:06:06,379 on a top -level constraint and you're out, regardless 136 00:06:06,379 --> 00:06:08,339 of anything else. You're out of the running against 137 00:06:08,339 --> 00:06:10,519 any competitor that satisfies that top constraint. 138 00:06:10,660 --> 00:06:13,500 It's merciless. A higher -ranked constraint's 139 00:06:13,500 --> 00:06:15,819 decision is final for comparing any two candidates. 140 00:06:16,300 --> 00:06:18,500 That rigidity is what gives each language its 141 00:06:18,500 --> 00:06:20,860 specific profile. Adding up violations would 142 00:06:20,860 --> 00:06:22,980 just muddy the waters. That makes sense. That 143 00:06:22,980 --> 00:06:25,420 absolute priority is how you get different languages. 144 00:06:25,899 --> 00:06:28,459 Let's make this concrete with that classic example. 145 00:06:28,920 --> 00:06:32,279 Voice sounds at the end of syllables, like in 146 00:06:32,279 --> 00:06:35,040 Dutch versus English. Perfect example. It pits 147 00:06:35,040 --> 00:06:37,980 faithfulness against markedness. The markedness 148 00:06:37,980 --> 00:06:39,819 constraint here is often called something like 149 00:06:39,819 --> 00:06:43,589 text. Voiced coda. Basically, don't have voiced 150 00:06:43,589 --> 00:06:46,170 obstruents like B, D, G at the end of a syllable. 151 00:06:46,410 --> 00:06:49,110 It prefers voiceless ones like P, T, K there. 152 00:06:49,269 --> 00:06:51,870 Because voiceless ones are maybe simpler, less 153 00:06:51,870 --> 00:06:54,029 effort. That's the idea behind markedness, yes. 154 00:06:54,389 --> 00:06:56,449 Simpler, more common structures are preferred. 155 00:06:56,670 --> 00:06:58,990 The competing constraint is faithfulness, specifically 156 00:06:58,990 --> 00:07:02,230 text, ident, IO, voiced A. It says, hey, if the 157 00:07:02,230 --> 00:07:04,250 input segment was voiced, the output segment 158 00:07:04,250 --> 00:07:06,110 should be voiced too. Preserve that feature. 159 00:07:06,410 --> 00:07:08,569 Okay, so in English, if I have the word bed in 160 00:07:08,569 --> 00:07:10,930 my lexicon. The output is bed. You pronounce 161 00:07:10,930 --> 00:07:13,490 the D. English tolerates that voice sound in 162 00:07:13,490 --> 00:07:16,089 the coda. Which means English prioritizes faithfulness 163 00:07:16,089 --> 00:07:19,670 here. Exactly. The ranking is text, iodine, IO. 164 00:07:20,209 --> 00:07:22,769 Keeping the output true to the input D is more 165 00:07:22,769 --> 00:07:24,509 important than avoiding the market voice coda. 166 00:07:24,810 --> 00:07:27,689 So the contrast between bet and bet is maintained. 167 00:07:28,629 --> 00:07:31,029 Faithfulness wins. But Dutch does the opposite, 168 00:07:31,189 --> 00:07:34,050 right? Right. The same input bed comes out as 169 00:07:34,050 --> 00:07:36,689 bet in Dutch. The final D becomes a voiceless 170 00:07:36,689 --> 00:07:40,290 T. So for Dutch. The ranking is flipped. Avoiding 171 00:07:40,290 --> 00:07:43,209 that marked voice code is the top priority. The 172 00:07:43,209 --> 00:07:45,350 language is willing to sacrifice faithfulness, 173 00:07:45,470 --> 00:07:48,689 change the voicing feature to satisfy that higher 174 00:07:48,689 --> 00:07:51,189 ranked markedness constraint. And that leads 175 00:07:51,189 --> 00:07:53,689 to positional neutralization. The difference 176 00:07:53,689 --> 00:07:57,490 between underlying D and T is lost specifically 177 00:07:57,490 --> 00:07:59,689 in that code of position. Precisely. And here's 178 00:07:59,689 --> 00:08:02,209 a subtle but important point Kager makes. Even 179 00:08:02,209 --> 00:08:04,009 though text edit is ranked lower in Dutch and 180 00:08:04,009 --> 00:08:07,269 gets violated, it's still active. Imagine GEN. 181 00:08:07,689 --> 00:08:09,910 also produced a candidate -like pet for input 182 00:08:09,910 --> 00:08:13,730 bed. That candidate pet also satisfies the top 183 00:08:13,730 --> 00:08:16,509 constraint, text voice to coda. But compared 184 00:08:16,509 --> 00:08:18,829 to the actual winner bet, pet is worse because 185 00:08:18,829 --> 00:08:21,589 it violates two faithfulness constraints. It 186 00:08:21,589 --> 00:08:23,850 changes voicing and it changes the initial B 187 00:08:23,850 --> 00:08:27,670 to P. The real output bet only violates one faithfulness 188 00:08:27,670 --> 00:08:31,370 constraint, voicing. Ah, so even dominated constraints 189 00:08:31,370 --> 00:08:33,950 enforce minimal violation. You break the rules 190 00:08:33,950 --> 00:08:36,669 only as much as absolutely necessary to satisfy 191 00:08:36,669 --> 00:08:39,090 the higher -ups. You got it. It's not just about 192 00:08:39,090 --> 00:08:41,289 winning or losing. It's about losing gracefully, 193 00:08:41,389 --> 00:08:43,750 violating the lower constraints as little as 194 00:08:43,750 --> 00:08:46,429 possible. And this shows you, the listener, that 195 00:08:46,429 --> 00:08:48,629 language differences arise from these fine -grained 196 00:08:48,629 --> 00:08:50,750 rankings, not from having totally different constraint 197 00:08:50,750 --> 00:08:53,090 sets. Okay, that clarifies the core mechanism. 198 00:08:53,309 --> 00:08:55,669 Let's move into Part 2, Chapters 2 and 3. This 199 00:08:55,669 --> 00:08:58,149 idea of parallel evaluation, judging all candidates 200 00:08:58,149 --> 00:09:01,269 at once, is key to solving an old problem. Conspiracies. 201 00:09:01,269 --> 00:09:03,830 What were those? Conspiracies were a puzzle for 202 00:09:03,830 --> 00:09:06,769 the old rule -based systems. You'd see several 203 00:09:06,769 --> 00:09:09,490 completely different unrelated rules in a language. 204 00:09:09,850 --> 00:09:13,490 Maybe a deletion rule, an insertion rule, a feature 205 00:09:13,490 --> 00:09:15,990 -changing rule, all seeming to work together 206 00:09:15,990 --> 00:09:18,750 towards a single goal, like making sure every 207 00:09:18,750 --> 00:09:21,299 syllable starts with a consonant. The rules themselves 208 00:09:21,299 --> 00:09:24,100 look different, but the outcome was suspiciously 209 00:09:24,100 --> 00:09:27,419 similar. Exactly. The rule -based approach struggled 210 00:09:27,419 --> 00:09:29,899 because the rules were applied sequentially and 211 00:09:29,899 --> 00:09:32,679 were formally distinct. There was no way to capture 212 00:09:32,679 --> 00:09:34,820 the idea that they were all conspiring towards 213 00:09:34,820 --> 00:09:38,720 the same output target. And OT's solution? Parallelism. 214 00:09:38,899 --> 00:09:41,460 Since all candidates are evaluated simultaneously 215 00:09:41,460 --> 00:09:44,500 against the entire ranked constraint set, you 216 00:09:44,500 --> 00:09:47,480 don't need multiple rules. A single high -ranked 217 00:09:47,480 --> 00:09:50,090 markedness constraint? The trigger that dislikes 218 00:09:50,090 --> 00:09:52,250 a certain structure can force some kind of repair. 219 00:09:52,470 --> 00:09:54,690 And the type of repair chosen depends on? It 220 00:09:54,690 --> 00:09:56,710 depends on which faithfulness constraint is ranked 221 00:09:56,710 --> 00:09:59,049 the lowest, making it the cheapest one to violate. 222 00:09:59,269 --> 00:10:01,750 The language takes the path of least resistance 223 00:10:01,750 --> 00:10:04,409 to satisfy the high -ranked trigger. Let's use 224 00:10:04,409 --> 00:10:06,370 the example from the book, the text -NN conspiracy. 225 00:10:07,350 --> 00:10:09,470 That's the constraint against having a nasal 226 00:10:09,470 --> 00:10:11,970 sound followed immediately by a voiceless consonant, 227 00:10:12,169 --> 00:10:15,870 right? Like N -T or M. Yes, text -NN marks that 228 00:10:15,870 --> 00:10:18,730 sequence as bad, undesirable. So if a language 229 00:10:18,730 --> 00:10:21,210 ranks text stand highly, it must do something 230 00:10:21,210 --> 00:10:23,909 to fix an input. Like input, what can it do? 231 00:10:24,149 --> 00:10:26,190 Okay, what are the repair options? Well, it could 232 00:10:26,190 --> 00:10:28,409 just delete the nasal. That's nasal deletion. 233 00:10:28,690 --> 00:10:31,330 But doing that violates text stand N -I -X -I 234 00:10:31,330 --> 00:10:33,049 -O, the constraint that says don't delete segments. 235 00:10:33,470 --> 00:10:35,490 Or it could stick a vowel in between, like in 236 00:10:35,490 --> 00:10:38,070 view put. That's vowel parenthesis. But that 237 00:10:38,070 --> 00:10:40,190 violates text stand N -E -D -P -I -O, the constraint 238 00:10:40,190 --> 00:10:42,470 against inserting segments. What about changing 239 00:10:42,470 --> 00:10:45,220 the consonant? Making the voiceless P into a 240 00:10:45,220 --> 00:10:48,000 voiced B like im -but. That's post -nasal voicing. 241 00:10:48,120 --> 00:10:49,840 It satisfies texta -n -i -na -tis -na -e -hub 242 00:10:49,840 --> 00:10:53,200 -m, not m -bae. But it violates texta -n -a -o, 243 00:10:53,519 --> 00:10:55,820 the constraint demanding faithfulness to the 244 00:10:55,820 --> 00:10:58,139 original voicing feature of the obstruent. Okay. 245 00:10:58,200 --> 00:11:00,340 Any others? Could you change the nasal itself? 246 00:11:00,740 --> 00:11:03,120 You could have denasalization, changing into 247 00:11:03,120 --> 00:11:06,039 something else, violating texta -n -i -o -nasal. 248 00:11:06,059 --> 00:11:08,740 Or you could have nasal substitution or fusion, 249 00:11:09,039 --> 00:11:11,340 where the nasal merges with the following consonant. 250 00:11:11,720 --> 00:11:13,960 may be taking its place of articulation. This 251 00:11:13,960 --> 00:11:16,879 violates text linearity I .O., which prohibits 252 00:11:16,879 --> 00:11:20,139 merging or reordering segments. So multiple ways 253 00:11:20,139 --> 00:11:22,679 to fix the text problem, each violating a different 254 00:11:22,679 --> 00:11:25,980 faithfulness constraint. Precisely. And OT predicts 255 00:11:25,980 --> 00:11:28,360 that the typology of repairs across languages 256 00:11:28,360 --> 00:11:31,059 comes down to how they rank these faithfulness 257 00:11:31,059 --> 00:11:33,500 constraints relative to the high -ranked text 258 00:11:33,500 --> 00:11:35,320 and trigger. Give us an example. Indonesian. 259 00:11:35,600 --> 00:11:38,820 Indonesian often uses fusion or nasal substitution. 260 00:11:39,399 --> 00:11:42,159 Kager gives the example. tex manan pili becoming 261 00:11:42,159 --> 00:11:45,399 tex pili. The NP sequence is resolved by effectively 262 00:11:45,399 --> 00:11:48,259 merging the nasal into the P, which then services 263 00:11:48,259 --> 00:11:51,399 as M at the same place of articulation. This 264 00:11:51,399 --> 00:11:53,299 tells us that for Indonesian, tex ends ranked 265 00:11:53,299 --> 00:11:57,279 high, and crucially, tex linearity IO, the anti 266 00:11:57,279 --> 00:11:59,639 -fusion constraint, must be ranked lower than 267 00:11:59,639 --> 00:12:01,440 the other faithfulness constraints, like tex 268 00:12:01,440 --> 00:12:04,419 WIO or tex Aydin base. Fusion is the cheapest 269 00:12:04,419 --> 00:12:07,320 way out. Okay, contrast that with, say, puyu 270 00:12:07,320 --> 00:12:10,730 punggok kecoa. In Quechua, the input text complot 271 00:12:10,730 --> 00:12:13,590 becomes text comba. It uses boast nasal voicing. 272 00:12:13,750 --> 00:12:17,409 The P becomes B. So for Quechua, text can again 273 00:12:17,409 --> 00:12:20,750 be high -ranked. But this time, text ident I 274 00:12:20,750 --> 00:12:22,870 -O, the constraint against changing voicing, 275 00:12:23,009 --> 00:12:25,090 must be ranked lower than the constraints against 276 00:12:25,090 --> 00:12:28,539 deletion, insertion, or fusion. Changing the 277 00:12:28,539 --> 00:12:31,320 voicing is the least costly repair strategy for 278 00:12:31,320 --> 00:12:34,059 Quechua. It's really elegant how shuffling those 279 00:12:34,059 --> 00:12:36,059 faithfulness rankings below the main trigger 280 00:12:36,059 --> 00:12:38,179 generates all these different language -specific 281 00:12:38,179 --> 00:12:40,559 behaviors. It is. It captures the functional 282 00:12:40,559 --> 00:12:43,700 unity fix the bad sequence while allowing for 283 00:12:43,700 --> 00:12:45,659 structural diversity in how it's fixed. Moving 284 00:12:45,659 --> 00:12:47,980 to chapter three, this same logic applies to 285 00:12:47,980 --> 00:12:50,039 basic syllable structure, right? Things like 286 00:12:50,039 --> 00:12:52,259 preferring simple syllables. Absolutely. There 287 00:12:52,259 --> 00:12:54,620 are universal tendencies captured by markedness 288 00:12:54,620 --> 00:12:56,980 constraints. Syllables generally prefer to have 289 00:12:56,980 --> 00:12:59,879 an onset. a starting consonant, captured by tex 290 00:12:59,879 --> 00:13:02,360 -tum -pult. They prefer not to have complex consonant 291 00:13:02,360 --> 00:13:05,539 clusters, penalized by constraints like tex -complex 292 00:13:05,539 --> 00:13:07,720 -a -mo -govel or tex -complex -co -de -drew. 293 00:13:07,820 --> 00:13:10,440 And they often prefer simple vowel nuclei. And 294 00:13:10,440 --> 00:13:13,539 things like apenthesis inserting a vowel or consonant 295 00:13:13,539 --> 00:13:16,039 happen when satisfying one of these structural 296 00:13:16,039 --> 00:13:18,620 constraints is more important than avoiding insertion. 297 00:13:18,740 --> 00:13:21,820 Exactly. Take axon and cacampa again. It highly 298 00:13:21,820 --> 00:13:25,200 ranks textetic. So if an input word would surface 299 00:13:25,200 --> 00:13:28,210 without an onset, like a dot -a -a, the language 300 00:13:28,210 --> 00:13:30,710 inserts a consonant, often t, to create one, 301 00:13:30,889 --> 00:13:34,129 text -a -o. This shows the ranking text epic 302 00:13:34,129 --> 00:13:37,470 axonic -e -i -o. Avoiding an onsetless syllable 303 00:13:37,470 --> 00:13:40,250 is worth the cost of inserting a segment. And 304 00:13:40,250 --> 00:13:42,830 you mentioned earlier the idea of emergence of 305 00:13:42,830 --> 00:13:45,450 the unmarked with these inserted segments. Why 306 00:13:45,450 --> 00:13:47,509 is the inserted segment usually something simple, 307 00:13:47,590 --> 00:13:50,549 like t or a schwa? Because that inserted segment 308 00:13:50,549 --> 00:13:53,070 has no input counterpart. It doesn't have an 309 00:13:53,070 --> 00:13:55,549 underlying form it needs to be faithful to via 310 00:13:55,549 --> 00:13:58,009 text IEO faithfulness. So its features are determined 311 00:13:58,009 --> 00:14:00,769 purely by markedness constraints. And markedness 312 00:14:00,769 --> 00:14:03,169 universally favors the simplest, least specified 313 00:14:03,169 --> 00:14:05,190 segments. So it's unmarked because there's no 314 00:14:05,190 --> 00:14:08,190 memory constraint forcing it to be complex. Precisely. 315 00:14:08,370 --> 00:14:10,809 Unless there's some assimilation process driven 316 00:14:10,809 --> 00:14:13,669 by context, the epithetic segment will default 317 00:14:13,669 --> 00:14:16,649 to the blandest, most basic sound available in 318 00:14:16,649 --> 00:14:20,029 the language's inventory. This is... Process 319 00:14:20,029 --> 00:14:22,669 economy changes happen only when necessary, and 320 00:14:22,669 --> 00:14:25,230 the change itself is minimal. Before leaving 321 00:14:25,230 --> 00:14:26,929 this section, we need to touch on generalized 322 00:14:26,929 --> 00:14:30,029 alignment from 3 .4. This seems important for 323 00:14:30,029 --> 00:14:32,029 connecting different parts of the grammar. It's 324 00:14:32,029 --> 00:14:34,769 the key formal tool for the morphology -phenology 325 00:14:34,769 --> 00:14:37,529 interface. Alignment constraints demand that 326 00:14:37,529 --> 00:14:39,289 the edges of different types of constituents 327 00:14:39,289 --> 00:14:44,809 line up. For instance, Alig -NLL word PRWD wants 328 00:14:44,809 --> 00:14:47,450 the left edge of a morphological word to coincide 329 00:14:47,450 --> 00:14:50,240 with the left edge of a prosodic word. phonological 330 00:14:50,240 --> 00:14:52,519 grouping. And these alignment constraints help 331 00:14:52,519 --> 00:14:55,620 explain things like where affixes go or how stress 332 00:14:55,620 --> 00:14:58,039 domains are determined? Exactly. They are viable 333 00:14:58,039 --> 00:15:00,639 rank constraints that mediate between morphological 334 00:15:00,639 --> 00:15:02,919 structure and phonological structure, crucial 335 00:15:02,919 --> 00:15:05,700 for phenomena like stress, reduplication, and 336 00:15:05,700 --> 00:15:07,899 affix placement. Okay, structure established. 337 00:15:08,340 --> 00:15:11,700 Let's dive into part 3, chapters 4, 5, and 6. 338 00:15:11,980 --> 00:15:14,840 This is where OT really gets tested against complex 339 00:15:14,840 --> 00:15:17,100 phenomena like stress and morphological identity. 340 00:15:17,690 --> 00:15:20,289 Starting with Chapter 4, Metrical Structure Stress. 341 00:15:20,610 --> 00:15:22,830 Stress systems were a major challenge for rule 342 00:15:22,830 --> 00:15:25,110 -based phonology. They often required really 343 00:15:25,110 --> 00:15:27,629 complex, multi -step derivations to build feet, 344 00:15:27,769 --> 00:15:30,970 rhythmic units, and assign prominence. OT aims 345 00:15:30,970 --> 00:15:33,769 to do this in parallel. A fundamental constraint 346 00:15:33,769 --> 00:15:38,090 is culminativity, often formalized as text GRWDPRWDT. 347 00:15:38,110 --> 00:15:40,549 It basically says every grammatical word needs 348 00:15:40,549 --> 00:15:42,929 to correspond to a prosodic word, which implies 349 00:15:42,929 --> 00:15:45,970 it must contain at least one foot and thus primary 350 00:15:45,970 --> 00:15:48,309 stress. The case study mentioned is Hicks -Gariana 351 00:15:48,309 --> 00:15:50,629 rhythmic lengthening. How does OT handle the 352 00:15:50,629 --> 00:15:52,649 interaction between needing well -formed feet 353 00:15:52,649 --> 00:15:54,850 and changing vowel length? Hicks -Gariana shows 354 00:15:54,850 --> 00:15:56,950 how foot structure constraints can force changes. 355 00:15:57,210 --> 00:16:00,470 It has constraints like Tex FTBN, requiring feet 356 00:16:00,470 --> 00:16:03,330 to be binary, two syllables or two more as. To 357 00:16:03,330 --> 00:16:05,309 satisfy this, the language might lengthen a vowel 358 00:16:05,309 --> 00:16:07,429 in a potential foot, even though that violates 359 00:16:07,429 --> 00:16:10,070 Tex IOET, don't insert Morris length in vowels. 360 00:16:10,289 --> 00:16:12,570 So the need for a well -shaped foot overrides 361 00:16:12,570 --> 00:16:15,730 the constraint against lengthening. If text FTBI 362 00:16:15,730 --> 00:16:19,950 is ranked higher than text IO, yes. Also relevant 363 00:16:19,950 --> 00:16:23,009 is text non -finality, a constraint that discourages 364 00:16:23,009 --> 00:16:26,330 stress on the very last syllable of a word. In 365 00:16:26,330 --> 00:16:28,909 Higgs -Cariana, the ranking text non -finality 366 00:16:28,909 --> 00:16:31,129 cells means it's better to leave the final syllable 367 00:16:31,129 --> 00:16:34,429 unstressed, unparsed into a foot, than to force 368 00:16:34,429 --> 00:16:36,490 every syllable into a foot if it means stressing 369 00:16:36,490 --> 00:16:39,190 the final one. And the key OT advantage here 370 00:16:39,190 --> 00:16:42,950 is... Parallelism. The old way needed rules to 371 00:16:42,950 --> 00:16:46,000 build feet. then later rules to fix feet by lengthening 372 00:16:46,000 --> 00:16:49,320 vols. OT evaluates candidates that already have 373 00:16:49,320 --> 00:16:51,019 the lengthening and the correct foot structure 374 00:16:51,019 --> 00:16:53,740 simultaneously. The winning candidate satisfies 375 00:16:53,740 --> 00:16:55,700 the high -ranked foot constraints from the get 376 00:16:55,700 --> 00:16:57,759 -go, accepting the lower -ranked lengthening 377 00:16:57,759 --> 00:17:00,399 violation. No intermediate repair step needed. 378 00:17:00,639 --> 00:17:02,860 Got it. Eliminates that abstract middle stage. 379 00:17:03,100 --> 00:17:05,619 Now, chapter five, correspondence in reduplication. 380 00:17:05,619 --> 00:17:07,559 This is about copying parts of words, right? 381 00:17:07,619 --> 00:17:09,700 Like walkie -talkie, how does OT handle the copy 382 00:17:09,700 --> 00:17:12,170 part? This requires a whole new set of faithfulness 383 00:17:12,170 --> 00:17:15,650 constraints. We still have the standard input 384 00:17:15,650 --> 00:17:19,049 out IO faithfulness relating the underlying form 385 00:17:19,049 --> 00:17:21,769 to the surface form. But for reduplication, we 386 00:17:21,769 --> 00:17:24,170 add base reduplicant, BR faithfulness constraints. 387 00:17:24,490 --> 00:17:27,309 So constraints that specifically demand similarity 388 00:17:27,309 --> 00:17:30,490 between the original bit, the base, and the copied 389 00:17:30,490 --> 00:17:32,849 bit, the reduplicant. Exactly. Constraints like 390 00:17:32,849 --> 00:17:35,069 text they had to bear say feature F in the base 391 00:17:35,069 --> 00:17:37,549 must be identical to feature F in the reduplicant. 392 00:17:37,720 --> 00:17:40,579 And this dual system, IOFaith and BRFaith, helps 393 00:17:40,579 --> 00:17:43,059 explain that emergence of the unmarked thing 394 00:17:43,059 --> 00:17:46,279 we saw earlier, but now in reduplicants. Why, 395 00:17:46,359 --> 00:17:48,500 copied bits are sometimes simpler than the original. 396 00:17:48,740 --> 00:17:51,400 Yes, it provides a really neat account. Imaginal 397 00:17:51,400 --> 00:17:53,799 language allows complex codas in general, but 398 00:17:53,799 --> 00:17:55,700 in the reduplicant part, those codas simplify. 399 00:17:56,200 --> 00:17:58,500 The ranking explanation is often this three -way 400 00:17:58,500 --> 00:18:01,839 interaction. First, text IOFaithfulness, like 401 00:18:01,839 --> 00:18:04,640 text IOFaithfulness for consonants, is ranked 402 00:18:04,640 --> 00:18:06,900 high, ensuring the base cues its complex coda. 403 00:18:07,220 --> 00:18:10,279 preserving the input information. Second, this 404 00:18:10,279 --> 00:18:12,799 high -ranked text IO faithfulness dominates the 405 00:18:12,799 --> 00:18:14,880 markedness constraint against complex coda, say, 406 00:18:15,019 --> 00:18:18,079 text complex code. This allows the base to surface 407 00:18:18,079 --> 00:18:21,319 with the marked structure. But third, text complex 408 00:18:21,319 --> 00:18:23,740 IO gate is ranked above the relevant text DR 409 00:18:23,740 --> 00:18:26,420 identity constraint. So the reduplicant is caught 410 00:18:26,420 --> 00:18:29,119 in the middle. Right. To satisfy the higher -ranked 411 00:18:29,119 --> 00:18:31,960 text complex coda, the reduplicant violates identity 412 00:18:31,960 --> 00:18:35,410 with the base and simplifies its coda. The unmarked 413 00:18:35,410 --> 00:18:37,470 structure emerges in the copy because faithfulness 414 00:18:37,470 --> 00:18:39,950 to the base is sacrificed for structural well 415 00:18:39,950 --> 00:18:42,369 -formedness. That's clever. What about those 416 00:18:42,369 --> 00:18:45,049 weird cases of over -application and under -application 417 00:18:45,049 --> 00:18:48,450 mentioned in 5 .6? What's over -application? 418 00:18:48,609 --> 00:18:51,470 Over -application is when a phonological process 419 00:18:51,470 --> 00:18:54,450 applies in the rediblicate in a place where the 420 00:18:54,450 --> 00:18:56,509 conditioning environment for that process isn't 421 00:18:56,509 --> 00:18:59,200 actually met. The example given is often Malay 422 00:18:59,200 --> 00:19:01,339 nasal harmony. How does that work? Let's say 423 00:19:01,339 --> 00:19:04,059 nasal harmony spreads nasality onto vowels following 424 00:19:04,059 --> 00:19:07,799 a nasal consonant in the base. In over -application, 425 00:19:07,880 --> 00:19:10,160 the vowel in the reduplicant might also become 426 00:19:10,160 --> 00:19:13,339 nasalized, just to match the base vowel, even 427 00:19:13,339 --> 00:19:14,920 if the reduplicant doesn't have the triggering 428 00:19:14,920 --> 00:19:17,259 nasal consonant nearby. Why would it do that? 429 00:19:17,359 --> 00:19:19,559 It seems illogical. Because the text -ident BR 430 00:19:19,559 --> 00:19:22,500 nasal constraint, demanding identity for nasality 431 00:19:22,500 --> 00:19:24,619 between base and reduplicant, is ranked very 432 00:19:24,619 --> 00:19:27,700 high. It's more important for the copy to sound 433 00:19:27,700 --> 00:19:29,859 identical to the original in terms of nasality 434 00:19:29,859 --> 00:19:32,980 than it is to obey the normal local rules for 435 00:19:32,980 --> 00:19:35,839 when nasalization should happen. Identity trumps 436 00:19:35,839 --> 00:19:37,940 phonotactics. Okay, that makes sense under OT 437 00:19:37,940 --> 00:19:40,819 logic. Then chapter 6 extends this identity idea 438 00:19:40,819 --> 00:19:43,740 even further with output -to -output correspondence. 439 00:19:43,940 --> 00:19:46,740 What's that alluding? Instead of base -to -reduplicant 440 00:19:46,740 --> 00:19:50,500 within one word, OO correspondence relates two 441 00:19:50,500 --> 00:19:53,440 separate, fully -formed output words that have 442 00:19:53,440 --> 00:19:56,220 a morphological relationship. like a full word 443 00:19:56,220 --> 00:19:59,519 and its shortened truncated version, or a stem 444 00:19:59,519 --> 00:20:01,880 and a word derived from that stem. So we have 445 00:20:01,880 --> 00:20:05,000 constraints like BT identity base truncated. 446 00:20:05,180 --> 00:20:07,980 Exactly. These demand that the truncated form 447 00:20:07,980 --> 00:20:10,440 maintain faithfulness to the full base form it's 448 00:20:10,440 --> 00:20:13,039 derived from. This can lead to underapplication, 449 00:20:13,059 --> 00:20:15,559 where a process that should apply in the truncated 450 00:20:15,559 --> 00:20:19,230 form fails to apply. precisely because applying 451 00:20:19,230 --> 00:20:21,690 it would make the truncated form less similar 452 00:20:21,690 --> 00:20:24,069 to its base. Or like keeping a vowel sound from 453 00:20:24,069 --> 00:20:26,029 the base, even if the truncation changes the 454 00:20:26,029 --> 00:20:28,650 environment. Could be. Or it can lead to over 455 00:20:28,650 --> 00:20:31,009 -application in derived forms, where some phonetic 456 00:20:31,009 --> 00:20:34,130 detail, like allophonic variation, from the stem 457 00:20:34,130 --> 00:20:37,130 carries over into an affixed form, even if the 458 00:20:37,130 --> 00:20:39,450 conditioning environment is gone, purely to maintain 459 00:20:39,450 --> 00:20:42,230 identity between the stem output and the derived 460 00:20:42,230 --> 00:20:44,210 output. And the big advantage here compared to 461 00:20:44,210 --> 00:20:46,349 older theories. It captures effects that used 462 00:20:46,349 --> 00:20:49,740 to be handled by cyclic rule application. The 463 00:20:49,740 --> 00:20:51,759 idea that rules applied first to inner structures, 464 00:20:51,920 --> 00:20:55,079 then reapplied to outer structures. OO correspondence 465 00:20:55,079 --> 00:20:57,819 does this without needing those abstract intermediate 466 00:20:57,819 --> 00:21:01,039 stages or the machinery of the cycle. It keeps 467 00:21:01,039 --> 00:21:03,380 the analysis closer to the surface forms we actually 468 00:21:03,380 --> 00:21:06,279 observe. Okay, we've covered the core phonology 469 00:21:06,279 --> 00:21:09,019 and morphology interface. Now, part four feels 470 00:21:09,019 --> 00:21:11,869 like zooming out. Chapter 7 and 8, looking at 471 00:21:11,869 --> 00:21:14,710 learning and extending to syntax. Let's start 472 00:21:14,710 --> 00:21:17,690 with learning, Chapter 7. If rankings define 473 00:21:17,690 --> 00:21:20,369 languages, how does a kid figure out the ranking, 474 00:21:20,589 --> 00:21:22,430 especially since they only hear the winners, 475 00:21:22,670 --> 00:21:25,009 the actual outputs? That's the huge acquisition 476 00:21:25,009 --> 00:21:27,190 challenge. The child gets positive evidence, 477 00:21:27,390 --> 00:21:30,309 only they hear bet in Dutch, but never the candidate 478 00:21:30,309 --> 00:21:33,009 bed labeled as wrong. How do they infer the hierarchy? 479 00:21:33,589 --> 00:21:36,269 The main proposal from Tease and Smolenski involves 480 00:21:36,269 --> 00:21:39,130 an algorithm based on constraint demotion. Constraint 481 00:21:39,130 --> 00:21:41,450 demotion. How does that work? The learner basically 482 00:21:41,450 --> 00:21:43,869 uses the winner to figure out why the loser's 483 00:21:43,869 --> 00:21:46,670 lost. When they hear an optimal output, say, 484 00:21:46,809 --> 00:21:50,589 old dollar. for some input answer, they implicitly 485 00:21:50,589 --> 00:21:52,509 compare it to other potential candidates, like 486 00:21:52,509 --> 00:21:56,230 suboptimal dollars. If dollar beat dollars, there 487 00:21:56,230 --> 00:21:58,410 must be some constraint that dollar satisfies 488 00:21:58,410 --> 00:22:01,430 but dollar violates. And this constraint must 489 00:22:01,430 --> 00:22:03,390 outrank any constraint that dollar satisfies 490 00:22:03,390 --> 00:22:06,230 but a dollar violates. OK, so the comparison 491 00:22:06,230 --> 00:22:09,589 reveals crucial ranking information. Yes. Each 492 00:22:09,589 --> 00:22:12,109 such comparison provides evidence to demote constraints 493 00:22:12,109 --> 00:22:14,809 violated by the winner below, constraints violated 494 00:22:14,809 --> 00:22:17,420 only by the loser. The algorithm starts with 495 00:22:17,420 --> 00:22:19,880 all constraints unranked and gradually stratifies 496 00:22:19,880 --> 00:22:21,740 them based on these comparisons derived from 497 00:22:21,740 --> 00:22:23,980 the input data until the correct hierarchy emerges. 498 00:22:24,339 --> 00:22:26,440 But what about when words change form, like the 499 00:22:26,440 --> 00:22:28,660 Dutch bed becoming bet sometimes and be do on 500 00:22:28,660 --> 00:22:30,740 other times? This is the alternation problem. 501 00:22:31,079 --> 00:22:33,500 How does the learner figure out the single underlying 502 00:22:33,500 --> 00:22:36,579 representation, bed with a D, when the surface 503 00:22:36,579 --> 00:22:38,720 forms vary? This requires another principle, 504 00:22:38,940 --> 00:22:42,119 lexicon optimization. The learner, in essence, 505 00:22:42,339 --> 00:22:45,299 tries out different possible URs. The goal is 506 00:22:45,299 --> 00:22:47,900 to choose the UR that allows the observed surface 507 00:22:47,900 --> 00:22:50,200 forms to be generated with the minimal overall 508 00:22:50,200 --> 00:22:52,720 violation cost, given the constraint hierarchy 509 00:22:52,720 --> 00:22:55,140 they are learning. So if the child guessed the 510 00:22:55,140 --> 00:22:58,220 UR for Dutch was bed, with a T, they'd have trouble 511 00:22:58,220 --> 00:23:01,380 explaining the plural B, de en. Big trouble. 512 00:23:01,700 --> 00:23:05,460 They'd need some unusual voicing rule. Hypothesizing 513 00:23:05,460 --> 00:23:08,609 the UR as bed, with a D, makes the singular bed 514 00:23:08,609 --> 00:23:11,569 easy to explain due to high -ranked Pentext -voiced 515 00:23:11,569 --> 00:23:14,589 coda. And the plural bait, also straightforward, 516 00:23:14,910 --> 00:23:17,250 voicing is preserved when not in the coda. That 517 00:23:17,250 --> 00:23:20,029 UR, bed, leads to a simpler overall analysis 518 00:23:20,029 --> 00:23:22,230 across the paradigm. So lexicon optimization 519 00:23:22,230 --> 00:23:24,470 pushes the learner towards the UR that requires 520 00:23:24,470 --> 00:23:27,089 the least magic or the fewest faithfulness violations 521 00:23:27,089 --> 00:23:29,730 overall to generate all the related surface forms. 522 00:23:30,009 --> 00:23:33,009 Essentially, yes. The system favors URs that 523 00:23:33,009 --> 00:23:34,910 are maximally faithful to the range of outputs. 524 00:23:35,349 --> 00:23:37,009 guided by the developing constraint ranking. 525 00:23:37,349 --> 00:23:40,130 There are safeguards, like the anti -loop provision 526 00:23:40,130 --> 00:23:42,829 mentioned in Kager, to prevent the learner from 527 00:23:42,829 --> 00:23:45,690 getting stuck on incorrect URs when faced with 528 00:23:45,690 --> 00:23:48,410 conflicting evidence. It encourages finding the 529 00:23:48,410 --> 00:23:51,109 most economical underlying form. Fascinating. 530 00:23:51,450 --> 00:23:53,869 Okay, final stretch, chapter eight, extending 531 00:23:53,869 --> 00:23:57,089 OT to syntax. Can this constraint ranking system 532 00:23:57,089 --> 00:23:59,789 really handle sentence structure, word order, 533 00:23:59,890 --> 00:24:02,849 movement? That's the claim, and yes, it applies 534 00:24:02,849 --> 00:24:06,130 the same core logic. G -E -N creates candidate 535 00:24:06,130 --> 00:24:08,630 syntactic structures, entire sentences or phrases. 536 00:24:09,150 --> 00:24:11,349 Eval uses ranked constraints to pick the best 537 00:24:11,349 --> 00:24:13,130 one. The constraints are just different. They 538 00:24:13,130 --> 00:24:15,609 refer to syntactic well -formedness. What kinds 539 00:24:15,609 --> 00:24:17,890 of constraints operate in syntax? You have constraints 540 00:24:17,890 --> 00:24:20,089 demanding certain configurations, like text OB 541 00:24:20,089 --> 00:24:22,650 spec, which requires operators, like who or what, 542 00:24:22,809 --> 00:24:25,250 to be in a specific structural position, the 543 00:24:25,250 --> 00:24:29,230 specifier CP, or text OB HD, obligatory head, 544 00:24:29,410 --> 00:24:31,839 which says projections need heads. These function 545 00:24:31,839 --> 00:24:34,160 like markedness constraints. And the faithfulness 546 00:24:34,160 --> 00:24:37,380 side in syntax. The big one is text stay, which 547 00:24:37,380 --> 00:24:41,019 is basically don't move. It penalizes any movement 548 00:24:41,019 --> 00:24:43,339 of elements from their base position. Think of 549 00:24:43,339 --> 00:24:45,940 it as syntactic inertia. So syntactic movement, 550 00:24:46,119 --> 00:24:48,220 like asking a question in English, what did she 551 00:24:48,220 --> 00:24:51,480 say, happens when the need to satisfy a structural 552 00:24:51,480 --> 00:24:54,140 constraint outweighs the cost of movement. Exactly. 553 00:24:54,220 --> 00:24:57,579 Let's trace what did Mary say. The underlined 554 00:24:57,579 --> 00:25:00,180 structure might have what inside the verb phrase, 555 00:25:00,420 --> 00:25:03,799 but TextActXFS is high -ranked in English, demanding 556 00:25:03,799 --> 00:25:07,059 what move to the front to specifier of CP. Okay, 557 00:25:07,099 --> 00:25:10,220 so what moves violating text stay. Right, but 558 00:25:10,220 --> 00:25:13,700 that movement leaves the head of CP empty. Now, 559 00:25:13,759 --> 00:25:18,059 TextOBHD is violated. Since TextOBHD is also 560 00:25:18,059 --> 00:25:20,079 high -ranked in English and higher than text 561 00:25:20,079 --> 00:25:22,619 stay, the auxiliary deed has to move into that 562 00:25:22,619 --> 00:25:24,869 head position to satisfy it. And that's the inversion 563 00:25:24,869 --> 00:25:27,210 we see, but moving did violates text decade. 564 00:25:27,349 --> 00:25:29,329 Correct. The winning candidate violates text 565 00:25:29,329 --> 00:25:31,890 twice, but crucially it satisfies the higher 566 00:25:31,890 --> 00:25:34,869 rank structural demands of text OB spec and text 567 00:25:34,869 --> 00:25:38,329 OB HD. Any candidate that didn't move what or 568 00:25:38,329 --> 00:25:41,269 did would fatally violate those top constraints. 569 00:25:41,569 --> 00:25:43,910 So the English ranking is something like text 570 00:25:43,910 --> 00:25:48,140 OB spec, text OB HD. That's the idea. And this 571 00:25:48,140 --> 00:25:50,880 naturally explains syntactic economy. You only 572 00:25:50,880 --> 00:25:53,700 move things as much as absolutely necessary to 573 00:25:53,700 --> 00:25:56,900 satisfy higher -ranked constraints. Any extra 574 00:25:56,900 --> 00:25:59,559 unnecessary movement would just incur extra text 575 00:25:59,559 --> 00:26:02,380 violations for no gain, so that candidate would 576 00:26:02,380 --> 00:26:05,200 lose. And changing the ranking explains different 577 00:26:05,200 --> 00:26:07,960 language types, like languages without Y movement. 578 00:26:08,460 --> 00:26:10,779 Precisely. If a language ranked higher than text 579 00:26:10,779 --> 00:26:12,720 type X, then moving the wayward would be too 580 00:26:12,720 --> 00:26:14,920 costly. The optimal candidate would leave it 581 00:26:14,920 --> 00:26:17,650 in place, Y in situ. as you see in languages 582 00:26:17,650 --> 00:26:20,589 like Chinese or Japanese. The factorial typology, 583 00:26:20,730 --> 00:26:22,950 all possible rankings, predicts the range of 584 00:26:22,950 --> 00:26:25,890 observed syntactic variation. Wow. Okay, let's 585 00:26:25,890 --> 00:26:27,670 try to wrap this all up. This deep dive really 586 00:26:27,670 --> 00:26:30,670 shows OT as a fundamental shift. Definitely. 587 00:26:30,769 --> 00:26:32,650 It threw out the step -by -step derivations and 588 00:26:32,650 --> 00:26:35,009 replaced them with this parallel evaluation system 589 00:26:35,009 --> 00:26:38,589 driven by universal but viable and ranked constraints. 590 00:26:38,970 --> 00:26:41,750 The core tension is always markedness versus 591 00:26:41,750 --> 00:26:44,029 faithfulness structure versus memory. And the 592 00:26:44,029 --> 00:26:46,579 key insight is that All the differences between 593 00:26:46,579 --> 00:26:49,119 languages boil down to just shuffling the priority 594 00:26:49,119 --> 00:26:51,539 list, the ranking of these universal constraints. 595 00:26:51,859 --> 00:26:54,799 Strict domination makes that ranking absolute. 596 00:26:55,119 --> 00:26:57,420 And we saw how this accounts for so much phonological 597 00:26:57,420 --> 00:27:00,740 patterns, conspiracies, stress, reduplication 598 00:27:00,740 --> 00:27:04,339 complexities via BR and OO correspondence, even 599 00:27:04,339 --> 00:27:07,160 syntactic movement. And it offers a formal model 600 00:27:07,160 --> 00:27:10,059 for how kids could learn these rankings via constraint 601 00:27:10,059 --> 00:27:13,309 demotion and lexicon optimization. But as the 602 00:27:13,309 --> 00:27:15,369 sources acknowledge, especially towards the end, 603 00:27:15,509 --> 00:27:18,670 it's not perfect. The standard model, this direct 604 00:27:18,670 --> 00:27:21,289 input output mapping, runs into trouble with 605 00:27:21,289 --> 00:27:23,730 certain phenomena, particularly something called 606 00:27:23,730 --> 00:27:26,210 opacity. Right. Opacity is tricky. It's when 607 00:27:26,210 --> 00:27:27,809 the surface form looks like it must have gone 608 00:27:27,809 --> 00:27:30,490 through some intermediate stage that isn't transparent 609 00:27:30,490 --> 00:27:33,369 anymore. Like a rule seems to have applied, but 610 00:27:33,369 --> 00:27:35,250 its effects are somehow masked or unexpected 611 00:27:35,250 --> 00:27:38,269 in the final output, making the winner look suboptimal 612 00:27:38,269 --> 00:27:40,089 according to the straightforward ranking. So 613 00:27:40,089 --> 00:27:42,190 the elegant direct mapping breaks down a bit. 614 00:27:42,430 --> 00:27:44,809 It seems to in these cases. Standard OT struggles 615 00:27:44,809 --> 00:27:46,589 because it doesn't have intermediate stages. 616 00:27:47,109 --> 00:27:49,509 This has pushed the theory to evolve. Leading 617 00:27:49,509 --> 00:27:51,690 to more advanced, maybe more complex versions 618 00:27:51,690 --> 00:27:54,769 of OT. Exactly. Things like sympathy theory, 619 00:27:54,910 --> 00:27:57,289 where the winner has to bear some resemblance 620 00:27:57,289 --> 00:28:00,049 to another sympathetic candidate, even if that 621 00:28:00,049 --> 00:28:03,470 candidate isn't the true optimal one. Or... Local 622 00:28:03,470 --> 00:28:06,109 conjunction, where violating two specific constraints 623 00:28:06,109 --> 00:28:08,789 at the same time is much worse than violating 624 00:28:08,789 --> 00:28:11,809 them individually. These are ways to handle opacity 625 00:28:11,809 --> 00:28:15,089 without bringing back full serial derivations. 626 00:28:15,640 --> 00:28:18,539 So even this revolutionary theory is constantly 627 00:28:18,539 --> 00:28:21,140 being tested and refined by the sheer complexity 628 00:28:21,140 --> 00:28:23,839 of language. Always. That's linguistics. It's 629 00:28:23,839 --> 00:28:26,240 incredible to think about. So for you listening, 630 00:28:26,359 --> 00:28:28,740 the next time you speak or hear language, remember 631 00:28:28,740 --> 00:28:31,859 this hidden battle. Countless universal constraints 632 00:28:31,859 --> 00:28:34,480 pushing for simplicity, pulling for faithfulness, 633 00:28:34,680 --> 00:28:37,680 all evaluated in parallel, with the winner being 634 00:28:37,680 --> 00:28:40,299 the single utterance that emerges, having minimally 635 00:28:40,299 --> 00:28:42,819 violated the specific priority list that defines 636 00:28:42,819 --> 00:28:46,000 your language. A constant, dynamic conflict shaping 637 00:28:46,000 --> 00:28:48,640 every sound, word, and sentence. It's quite something.