1 00:00:00,000 --> 00:00:02,640 Welcome back to the Deep Dive. Today we are strapping 2 00:00:02,640 --> 00:00:04,839 in for a fascinating journey into one of the 3 00:00:04,839 --> 00:00:07,919 most hidden mechanisms of your own brain, the 4 00:00:07,919 --> 00:00:10,019 sound of silence. It really does sound like a 5 00:00:10,019 --> 00:00:13,179 paradox, doesn't it? But it's deeply real. We're 6 00:00:13,179 --> 00:00:15,539 talking about the unseen rhythmic, the phrasing, 7 00:00:15,560 --> 00:00:19,359 and the international structure that you automatically 8 00:00:19,359 --> 00:00:21,859 assign to sentences the moment you read them 9 00:00:21,859 --> 00:00:23,780 silently. And for this, we're really digging 10 00:00:23,780 --> 00:00:26,219 into one source. We are. We're synthesizing the 11 00:00:26,219 --> 00:00:28,280 monumental work from the 2022 investigation, 12 00:00:28,739 --> 00:00:31,879 The Sound of Silence, Investigations of Implicit 13 00:00:31,879 --> 00:00:34,460 Prosody. That's right. So our mission today is 14 00:00:34,460 --> 00:00:37,119 to explore how this silent sound, this implicit 15 00:00:37,119 --> 00:00:40,240 prosody, is constructed word by word in real 16 00:00:40,240 --> 00:00:43,420 time. And crucially, how it acts as, well, maybe 17 00:00:43,420 --> 00:00:46,719 the primary driver for resolving ambiguity in 18 00:00:46,719 --> 00:00:49,000 how we parse sentences. It turns out that the 19 00:00:49,000 --> 00:00:51,420 phantom voice in your head is less a quiet narrator 20 00:00:51,420 --> 00:00:55,200 and more a very strict foreman deciding how phrases 21 00:00:55,200 --> 00:00:57,640 get built. That's a great way to put it. So let's 22 00:00:57,640 --> 00:01:00,210 start at the beginning. To understand the silent 23 00:01:00,210 --> 00:01:02,409 voice, I guess we first have to understand the 24 00:01:02,409 --> 00:01:05,730 spoken one. Let's establish the ground truth 25 00:01:05,730 --> 00:01:09,049 of prosody in overt spoken language. Right. When 26 00:01:09,049 --> 00:01:10,769 we talk about prosody, we're basically talking 27 00:01:10,769 --> 00:01:13,049 about the music of speech, correct? Precisely, 28 00:01:13,049 --> 00:01:15,849 yeah. Music is the perfect word. Prosody covers, 29 00:01:16,129 --> 00:01:18,849 you could say, three main elements. First, you 30 00:01:18,849 --> 00:01:21,510 have the rhythm, the pattern of stressed and 31 00:01:21,510 --> 00:01:24,930 unstressed syllables. Like convict versus convicity. 32 00:01:25,230 --> 00:01:27,849 Exactly. Right. Then there's phrasing, how we 33 00:01:27,849 --> 00:01:30,790 break up the speech stream into chunks, or what 34 00:01:30,790 --> 00:01:34,489 linguists call prosodic constituents. The pauses, 35 00:01:34,489 --> 00:01:36,950 the grouping of words. And finally, the international 36 00:01:36,950 --> 00:01:39,969 contour. That's the rise and fall of pitch, which 37 00:01:39,969 --> 00:01:42,670 conveys all sorts of things like focus, emotion, 38 00:01:42,989 --> 00:01:45,250 whether you're asking a question. And the key 39 00:01:45,250 --> 00:01:47,989 thing here, which the research makes very clear, 40 00:01:48,109 --> 00:01:50,590 is that these elements are not just decorative. 41 00:01:50,750 --> 00:01:53,090 They're not just flavor. Not at all. They're 42 00:01:53,090 --> 00:01:55,769 completely integral to comprehension. So the 43 00:01:55,769 --> 00:01:57,790 source highlights that prosody influences how 44 00:01:57,790 --> 00:02:00,640 we understand things in two main ways. Boundaries 45 00:02:00,640 --> 00:02:03,480 and accents. Let's tackle boundaries first. If 46 00:02:03,480 --> 00:02:06,640 I'm listening to you speak and you, you pause 47 00:02:06,640 --> 00:02:10,219 for a second. How does that pause, that boundary 48 00:02:10,219 --> 00:02:12,319 change how I structure the sentence in my head? 49 00:02:12,479 --> 00:02:15,030 It basically builds a wall. A structural wall. 50 00:02:15,210 --> 00:02:18,030 I mean, decades of research show that when listeners 51 00:02:18,030 --> 00:02:21,889 hear a major prosodic boundary like a longer 52 00:02:21,889 --> 00:02:25,370 pause or a real distinct change in pitch, they're 53 00:02:25,370 --> 00:02:28,110 consistently less likely to attach any new information 54 00:02:28,110 --> 00:02:30,830 to the word that came right before that boundary. 55 00:02:31,009 --> 00:02:33,069 OK, so the purser in my brain just assumes that 56 00:02:33,069 --> 00:02:36,090 phrase is done. It's closed off. It is. The boundary, 57 00:02:36,129 --> 00:02:39,310 in effect, repels new information. It says, don't 58 00:02:39,310 --> 00:02:41,650 attach here. This part is finished. So if I'm 59 00:02:41,650 --> 00:02:44,069 dealing with an ambiguous sentence. Something 60 00:02:44,069 --> 00:02:46,849 like, the audience enjoyed the performance of 61 00:02:46,849 --> 00:02:48,909 the singer, pause, who was touring Europe. Right. 62 00:02:48,949 --> 00:02:51,569 That pause makes me strongly prefer that who 63 00:02:51,569 --> 00:02:53,569 was touring Europe modifies the performance, 64 00:02:53,689 --> 00:02:57,210 not the singer. Exactly. The boundary isolates 65 00:02:57,210 --> 00:02:59,210 the singer, so the attachment has to go up higher 66 00:02:59,210 --> 00:03:01,889 in the structure to performance. We call that 67 00:03:01,889 --> 00:03:03,590 high attachment. Okay, so that's the pushing 68 00:03:03,590 --> 00:03:06,409 force. The boundary pushes away. But now let's 69 00:03:06,409 --> 00:03:08,550 talk about the pulling force. Pitch accents. 70 00:03:09,080 --> 00:03:11,259 The accents, right? These are all about information 71 00:03:11,259 --> 00:03:14,020 structure, highlighting what's new or important, 72 00:03:14,240 --> 00:03:18,199 what we call focus. And this is where the focus 73 00:03:18,199 --> 00:03:21,360 attraction hypothesis, or FAH, comes in. Focus 74 00:03:21,360 --> 00:03:23,800 attraction. So that sounds pretty self -explanatory. 75 00:03:24,039 --> 00:03:27,979 It is. The FAH argues that focus... Well, it 76 00:03:27,979 --> 00:03:30,039 attracts attachment. If a sentence has two possible 77 00:03:30,039 --> 00:03:33,219 nouns, a phrase could modify, and you put a prominent 78 00:03:33,219 --> 00:03:35,460 pitch movement on one of them, the parser gets 79 00:03:35,460 --> 00:03:37,759 pulled toward that accented noun. So in that 80 00:03:37,759 --> 00:03:40,039 same sentence, if you said the audience enjoyed 81 00:03:40,039 --> 00:03:42,900 the performance of the singer with that emphasis. 82 00:03:43,219 --> 00:03:45,669 Yep. then the relative clause, who is touring 83 00:03:45,669 --> 00:03:48,110 Europe, is almost certainly going to be interpreted 84 00:03:48,110 --> 00:03:50,550 as modifying the accented singer. We call that 85 00:03:50,550 --> 00:03:53,610 low attachment. Focus is a very strong attractor. 86 00:03:53,629 --> 00:03:56,110 Okay, so this is fascinating. In spoken language, 87 00:03:56,289 --> 00:03:58,569 we have these two powerful forces that can sometimes 88 00:03:58,569 --> 00:04:01,030 compete. We have boundaries that push material 89 00:04:01,030 --> 00:04:03,629 away repulsion. Right. And we have accents that 90 00:04:03,629 --> 00:04:05,770 pull material toward the fraction. So what happens 91 00:04:05,770 --> 00:04:07,949 when those two forces clash in the same sentence? 92 00:04:08,129 --> 00:04:11,599 Which one wins? And that. That is the exact question 93 00:04:11,599 --> 00:04:13,280 where the source material builds its foundation. 94 00:04:13,599 --> 00:04:16,600 It challenges some older assumptions and establishes 95 00:04:16,600 --> 00:04:18,199 what's called the visibility first hypothesis 96 00:04:18,199 --> 00:04:22,420 or VFH. Visibility first. OK, that name suggests 97 00:04:22,420 --> 00:04:24,800 there's a hierarchy. One of these is dominant. 98 00:04:24,920 --> 00:04:28,379 It is. The VFH argues that attachment at its 99 00:04:28,379 --> 00:04:33,339 core prioritizes the most visible site. And visibility 100 00:04:33,339 --> 00:04:35,220 here is a technical term. What does it mean? 101 00:04:35,420 --> 00:04:37,699 It's defined by the absence of a major prosodic 102 00:04:37,699 --> 00:04:41,639 boundary. So crucially. The VFH sets up a clear 103 00:04:41,639 --> 00:04:44,120 pecking order. Prosodic boundaries are the top 104 00:04:44,120 --> 00:04:46,500 priority. They are the boss. So if a boundary 105 00:04:46,500 --> 00:04:49,480 separates a potential attachment site from the 106 00:04:49,480 --> 00:04:51,279 incoming words... Then that site is basically 107 00:04:51,279 --> 00:04:54,040 rendered invisible. It's structurally inaccessible. 108 00:04:54,160 --> 00:04:56,980 So the phrasing, the boundary, actually overrides 109 00:04:56,980 --> 00:05:00,100 the focus from the accent. Absolutely. The VFH 110 00:05:00,100 --> 00:05:02,600 posits that pitch accent of focus only get a 111 00:05:02,600 --> 00:05:04,670 say in the matter. They only start to modulate 112 00:05:04,670 --> 00:05:06,649 our preferences when both potential attachment 113 00:05:06,649 --> 00:05:08,970 sites are equally visible. Meaning when they're 114 00:05:08,970 --> 00:05:12,250 both inside the same uninterrupted phrase? Exactly. 115 00:05:12,269 --> 00:05:14,069 If there's a strong boundary in the way, the 116 00:05:14,069 --> 00:05:16,170 parser doesn't care how brightly you accent a 117 00:05:16,170 --> 00:05:18,649 word. It just defaults to the site that's on 118 00:05:18,649 --> 00:05:21,629 the near side of that boundary wall. It's a fundamental 119 00:05:21,629 --> 00:05:24,709 reordering of how we thought these prosodic cues 120 00:05:24,709 --> 00:05:27,000 work together. Okay, this is a perfect foundation. 121 00:05:27,240 --> 00:05:29,620 We have the rules of engagement for spoken language. 122 00:05:29,879 --> 00:05:32,399 Now we have to transpose all of this onto silent 123 00:05:32,399 --> 00:05:34,379 reading. Right. And this is where we get to the 124 00:05:34,379 --> 00:05:38,759 core concept. Implicit praesity. The mental blueprint 125 00:05:38,759 --> 00:05:42,000 of rhythm, phrasing, and contour that the reader 126 00:05:42,000 --> 00:05:46,029 is just generating internally. And this internal 127 00:05:46,029 --> 00:05:48,610 blueprint, this silent music, is what drives 128 00:05:48,610 --> 00:05:50,810 the whole theory. The implicit prosody hypothesis, 129 00:05:51,189 --> 00:05:54,509 or IPH. What's the core principle of the IPH? 130 00:05:54,829 --> 00:05:56,769 It's pretty straightforward, actually. The IPH 131 00:05:56,769 --> 00:05:58,649 just says that when a reader encounters a sentence 132 00:05:58,649 --> 00:06:00,310 that could be structured in two different ways, 133 00:06:00,430 --> 00:06:04,569 a syntactic ambiguity, the parser favors the 134 00:06:04,569 --> 00:06:06,850 analysis that aligns with the most natural, the 135 00:06:06,850 --> 00:06:09,050 most default, the most expected prosodic contour 136 00:06:09,050 --> 00:06:11,350 for that string of words. So the inner voice 137 00:06:11,350 --> 00:06:13,850 isn't just, it's not a passive narrator just 138 00:06:13,850 --> 00:06:15,589 reading the words back to you. It's actively 139 00:06:15,589 --> 00:06:18,490 structuring the text as you go. It's structuring 140 00:06:18,490 --> 00:06:20,870 it. And that structure determines the meaning 141 00:06:20,870 --> 00:06:23,629 before the syntax is even fully resolved. Wow. 142 00:06:23,709 --> 00:06:26,370 And we know that this structure, it has a cognitive 143 00:06:26,370 --> 00:06:28,769 cost, especially when it needs to be undone, 144 00:06:28,769 --> 00:06:31,610 right? A huge cost. And this is captured in what's 145 00:06:31,610 --> 00:06:33,990 called the prosodic constraint on reanalysis, 146 00:06:33,990 --> 00:06:38,370 or PCR. Let's unpack the PCR a bit. How much 147 00:06:38,370 --> 00:06:41,009 harder is it to read when you have to perform 148 00:06:41,009 --> 00:06:44,449 both a syntactic and a prosodic reanalysis? It's 149 00:06:44,449 --> 00:06:47,009 significantly harder. I mean, think about a classic... 150 00:06:47,160 --> 00:06:49,560 garden path sentence. The kind that leads you 151 00:06:49,560 --> 00:06:52,279 down one path and then a later word reveals you 152 00:06:52,279 --> 00:06:55,680 took a wrong turn. Exactly. That process of revising 153 00:06:55,680 --> 00:06:58,259 your interpretation is already costly. But the 154 00:06:58,259 --> 00:07:01,339 PCR says that cost is compounded if you also 155 00:07:01,339 --> 00:07:03,800 have to dismantle and rebuild the rhythmic structure 156 00:07:03,800 --> 00:07:05,839 your inner voice is already assigned to the first 157 00:07:05,839 --> 00:07:07,459 part of the sentence. So if you only have to 158 00:07:07,459 --> 00:07:10,040 change the grammar, that's one level of difficulty. 159 00:07:10,439 --> 00:07:13,350 It's manageable. But if you have to tear down 160 00:07:13,350 --> 00:07:15,670 the mental phrasing and the intonation pattern 161 00:07:15,670 --> 00:07:18,269 you've already built, the process becomes much 162 00:07:18,269 --> 00:07:21,089 slower, much more error prone. And the dissertation 163 00:07:21,089 --> 00:07:23,850 points to some German self -paced reading studies 164 00:07:23,850 --> 00:07:26,569 to show this. It does. They had these ambiguous 165 00:07:26,569 --> 00:07:29,230 relative clauses. And when the sentence resolved 166 00:07:29,230 --> 00:07:31,629 in a way that required a complete shift in the 167 00:07:31,629 --> 00:07:34,689 expected intonational phrasing, the reading slowdowns 168 00:07:34,689 --> 00:07:38,000 were just much more pronounced. So the brain 169 00:07:38,000 --> 00:07:40,720 is really investing effort in building this. 170 00:07:41,089 --> 00:07:44,189 prosodic show. And it hates wasting that investment. 171 00:07:44,410 --> 00:07:46,850 The big takeaway for the field, really, is that 172 00:07:46,850 --> 00:07:49,209 we cannot treat silent reading as a purely visual, 173 00:07:49,529 --> 00:07:53,089 a purely syntactic process. The prosodic that 174 00:07:53,089 --> 00:07:56,310 we generate is this mandatory, instantaneous 175 00:07:56,310 --> 00:07:59,449 layer of interpretation that actively governs 176 00:07:59,449 --> 00:08:01,699 how we resolve conflict. This all sounds incredibly 177 00:08:01,699 --> 00:08:04,420 neat in theory, but now we have to face the central 178 00:08:04,420 --> 00:08:06,800 challenge, right? The idea of a single default 179 00:08:06,800 --> 00:08:09,379 parsity. Ah, yes. I mean, if I read a sentence 180 00:08:09,379 --> 00:08:11,120 at a normal pace, there are often a few different 181 00:08:11,120 --> 00:08:12,860 ways I could phrase it that all sound perfectly 182 00:08:12,860 --> 00:08:15,699 fine. They're not weird or marked. So how can 183 00:08:15,699 --> 00:08:17,819 the parser rely on a default if several options 184 00:08:17,819 --> 00:08:21,120 seem possible? And you just hit on the crux of 185 00:08:21,120 --> 00:08:23,660 the incrementality problem. This is the core 186 00:08:23,660 --> 00:08:26,819 issue. The prosody that you assign during what 187 00:08:26,819 --> 00:08:29,759 we call first pass parsing, that's the rapid. 188 00:08:30,220 --> 00:08:32,879 word -by -word decision -making process, has 189 00:08:32,879 --> 00:08:35,399 to operate with incomplete information. It doesn't 190 00:08:35,399 --> 00:08:38,019 know how the sentence ends yet. Exactly. So that 191 00:08:38,019 --> 00:08:40,620 initial incremental prosody you assign might 192 00:08:40,620 --> 00:08:43,080 be fundamentally different from the final default 193 00:08:43,080 --> 00:08:45,519 prosody that a speaker would assign once they 194 00:08:45,519 --> 00:08:48,159 can see the whole finished sentence. The parser 195 00:08:48,159 --> 00:08:50,360 has to make a choice now, not later. So what 196 00:08:50,360 --> 00:08:52,559 happens if the prosodic structure starts building 197 00:08:52,559 --> 00:08:55,399 in the first few words? What if that contradicts 198 00:08:55,399 --> 00:08:56,799 the structure that would be best for the whole 199 00:08:56,799 --> 00:08:58,950 sentence once you see it all? That divergence, 200 00:08:59,110 --> 00:09:01,350 that conflict, is exactly where parsing difficulty 201 00:09:01,350 --> 00:09:04,269 and reanalysis costs come from. Let's use a concrete 202 00:09:04,269 --> 00:09:07,330 example. The dissertation brings up Fodor's relative 203 00:09:07,330 --> 00:09:10,009 clause length effects. This is a classic finding. 204 00:09:10,370 --> 00:09:12,490 It's a crucial one because it's one of the first 205 00:09:12,490 --> 00:09:15,690 big pieces of evidence that silent reading is 206 00:09:15,690 --> 00:09:18,269 sensitive to these subtle structural properties 207 00:09:18,269 --> 00:09:20,490 that we usually associate with spoken prosody. 208 00:09:20,629 --> 00:09:23,190 So can you refresh us on that pattern? Sure. 209 00:09:23,269 --> 00:09:26,679 In a complex noun phrase like, say, The professor 210 00:09:26,679 --> 00:09:29,659 saw the friend of the senior. What comes next 211 00:09:29,659 --> 00:09:32,919 is key. If the relative clause is short, like 212 00:09:32,919 --> 00:09:36,580 who studied, readers almost always prefer low 213 00:09:36,580 --> 00:09:39,259 attachment. Meaning the clause modifies senior, 214 00:09:39,379 --> 00:09:42,320 the closest noun. Right. But if the relative 215 00:09:42,320 --> 00:09:45,500 clause is long, who studied for hours and hours 216 00:09:45,500 --> 00:09:48,139 in the library, they suddenly flip their preference 217 00:09:48,139 --> 00:09:50,620 and prefer high attachment. So now it modifies 218 00:09:50,620 --> 00:09:52,950 friend. Correct. And the prosodic explanation 219 00:09:52,950 --> 00:09:56,169 for this is that readers are mentally, implicitly 220 00:09:56,169 --> 00:09:59,509 inserting a prosodic boundary before a long, 221 00:09:59,549 --> 00:10:02,490 heavy relative clause. That's the theory. That 222 00:10:02,490 --> 00:10:05,649 implicit boundary repels the RC, pushing it up 223 00:10:05,649 --> 00:10:07,950 to the higher attachment site, friend. But this 224 00:10:07,950 --> 00:10:10,669 brings us right back to the paradox of incrementality. 225 00:10:10,889 --> 00:10:13,509 It does. Because when the parser hits that word 226 00:10:13,509 --> 00:10:16,399 who, It has absolutely no idea if the relative 227 00:10:16,399 --> 00:10:18,700 clause is going to be short or long. It has to 228 00:10:18,700 --> 00:10:21,299 decide right then and there whether to introduce 229 00:10:21,299 --> 00:10:23,340 a boundary or not. It's playing chicken with 230 00:10:23,340 --> 00:10:25,620 the boundary. It is. Does it just go the default, 231 00:10:25,740 --> 00:10:27,460 which is usually attaching to the nearest lowest 232 00:10:27,460 --> 00:10:29,740 site, a principle we call late closure? Just 233 00:10:29,740 --> 00:10:32,320 keep the phrase open. Or does it wait a few words 234 00:10:32,320 --> 00:10:34,559 to see if the phrase gets heavy enough to justify 235 00:10:34,559 --> 00:10:37,399 putting in a boundary? And this introduces all 236 00:10:37,399 --> 00:10:40,220 these complex timing questions. We need to figure 237 00:10:40,220 --> 00:10:42,600 out the rules the parser is using to build this 238 00:10:42,600 --> 00:10:45,299 prosody on the fly to anticipate the final structure 239 00:10:45,299 --> 00:10:47,720 without knowing the full input. So the goal of 240 00:10:47,720 --> 00:10:49,960 the whole investigation, really, is to define 241 00:10:49,960 --> 00:10:52,940 the rules of that inner voice, to model the constraints 242 00:10:52,940 --> 00:10:55,220 on how that rhythmic structure is built word 243 00:10:55,220 --> 00:10:58,799 by word, to predict when and why the parser decides 244 00:10:58,799 --> 00:11:01,539 to insert that invisible but structure -altering 245 00:11:01,539 --> 00:11:04,559 pause. That's the challenge. Okay, so before 246 00:11:04,559 --> 00:11:06,879 we can trust any experimental results about this 247 00:11:06,879 --> 00:11:09,440 incremental process, the dissertation has to 248 00:11:09,440 --> 00:11:12,100 deal with a big methodological fear. A very big 249 00:11:12,100 --> 00:11:14,820 one. The fear that the experimental task itself 250 00:11:14,820 --> 00:11:17,059 might be corrupting the very thing we're trying 251 00:11:17,059 --> 00:11:19,419 to measure. This brings us to a comparison between 252 00:11:19,419 --> 00:11:22,940 two main reading tasks, the maze task and self 253 00:11:22,940 --> 00:11:25,179 -paced reading. This chapter is so important 254 00:11:25,179 --> 00:11:27,789 for the methodology. Because as researchers, 255 00:11:27,909 --> 00:11:30,110 we need tools that force people to make immediate, 256 00:11:30,289 --> 00:11:33,289 local processing decisions. The maze task is 257 00:11:33,289 --> 00:11:37,240 great for that, but it's design. It poses a serious 258 00:11:37,240 --> 00:11:39,460 threat to validity when you're studying something 259 00:11:39,460 --> 00:11:42,320 as subtle as prosody. So let's detail the maze 260 00:11:42,320 --> 00:11:44,799 task first. What does a participant actually 261 00:11:44,799 --> 00:11:48,200 do? In the maze task, the sentence is shown one 262 00:11:48,200 --> 00:11:51,929 word at a time, but at every single step, The 263 00:11:51,929 --> 00:11:53,710 participant has to choose between the correct 264 00:11:53,710 --> 00:11:57,009 next word and a foil word, a word that's plausible 265 00:11:57,009 --> 00:12:00,210 but grammatically wrong. Ah, so it forces a choice 266 00:12:00,210 --> 00:12:02,929 point, a pause, after every single word. A mandatory 267 00:12:02,929 --> 00:12:04,909 pause, yes. And how does that compare to the 268 00:12:04,909 --> 00:12:07,789 more traditional self -paced reading or SPR task? 269 00:12:08,190 --> 00:12:09,929 SPR is much simpler. You just press a button 270 00:12:09,929 --> 00:12:11,889 to see the next word. It measures how long you 271 00:12:11,889 --> 00:12:14,009 spend reading that word, and you often get what 272 00:12:14,009 --> 00:12:16,289 are called spillover effects. Meaning the difficulty 273 00:12:16,289 --> 00:12:18,289 from one word shows up in the reading time of 274 00:12:18,289 --> 00:12:20,110 the next word or two? Exactly. The processing 275 00:12:20,110 --> 00:12:22,610 cost can be delayed. So the big problem with 276 00:12:22,610 --> 00:12:24,730 the maze, which this research had to face head 277 00:12:24,730 --> 00:12:28,309 on, is the risk that all those mandatory pauses 278 00:12:28,309 --> 00:12:32,029 would induce a totally artificial unnatural prosody. 279 00:12:32,149 --> 00:12:34,649 A staccato rhythm, as they call it. Right. And 280 00:12:34,649 --> 00:12:36,710 if the maze forces a boundary after every word, 281 00:12:36,809 --> 00:12:39,250 then we can't possibly study how readers naturally 282 00:12:39,250 --> 00:12:42,509 generate boundaries. The task would be creating 283 00:12:42,509 --> 00:12:45,570 the effect. Precisely. So the first set of experiments 284 00:12:45,570 --> 00:12:48,240 were all about validating the maze. They had 285 00:12:48,240 --> 00:12:50,419 to prove it was sensitive, first to the really 286 00:12:50,419 --> 00:12:52,980 fine -grained parts of porosity, and then to 287 00:12:52,980 --> 00:12:55,399 the bigger, established phrasal effects. They 288 00:12:55,399 --> 00:12:57,779 started small, then, with metrical porosity, 289 00:12:57,980 --> 00:13:00,700 lexical stress. The smallest unit, right. This 290 00:13:00,700 --> 00:13:02,799 is about the rhythm inside individual words. 291 00:13:03,120 --> 00:13:06,399 The classic example is the noun convict, with 292 00:13:06,399 --> 00:13:08,480 stress on the first syllable, versus the verb 293 00:13:08,480 --> 00:13:11,820 convict, with stress on the second. And the idea 294 00:13:11,820 --> 00:13:13,879 for the experiment here was pretty clever. It 295 00:13:13,879 --> 00:13:16,460 was ingenious. They built these temporarily ambiguous 296 00:13:16,460 --> 00:13:19,039 sentences. For example, they'd lead a reader 297 00:13:19,039 --> 00:13:21,240 to first interpret a word like abstract as a 298 00:13:21,240 --> 00:13:24,299 noun. A -B -stract. And then a few words later, 299 00:13:24,419 --> 00:13:26,639 they'd hit a disambiguating word that forced 300 00:13:26,639 --> 00:13:29,919 them to realize it was actually a verb. Abstract. 301 00:13:30,200 --> 00:13:32,840 So the parser has to do a syntactic reanalysis. 302 00:13:32,919 --> 00:13:34,840 Yeah. It has to switch from noun to verb. But 303 00:13:34,840 --> 00:13:37,240 critically, in some of the sentences, that syntactic 304 00:13:37,240 --> 00:13:40,240 reanalysis also forced a metrical reanalysis. 305 00:13:40,240 --> 00:13:42,039 You had to change the stress pattern in your 306 00:13:42,039 --> 00:13:45,120 head. And the prediction is, if implicit prosody 307 00:13:45,120 --> 00:13:47,960 really matters. Then forcing someone to do both 308 00:13:47,960 --> 00:13:50,340 a syntactic and a metrical reanalysis should 309 00:13:50,340 --> 00:13:53,019 be way more costly than just the syntactic one 310 00:13:53,019 --> 00:13:54,860 alone. So what did they find? This is experiment 311 00:13:54,860 --> 00:13:58,100 one in the maze and two in SPR. The results were 312 00:13:58,100 --> 00:14:01,139 clear. Both tasks were sensitive to this tiny 313 00:14:01,139 --> 00:14:04,320 prosodic detail. There was consistently more 314 00:14:04,320 --> 00:14:06,720 processing difficulty, longer response times, 315 00:14:06,799 --> 00:14:09,340 more errors in the maze when the parser had to 316 00:14:09,340 --> 00:14:11,600 switch both the syntax and the stress pattern. 317 00:14:11,759 --> 00:14:13,720 Which confirms that even the super localized 318 00:14:13,720 --> 00:14:16,960 word -by -word decision in the maze task is incorporating 319 00:14:16,960 --> 00:14:20,840 implicit stress. Yes. It's not ignoring it. And 320 00:14:20,840 --> 00:14:22,759 what I find really fascinating is that the maze 321 00:14:22,759 --> 00:14:24,940 showed more localized effects. What do you mean 322 00:14:24,940 --> 00:14:27,159 by that? The slowdown, the moment of difficulty, 323 00:14:27,419 --> 00:14:30,080 happened immediately. Right on the word that 324 00:14:30,080 --> 00:14:33,769 forced the reanalysis. In the SPR task, as expected, 325 00:14:33,970 --> 00:14:36,230 the cost kind of spilled over onto the next couple 326 00:14:36,230 --> 00:14:39,149 of words. Ah, I see. So the maze might actually 327 00:14:39,149 --> 00:14:41,970 be better at pinpointing the exact moment to 328 00:14:41,970 --> 00:14:44,110 difficulty. It supports the idea that the maze 329 00:14:44,110 --> 00:14:46,470 is great for studying these rapid incremental 330 00:14:46,470 --> 00:14:50,070 decisions because its forced choices prevent 331 00:14:50,070 --> 00:14:53,070 the parser from sort of hiding the cost downstream. 332 00:14:53,450 --> 00:14:55,789 Okay, so the maze is good for word stress, but 333 00:14:55,789 --> 00:14:57,549 now they had to move up to the bigger question. 334 00:14:58,350 --> 00:15:01,029 Phrasal prosody. And this is where they test 335 00:15:01,029 --> 00:15:03,690 that core fear. Does the maze force a staccato 336 00:15:03,690 --> 00:15:05,929 prosody? This is the critical test for the whole 337 00:15:05,929 --> 00:15:08,289 methodology. And they used a classic sentence 338 00:15:08,289 --> 00:15:11,309 structure for it, the NPV garden path. An example 339 00:15:11,309 --> 00:15:13,190 would be something like, since the comedian left 340 00:15:13,190 --> 00:15:15,789 the manager. Exactly. The garden path effect 341 00:15:15,789 --> 00:15:17,870 there relies on the reader initially assuming 342 00:15:17,870 --> 00:15:21,169 one single uninterrupted phrase. The comedian 343 00:15:21,169 --> 00:15:23,480 left the manager. That's the NP interpretation, 344 00:15:23,720 --> 00:15:26,620 where manager is the object of left. Right. But 345 00:15:26,620 --> 00:15:28,519 the sentence could continue differently, making 346 00:15:28,519 --> 00:15:31,159 manager the subject of a new clause. That's the 347 00:15:31,159 --> 00:15:34,600 Z interpretation. Correct. And that Z interpretation 348 00:15:34,600 --> 00:15:38,200 is strongly dispreferred unless you have a really 349 00:15:38,200 --> 00:15:41,379 clear Q, like a comma. So the manipulation was 350 00:15:41,379 --> 00:15:44,740 just presence or absence of a comma. Now let's 351 00:15:44,740 --> 00:15:47,080 go back to that interference hypothesis. The 352 00:15:47,080 --> 00:15:51,379 fear was this. If the maze task forces an unnatural 353 00:15:52,110 --> 00:15:55,090 pause -heavy prosody on the reader, then it should 354 00:15:55,090 --> 00:15:57,570 automatically be inserting an implicit boundary 355 00:15:57,570 --> 00:16:00,350 after the verb left. And a boundary there would 356 00:16:00,350 --> 00:16:02,129 actually help the z -parse. It would separate 357 00:16:02,129 --> 00:16:04,490 manager from left, making it easy to see it as 358 00:16:04,490 --> 00:16:06,909 a new subject. You got it. So if the maze was 359 00:16:06,909 --> 00:16:09,690 flawed in the way people feared, it should have 360 00:16:09,690 --> 00:16:12,269 completely reversed the classic garden path effect. 361 00:16:12,450 --> 00:16:14,389 It should have made the z -interpretation easier. 362 00:16:14,590 --> 00:16:18,309 So? Not even close. The research clearly replicated 363 00:16:18,309 --> 00:16:21,419 the classic garden path in the maze task. Participants 364 00:16:21,419 --> 00:16:23,620 showed a huge slowdown when the sentence forced 365 00:16:23,620 --> 00:16:25,700 that Z interpretation without a calmer. Which 366 00:16:25,700 --> 00:16:27,679 proves they were going down the default MP path 367 00:16:27,679 --> 00:16:30,399 first and getting tripped up. Exactly. And that 368 00:16:30,399 --> 00:16:33,820 is a huge methodological takeaway. It shows the 369 00:16:33,820 --> 00:16:36,299 maze does not inherently introduce some weird 370 00:16:36,299 --> 00:16:39,419 boundary -heavy prosody that overrides these 371 00:16:39,419 --> 00:16:41,919 established parsing effects. It seems to capture 372 00:16:41,919 --> 00:16:45,600 the default natural contour just fine. Okay, 373 00:16:45,639 --> 00:16:48,659 so one final check. To really cement the maze's 374 00:16:48,659 --> 00:16:51,179 reliability for this kind of work, they had to 375 00:16:51,179 --> 00:16:53,620 tackle Fodor's length effect. The cornerstone 376 00:16:53,620 --> 00:16:56,059 fight. Right. The idea that long relative clauses 377 00:16:56,059 --> 00:16:58,399 drive high attachment because of the implicit 378 00:16:58,399 --> 00:17:01,320 boundary they generate. If the maze can't capture 379 00:17:01,320 --> 00:17:04,380 that subtle length -based effect, then it's kind 380 00:17:04,380 --> 00:17:06,319 of useless for the rest of this investigation. 381 00:17:06,740 --> 00:17:08,880 It would have been a deal breaker. But fortunately, 382 00:17:08,980 --> 00:17:11,839 both the maze in Experiment 5 and the SPR in 383 00:17:11,839 --> 00:17:14,380 Experiment 6 replicated the effect beautifully. 384 00:17:14,859 --> 00:17:17,819 So participants consistently chose the high attachment 385 00:17:17,819 --> 00:17:20,579 interpretation when the RCs were longer. Consistently. 386 00:17:20,619 --> 00:17:23,259 It validated the entire experimental approach. 387 00:17:23,599 --> 00:17:26,019 The maze is sensitive enough to pick up on the 388 00:17:26,019 --> 00:17:28,420 implicit boundary generated by sheer constituent 389 00:17:28,420 --> 00:17:30,640 size. And this is where the reading time data 390 00:17:30,640 --> 00:17:32,859 gets really provocative, and it links back to 391 00:17:32,859 --> 00:17:35,680 this idea that prosody can actually help processing. 392 00:17:35,880 --> 00:17:37,720 We're talking about the anti -locality effect. 393 00:17:38,059 --> 00:17:41,250 Yes. This was a surprising finding. They found 394 00:17:41,250 --> 00:17:43,589 that the region of the sentence after the long 395 00:17:43,589 --> 00:17:46,150 relative clause, so the main verb of the sentence, 396 00:17:46,349 --> 00:17:48,950 was actually read faster than the same region 397 00:17:48,950 --> 00:17:51,630 after the short relative clause. Which is completely 398 00:17:51,630 --> 00:17:53,930 counterintuitive. It's the definition of anti 399 00:17:53,930 --> 00:17:57,190 -locality. Making something longer and more complex 400 00:17:57,190 --> 00:17:59,450 actually makes a later part of the sentence easier 401 00:17:59,450 --> 00:18:02,829 to process. Right. Your intuition says a longer, 402 00:18:02,849 --> 00:18:06,049 more complex phrase should just exhaust the parser, 403 00:18:06,089 --> 00:18:08,640 slowing everything down later. So why does this 404 00:18:08,640 --> 00:18:11,140 happen? Well, the speculation centers on a couple 405 00:18:11,140 --> 00:18:14,940 of possibilities. One is about reduced surprisal. 406 00:18:15,119 --> 00:18:19,160 A very long, complex RC might just strongly constrain 407 00:18:19,160 --> 00:18:22,140 what can possibly come next grammatically. So 408 00:18:22,140 --> 00:18:24,039 the main verb becomes highly predictable. Okay, 409 00:18:24,119 --> 00:18:25,859 that's one idea. But the other possibility, the 410 00:18:25,859 --> 00:18:27,460 one that's really relevant to our deep dive, 411 00:18:27,599 --> 00:18:30,220 is implicit prosody. The mental pause button 412 00:18:30,220 --> 00:18:33,079 again. The mental pause button. The implicit 413 00:18:33,079 --> 00:18:35,599 boundary that the parser places after that long 414 00:18:35,599 --> 00:18:38,559 relative clause isn't just for attachment. It 415 00:18:38,559 --> 00:18:40,339 also acts as a signal to the parser. It says, 416 00:18:40,440 --> 00:18:43,440 OK, this complex dependent clause is structurally 417 00:18:43,440 --> 00:18:45,319 complete. We're done with it. It's like a mental 418 00:18:45,319 --> 00:18:47,539 restart button. That's a great way to think of 419 00:18:47,539 --> 00:18:50,180 it. It allows the parser to efficiently shift 420 00:18:50,180 --> 00:18:52,519 its focus back to the main clause and quickly 421 00:18:52,519 --> 00:18:55,000 process that matrix verb when it finally shows 422 00:18:55,000 --> 00:18:58,220 up. The cognitive cost of assigning the boundary 423 00:18:58,220 --> 00:19:00,880 is paid early, and the reward is a smoother, 424 00:19:00,960 --> 00:19:04,099 faster ride later on. That's a really powerful 425 00:19:04,099 --> 00:19:06,519 idea that the structural boundary we mentally 426 00:19:06,519 --> 00:19:09,640 impose serves this therapeutic dependency closing 427 00:19:09,640 --> 00:19:12,480 function, and it actually accelerates our overall 428 00:19:12,480 --> 00:19:14,579 parsing speed. It's a key part of the puzzle. 429 00:19:14,779 --> 00:19:16,180 All right, so we've established that implicit 430 00:19:16,180 --> 00:19:19,039 prosody is real, it's measurable, and it doesn't 431 00:19:19,039 --> 00:19:22,000 get messed up by the maze task. Now we can finally 432 00:19:22,000 --> 00:19:24,750 get into the theory. The fun part. How does the 433 00:19:24,750 --> 00:19:27,509 inner voice decide, word by word, where to put 434 00:19:27,509 --> 00:19:30,150 those crucial boundaries? This section of the 435 00:19:30,150 --> 00:19:32,509 work shifts into building a sort of toy model 436 00:19:32,509 --> 00:19:35,529 of incremental prosodic parsing. Right. We need 437 00:19:35,529 --> 00:19:37,690 to define the grammatical constraints that are 438 00:19:37,690 --> 00:19:40,329 guiding the parser. And the core tension here 439 00:19:40,329 --> 00:19:43,289 is that prosodic structure, the rhythmic phrasing, 440 00:19:43,369 --> 00:19:46,170 it generally mirrors syntactic structure, but... 441 00:19:46,730 --> 00:19:49,769 Not perfectly. You can have one grammatical structure 442 00:19:49,769 --> 00:19:53,450 with a few different possible rhythmic renderings. 443 00:19:53,509 --> 00:19:55,750 Exactly. So we need rules that govern the relationship 444 00:19:55,750 --> 00:19:58,210 between syntactic constituents, like a verb phrase, 445 00:19:58,390 --> 00:20:00,549 and prosodic constituents, like a phonological 446 00:20:00,549 --> 00:20:03,789 phrase. And the dissertation pulls from two major 447 00:20:03,789 --> 00:20:07,390 theoretical approaches here. It does. First, 448 00:20:07,470 --> 00:20:09,890 you have alignment theory. This theory is all 449 00:20:09,890 --> 00:20:12,210 about aligning the edges of syntactic constituents 450 00:20:12,210 --> 00:20:15,460 to the edges of prosodic constituents. For a 451 00:20:15,460 --> 00:20:17,680 parser that's working incrementally, this is 452 00:20:17,680 --> 00:20:19,920 a relatively simple calculation. Because it only 453 00:20:19,920 --> 00:20:22,319 needs to know where the current edge is. It doesn't 454 00:20:22,319 --> 00:20:24,000 need to have a picture of the whole structure 455 00:20:24,000 --> 00:20:26,099 yet. Correct. The alternative is match theory, 456 00:20:26,240 --> 00:20:28,980 which is a bit stricter. It requires a more direct 457 00:20:28,980 --> 00:20:31,559 one -to -one correspondence, where syntactic 458 00:20:31,559 --> 00:20:33,880 phrases have to directly match and cover the 459 00:20:33,880 --> 00:20:36,099 same material as prosodic phrases. And that would 460 00:20:36,099 --> 00:20:39,519 often require more complex, nested prosodic structures. 461 00:20:39,960 --> 00:20:42,950 It would. So for a parser that's working in real 462 00:20:42,950 --> 00:20:45,450 time with limited look -ahead, alignment constraints 463 00:20:45,450 --> 00:20:48,250 are just. They're easier to satisfy on the fly. 464 00:20:48,960 --> 00:20:52,059 But beyond just aligning to syntax, the parser 465 00:20:52,059 --> 00:20:54,660 also has to obey its own internal rhythmic preferences. 466 00:20:55,180 --> 00:20:57,619 And these are critical for explaining those length 467 00:20:57,619 --> 00:20:59,519 effects we just talked about. Absolutely critical. 468 00:20:59,759 --> 00:21:01,980 The internal rhythm of the parser, it seems, 469 00:21:02,180 --> 00:21:05,440 prefers balance and simplicity. Prosodic phrases 470 00:21:05,440 --> 00:21:07,779 tend to gravitate toward a binary structure. 471 00:21:08,319 --> 00:21:10,859 Meaning they prefer to contain just two prosodic 472 00:21:10,859 --> 00:21:13,339 words. That's the ideal, yeah. The inner voice 473 00:21:13,339 --> 00:21:15,799 is kind of a rhythm dictator that prefers simple 474 00:21:15,799 --> 00:21:18,200 two -beat measures. And this preference is enforced 475 00:21:18,200 --> 00:21:21,420 by specific constraints. It is. Constraints like 476 00:21:21,420 --> 00:21:25,960 bin max and bin max sight sime. Bin min penalizes 477 00:21:25,960 --> 00:21:28,579 phrases that are too short, less than two words. 478 00:21:28,900 --> 00:21:30,839 But the really important one for length effects 479 00:21:30,839 --> 00:21:33,680 is bin max, which penalizes phrases that are 480 00:21:33,680 --> 00:21:36,029 too long. Now, obviously, many phrases in English 481 00:21:36,029 --> 00:21:38,309 are longer than two words. Of course. But the 482 00:21:38,309 --> 00:21:40,890 idea behind bin max is that the penalty, and 483 00:21:40,890 --> 00:21:42,569 therefore the pressure on the parser to insert 484 00:21:42,569 --> 00:21:45,750 a boundary, increases exponentially as that constituent 485 00:21:45,750 --> 00:21:48,250 gets longer and longer. It's like pressure building 486 00:21:48,250 --> 00:21:51,549 up in a system. Exactly. As the current prosodic 487 00:21:51,839 --> 00:21:54,339 phrase keeps accumulating words, it gets heavy 488 00:21:54,339 --> 00:21:57,279 and starts violating binmax more and more. This 489 00:21:57,279 --> 00:21:59,240 violation makes it much more likely that the 490 00:21:59,240 --> 00:22:01,099 parser will just decide to break the phrase up 491 00:22:01,099 --> 00:22:03,500 and insert a boundary even before it's seen the 492 00:22:03,500 --> 00:22:05,299 end of the sentence. And this links directly 493 00:22:05,299 --> 00:22:08,740 back to Fodor's findings. A long RC is a heavy 494 00:22:08,740 --> 00:22:12,200 constituent. It's a huge binmax violation. It 495 00:22:12,200 --> 00:22:14,140 is. And on top of that, we also have what are 496 00:22:14,140 --> 00:22:17,240 called uniformity or balance constraints. Fodor's 497 00:22:17,240 --> 00:22:18,759 same -size sister principle is a good example. 498 00:22:19,150 --> 00:22:22,549 The parser likes adjacent prosodic phrases to 499 00:22:22,549 --> 00:22:25,109 be of singular structural weight. So a heavy 500 00:22:25,109 --> 00:22:27,650 phrase wants to pair up with another heavy phrase. 501 00:22:27,930 --> 00:22:31,029 Right. A long RC, which is a heavy constituent, 502 00:22:31,049 --> 00:22:33,430 searches for another heavy constituent to attach 503 00:22:33,430 --> 00:22:36,130 to for balance, like the high noun phrase, which 504 00:22:36,130 --> 00:22:39,069 contains both nouns. This whole prosodic balancing 505 00:22:39,069 --> 00:22:41,450 act is what's driving the attachment decision. 506 00:22:41,950 --> 00:22:44,049 Okay, let's put these rules into motion now and 507 00:22:44,049 --> 00:22:46,309 see how the parser actually builds prosody word 508 00:22:46,309 --> 00:22:48,730 by word. We have to start with a default behavior, 509 00:22:49,009 --> 00:22:51,190 which you mentioned earlier, prosodic late closure. 510 00:22:51,490 --> 00:22:54,309 Why is that the default? Why is it risk -averse? 511 00:22:54,589 --> 00:22:56,190 Well, because committing to a prosodic boundary 512 00:22:56,190 --> 00:22:59,349 too early is really risky. If you close off a 513 00:22:59,349 --> 00:23:01,710 phrase and then the sentence continues in a way 514 00:23:01,710 --> 00:23:03,690 that shows it should have stayed open, you've 515 00:23:03,690 --> 00:23:05,529 made a mistake and you have to go back and reanalyze. 516 00:23:05,589 --> 00:23:08,720 Which is costly. Very costly. So in the absence 517 00:23:08,720 --> 00:23:12,079 of a strong Q, like a comma, the parser's default 518 00:23:12,079 --> 00:23:14,220 is to just keep the current phonological phrase 519 00:23:14,220 --> 00:23:17,619 open, just keep adding material to it. This minimizes 520 00:23:17,619 --> 00:23:20,000 the risk of early reanalysis. Okay, let's walk 521 00:23:20,000 --> 00:23:22,259 through the short RC scenario first with a simple 522 00:23:22,259 --> 00:23:24,480 example. Let's use the brother of the psychiatrist 523 00:23:24,480 --> 00:23:28,059 who listened. So word by word. One, the brother. 524 00:23:28,259 --> 00:23:31,819 A new phonological phrase, a figat opens. Two, 525 00:23:32,099 --> 00:23:35,220 of the psychiatrist. This gets added to the open 526 00:23:35,220 --> 00:23:37,599 figure. It's not too heavy yet. Bean Max isn't 527 00:23:37,599 --> 00:23:39,900 screaming. Prosodic late closure keeps the phrase 528 00:23:39,900 --> 00:23:42,700 open. Three, who listened? This also gets added 529 00:23:42,700 --> 00:23:45,099 to the open figure. And the result is a single 530 00:23:45,099 --> 00:23:47,480 manageable prosodic phrase that covers the whole 531 00:23:47,480 --> 00:23:50,799 complex NP and the short RC. Right. No boundary 532 00:23:50,799 --> 00:23:53,039 is inserted. And because there's no boundary, 533 00:23:53,240 --> 00:23:55,980 the low attachment site, psychiatrist, is the 534 00:23:55,980 --> 00:23:58,519 most visible and local choice. The first pass 535 00:23:58,519 --> 00:24:00,779 prosody matches the final preferred prosody. 536 00:24:00,880 --> 00:24:04,380 Low attachment wins. Simple. Okay, now for the 537 00:24:04,380 --> 00:24:07,220 long RC scenario. This is the moment where the 538 00:24:07,220 --> 00:24:10,200 inner voice changes its mind. Let's say it's 539 00:24:10,200 --> 00:24:12,880 the brother of the psychiatrist who listened 540 00:24:12,880 --> 00:24:15,319 to the callers. The initial steps are identical. 541 00:24:15,599 --> 00:24:18,140 The parser is following late closure, so it's 542 00:24:18,140 --> 00:24:20,579 initially favoring low attachment. But as the 543 00:24:20,579 --> 00:24:23,059 phrase keeps growing, one, the brother of the 544 00:24:23,059 --> 00:24:26,059 psychiatrist, door is open. Two, who listened 545 00:24:26,059 --> 00:24:28,240 to the callers. This material just keeps getting 546 00:24:28,240 --> 00:24:30,559 added to that same open viola. The whole thing 547 00:24:30,559 --> 00:24:33,259 now contains many prosodic words. And the pressure 548 00:24:33,259 --> 00:24:35,279 is building. The pressure is building. The phrase 549 00:24:35,279 --> 00:24:38,480 becomes structurally heavy. The bin max constraint 550 00:24:38,480 --> 00:24:41,400 starts to impose a significant penalty. This 551 00:24:41,400 --> 00:24:43,339 makes the insertion of a boundary much, much 552 00:24:43,339 --> 00:24:45,619 more likely, even though the parser still hasn't 553 00:24:45,619 --> 00:24:47,920 seen the end of the sentence. So the sheer weight 554 00:24:47,920 --> 00:24:50,779 of the phrase forces the parser's hand. The parser 555 00:24:50,779 --> 00:24:53,140 knows, based on these internal rhythmic rules, 556 00:24:53,359 --> 00:24:55,900 that this long string of words is not a default. 557 00:24:56,349 --> 00:24:59,269 Simple phrase. That's it. So when the parser 558 00:24:59,269 --> 00:25:01,809 finally hits the main verb of the sentence, the 559 00:25:01,809 --> 00:25:04,190 unambiguous cue that the whole RC is finally 560 00:25:04,190 --> 00:25:07,089 over, it decides on its final prosodic structure. 561 00:25:07,740 --> 00:25:10,619 And given the huge binmax violation from the 562 00:25:10,619 --> 00:25:13,160 length, it's going to choose to insert a boundary 563 00:25:13,160 --> 00:25:16,839 between the complex MP and the RC. It partitions 564 00:25:16,839 --> 00:25:19,079 the sentence to satisfy those balance constraints. 565 00:25:19,339 --> 00:25:21,299 Right. And that boundary, once it's inserted, 566 00:25:21,440 --> 00:25:23,759 repels the RC, which drives high attachment. 567 00:25:24,509 --> 00:25:26,349 But this is where you see the crucial divergence. 568 00:25:26,769 --> 00:25:29,230 The first pass decision, driven by risk aversion 569 00:25:29,230 --> 00:25:31,849 and late closure, favored low attachment. But 570 00:25:31,849 --> 00:25:34,349 the final default prosody, driven by the length 571 00:25:34,349 --> 00:25:37,690 constraints and bin max, forces a boundary and 572 00:25:37,690 --> 00:25:39,890 mandates high attachment. And that potential 573 00:25:39,890 --> 00:25:42,109 for divergence, that's the key insight of the 574 00:25:42,109 --> 00:25:44,930 model. We can now hypothesize when the parser 575 00:25:44,930 --> 00:25:47,309 faces the highest cost of reanalysis, not just 576 00:25:47,309 --> 00:25:50,210 syntactically, but prosodically. It pinpoints 577 00:25:50,210 --> 00:25:52,329 the moment the inner voice realizes it has to 578 00:25:52,329 --> 00:25:55,049 change its tune. It clarifies everything. Okay, 579 00:25:55,109 --> 00:25:57,470 so the model successfully explains these length 580 00:25:57,470 --> 00:26:00,990 effects. But the dissertation, very wisely, looks 581 00:26:00,990 --> 00:26:03,950 ahead to the remaining mysteries. What foundational 582 00:26:03,950 --> 00:26:08,049 concepts did this initial model kind of put on 583 00:26:08,049 --> 00:26:10,309 the back burner? The initial model worked really 584 00:26:10,309 --> 00:26:12,470 well because it could rely mostly on alignment 585 00:26:12,470 --> 00:26:16,269 and binarity constraints to handle length. But 586 00:26:16,269 --> 00:26:19,130 future work has to grapple with deeply recursive 587 00:26:19,130 --> 00:26:21,750 structures. Complex sentences where phrases are 588 00:26:21,750 --> 00:26:24,509 nested inside other phrases. Exactly. And to 589 00:26:24,509 --> 00:26:26,930 model that, you really have to adopt stronger 590 00:26:26,930 --> 00:26:29,970 versions of match theory. And what specific constraints 591 00:26:29,970 --> 00:26:31,650 would come into play when you're dealing with 592 00:26:31,650 --> 00:26:34,230 that kind of nested structure? Well, things like 593 00:26:34,230 --> 00:26:36,690 strong start and equal sisters. Strong start 594 00:26:36,690 --> 00:26:39,069 is a constraint that basically ensures a major 595 00:26:39,069 --> 00:26:41,730 prosodic constituent begins with a strong metrical 596 00:26:41,730 --> 00:26:44,390 unit. And equal sisters is a constraint that 597 00:26:44,390 --> 00:26:47,109 pushes for prosodic sister constituents to be 598 00:26:47,109 --> 00:26:49,769 of the same prosodic category. Which reinforces 599 00:26:49,769 --> 00:26:52,309 that idea of structural balance and uniform recursion. 600 00:26:52,390 --> 00:26:55,089 It does, but it's a big jump in complexity, right? 601 00:26:55,170 --> 00:26:57,930 Moving from just deciding where to put a boundary. 602 00:26:58,470 --> 00:27:00,890 to ensuring all these nested rhythms match up 603 00:27:00,890 --> 00:27:03,190 perfectly. It seems like a huge jump. It is, 604 00:27:03,289 --> 00:27:05,450 but it's a necessary one if we want a complete 605 00:27:05,450 --> 00:27:09,190 model. If the parser is building complex syntactic 606 00:27:09,190 --> 00:27:11,670 structure incrementally, it must be building 607 00:27:11,670 --> 00:27:15,049 recursive prosodic structure in parallel. These 608 00:27:15,049 --> 00:27:17,089 constraints are the tools we'll need to model 609 00:27:17,089 --> 00:27:19,490 that process. We've built the theoretical engine. 610 00:27:19,589 --> 00:27:21,970 Now it's time for the ultimate experimental test. 611 00:27:22,349 --> 00:27:25,039 Chapter 4. Right. This is where the research 612 00:27:25,039 --> 00:27:26,980 moves away from just manipulating the length 613 00:27:26,980 --> 00:27:29,400 of the relative clause itself and focuses on 614 00:27:29,400 --> 00:27:32,519 modifying the noun phrase before the RC. The 615 00:27:32,519 --> 00:27:35,119 goal is to induce an implicit boundary precisely 616 00:27:35,119 --> 00:27:37,759 where they want it. Yes. The goal was to pit 617 00:27:37,759 --> 00:27:41,259 the two big competing forces, focus and attraction 618 00:27:41,259 --> 00:27:44,359 from the FAH versus boundaries and visibility 619 00:27:44,359 --> 00:27:46,880 from the VFH against each other in a controlled 620 00:27:46,880 --> 00:27:49,369 written environment. Can we use length modification 621 00:27:49,369 --> 00:27:52,289 to force an implicit boundary and then see if 622 00:27:52,289 --> 00:27:54,569 that boundary overrides the informational focus 623 00:27:54,569 --> 00:27:56,769 of the modified noun? That was the million -dollar 624 00:27:56,769 --> 00:27:59,150 question. Yeah. So instead of a simple phrase 625 00:27:59,150 --> 00:28:01,250 like the niece of the waitress, they created 626 00:28:01,250 --> 00:28:04,089 complex noun phrases like the niece of the incredibly 627 00:28:04,089 --> 00:28:06,549 diligent waitress. Adding that two -word modifier, 628 00:28:06,869 --> 00:28:09,549 incredibly diligent, dramatically increases the 629 00:28:09,549 --> 00:28:12,049 weight and the length of that N2 constituent. 630 00:28:12,329 --> 00:28:14,539 And... Based on the incremental model we just 631 00:28:14,539 --> 00:28:17,299 discussed, increasing the length of a constituent 632 00:28:17,299 --> 00:28:19,779 significantly increases the probability that 633 00:28:19,779 --> 00:28:22,539 the parser, under pressure from Benmax, will 634 00:28:22,539 --> 00:28:25,079 insert an implicit prosodic boundary right after 635 00:28:25,079 --> 00:28:27,839 it. Okay, so let's quickly review the three competing 636 00:28:27,839 --> 00:28:29,660 predictions for what would happen next. Sure. 637 00:28:29,720 --> 00:28:32,839 First, the repellent boundaries hypothesis, RBH. 638 00:28:33,319 --> 00:28:35,700 Modification creates a boundary, and that boundary 639 00:28:35,700 --> 00:28:38,660 repels the RC, pushing it high. Second. The modification 640 00:28:38,660 --> 00:28:43,059 attraction hypothesis, MA. F -A -H. The focus 641 00:28:43,059 --> 00:28:45,480 created by the modification attracts the RC to 642 00:28:45,480 --> 00:28:47,819 it. And third, the big one. The visibility -first 643 00:28:47,819 --> 00:28:51,240 hypothesis, the F -H. Boundaries are king. Attachment 644 00:28:51,240 --> 00:28:53,279 always targets the most visible, boundary -free 645 00:28:53,279 --> 00:28:55,380 site. Okay, let's get to the results. Experiment 646 00:28:55,380 --> 00:28:59,559 7, modified noun 2, N2. So the niece of the incredibly 647 00:28:59,559 --> 00:29:02,359 diligent waitress. And when N2 was modified, 648 00:29:02,619 --> 00:29:04,339 the interpretation data showed that participants 649 00:29:04,339 --> 00:29:07,140 preferred more low attachment. They attached 650 00:29:07,140 --> 00:29:10,039 the RSC to the modified N2, the waitress. Which 651 00:29:10,039 --> 00:29:13,130 is ambiguous. Completely ambiguous. That result 652 00:29:13,130 --> 00:29:15,930 could support the MAH focus attracts attachment 653 00:29:15,930 --> 00:29:19,730 to N2, or it could support the VFH if you assume 654 00:29:19,730 --> 00:29:22,349 the boundary lands after the modified N2, but 655 00:29:22,349 --> 00:29:25,789 N2 is still the most local visible site. So no 656 00:29:25,789 --> 00:29:28,309 clear winner there. So Experiment 7 didn't settle 657 00:29:28,309 --> 00:29:30,609 the debate, but Experiment 8 provided the critical 658 00:29:30,609 --> 00:29:34,369 test. This time they modified noun 1, N1. For 659 00:29:34,369 --> 00:29:36,789 example, the incredibly diligent niece of the 660 00:29:36,789 --> 00:29:38,670 waitress. And this is the moment where everything 661 00:29:38,670 --> 00:29:42,509 flips. Because if the MAHFAH were correct, modifying 662 00:29:42,509 --> 00:29:45,309 N1 should create focus on it, pulling the RC 663 00:29:45,309 --> 00:29:48,049 high to N1. But that's not what happened. Not 664 00:29:48,049 --> 00:29:49,970 at all. The response findings were completely 665 00:29:49,970 --> 00:29:52,730 counterintuitive. Modifying N1 led to more low 666 00:29:52,730 --> 00:29:55,170 attachment. Attachment to N2, the unmodified 667 00:29:55,170 --> 00:29:57,349 noun. Wait, hold on. Why would focusing on the 668 00:29:57,349 --> 00:29:59,289 niece make me think the relative clause refers 669 00:29:59,289 --> 00:30:01,910 to the waitress? That seems to completely contradict 670 00:30:01,910 --> 00:30:04,569 the idea that focus attracts attachment. It does. 671 00:30:04,710 --> 00:30:07,569 It shatters it. The MAHFEH failed this test. 672 00:30:07,730 --> 00:30:10,250 And this is the strong, almost undeniable evidence 673 00:30:10,250 --> 00:30:12,250 that elevates the visibility first hypothesis. 674 00:30:12,589 --> 00:30:14,569 Okay, walk us through the VFH explanation for 675 00:30:14,569 --> 00:30:16,529 this exact result. So when the parser encounters 676 00:30:16,529 --> 00:30:19,589 the incredibly diligent niece, the length of 677 00:30:19,589 --> 00:30:22,880 that N1 constituent goes way up. The internal 678 00:30:22,880 --> 00:30:25,180 rhythmic constraints, binax, are just screaming. 679 00:30:25,660 --> 00:30:28,220 This forces the parser to insert an implicit 680 00:30:28,220 --> 00:30:31,460 prosodic boundary right after N1. So there's 681 00:30:31,460 --> 00:30:34,460 an invisible wall after knees. Exactly. And that 682 00:30:34,460 --> 00:30:37,859 wall makes N1 invisible. Even though N1 carried 683 00:30:37,859 --> 00:30:40,200 the focus in the modification, that boundary 684 00:30:40,200 --> 00:30:43,359 renders it structurally inaccessible. Noun 2, 685 00:30:43,599 --> 00:30:46,180 waitress, is now immediately adjacent to the 686 00:30:46,180 --> 00:30:49,000 incoming relative pronoun who. It has become 687 00:30:49,000 --> 00:30:51,380 the most local and most visible attachment site. 688 00:30:51,670 --> 00:30:54,289 And the parser, prioritizing visibility over 689 00:30:54,289 --> 00:30:58,049 focus, attaches low. It attaches low to N2. This 690 00:30:58,049 --> 00:31:01,490 result confirms the hierarchy. Implicit prosodic 691 00:31:01,490 --> 00:31:03,670 boundaries, which are driven by constituent length, 692 00:31:03,930 --> 00:31:05,990 are the dominant factor in written ambiguity 693 00:31:05,990 --> 00:31:08,640 resolution. That's just a fundamental clarifying 694 00:31:08,640 --> 00:31:10,700 point on how we read. It's not about what we 695 00:31:10,700 --> 00:31:13,380 mentally highlight with focus. It's about where 696 00:31:13,380 --> 00:31:15,940 our internal rhythm forces us to take a pause. 697 00:31:16,140 --> 00:31:17,839 And to make sure this wasn't some kind of fluke, 698 00:31:17,859 --> 00:31:20,740 Experiment 9 generalized the findings. They crossed 699 00:31:20,740 --> 00:31:24,259 both NP modification types, N1 or N2, with the 700 00:31:24,259 --> 00:31:26,839 classic RC length effect. The findings generalized 701 00:31:26,839 --> 00:31:30,500 perfectly. N1 modification and N2 modification 702 00:31:30,500 --> 00:31:33,500 both consistently drove attachment to the lower. 703 00:31:33,869 --> 00:31:37,390 N2 site, no matter how long the RC was. It just 704 00:31:37,390 --> 00:31:40,349 reinforces the VFH. Length leads to boundaries. 705 00:31:40,730 --> 00:31:43,269 Boundaries lead to invisibility. And visibility 706 00:31:43,269 --> 00:31:46,009 guides attachment, period. Before we wrap up 707 00:31:46,009 --> 00:31:47,930 the experiments, we have to revisit the timing 708 00:31:47,930 --> 00:31:50,230 data because it offered this tantalizing clue 709 00:31:50,230 --> 00:31:53,450 about when the cognitive cost of this whole process 710 00:31:53,450 --> 00:31:55,650 happens. The reading times, right. Where did 711 00:31:55,650 --> 00:31:58,670 the parser actually slow down relative to where 712 00:31:58,670 --> 00:32:01,369 we think the boundary is? Across multiple experiments, 713 00:32:01,569 --> 00:32:04,170 there was this really suggestive pattern. The 714 00:32:04,170 --> 00:32:06,470 slowdown didn't happen before the boundary, which 715 00:32:06,470 --> 00:32:08,269 you might expect if the cost was in the decision 716 00:32:08,269 --> 00:32:10,769 -making. Instead, it happened on the word immediately 717 00:32:10,769 --> 00:32:13,450 following the hypothesized boundary. Can you 718 00:32:13,450 --> 00:32:15,890 give us an example of that? Sure. When N1 was 719 00:32:15,890 --> 00:32:18,190 modified, like in the incredibly diligent Nice, 720 00:32:18,430 --> 00:32:21,549 we predict a boundary right after Nice. The slowdown 721 00:32:21,549 --> 00:32:23,670 in reading time was observed on the very next 722 00:32:23,670 --> 00:32:27,390 word, the preposition of. And similarly for N2. 723 00:32:27,829 --> 00:32:30,529 Same thing. When N2 was modified, the boundary 724 00:32:30,529 --> 00:32:33,009 was predicted after the whole complex N2 phrase, 725 00:32:33,170 --> 00:32:35,650 and the slowdown was observed on the relative 726 00:32:35,650 --> 00:32:38,309 pronoun who, the word that came right after. 727 00:32:38,470 --> 00:32:42,640 This timing is so fascinating. If the slowdown 728 00:32:42,640 --> 00:32:45,460 happens after the decision point, what does that 729 00:32:45,460 --> 00:32:48,319 tell us about the parser's mechanism? It really 730 00:32:48,319 --> 00:32:50,440 supports that concept of prosodic late closure. 731 00:32:51,119 --> 00:32:53,519 It suggests the slowdown might be the cognitive 732 00:32:53,519 --> 00:32:56,099 cost of either realizing the expected boundary 733 00:32:56,099 --> 00:32:59,039 has now occurred or the cost of actively assigning 734 00:32:59,039 --> 00:33:01,519 that boundary once the constituent length has 735 00:33:01,519 --> 00:33:03,819 passed some critical threshold. So the parser 736 00:33:03,819 --> 00:33:06,000 is basically holding off on committing to the 737 00:33:06,000 --> 00:33:08,779 pause until the very last second. It waits until 738 00:33:08,779 --> 00:33:10,900 the structural pressure from BinMax builds up 739 00:33:10,900 --> 00:33:13,079 and then the cost is incurred at the moment of 740 00:33:13,079 --> 00:33:14,880 commitment. And that registers as the processing 741 00:33:14,880 --> 00:33:17,339 time on the very next word. It's just highly 742 00:33:17,339 --> 00:33:19,849 compatible with the whole model. Risk aversion 743 00:33:19,849 --> 00:33:22,630 keeps the phrase open, length increases the pressure 744 00:33:22,630 --> 00:33:24,789 to close it, and the cost of finally closing 745 00:33:24,789 --> 00:33:27,109 it shows up just after that crucial decision 746 00:33:27,109 --> 00:33:29,250 point. It all fits together. We've covered the 747 00:33:29,250 --> 00:33:31,809 methodology, the theory, and the definitive experimental 748 00:33:31,809 --> 00:33:35,029 findings. If we take a step back, what is the 749 00:33:35,029 --> 00:33:37,269 integrated picture of silent reading that we're 750 00:33:37,269 --> 00:33:39,730 left with? Well, first, the methodological point 751 00:33:39,730 --> 00:33:42,910 is rock solid. Okay. The maze task is a sensitive 752 00:33:42,910 --> 00:33:45,650 and reliable tool for studying implicit prosody. 753 00:33:46,109 --> 00:33:48,250 It gives you these localized effects that are 754 00:33:48,250 --> 00:33:50,710 incredibly valuable, and it does not force an 755 00:33:50,710 --> 00:33:53,009 unnatural rhythm. And second, the major theoretical 756 00:33:53,009 --> 00:33:56,349 takeaway is the hierarchy of influences. We have, 757 00:33:56,450 --> 00:33:58,589 I would say, overwhelming evidence now that the 758 00:33:58,589 --> 00:34:00,950 parser's primary guide during ambiguity resolution 759 00:34:00,950 --> 00:34:04,569 is the visibility -first hypothesis. Attachment 760 00:34:04,569 --> 00:34:07,190 preferences are governed first and foremost by 761 00:34:07,190 --> 00:34:09,590 implicit prosodic boundaries generated during 762 00:34:09,590 --> 00:34:12,610 reading. And the role of information, structure 763 00:34:12,610 --> 00:34:16,190 focus, modification, accent is really relegated 764 00:34:16,190 --> 00:34:18,929 to a secondary modulatory status. It only kicks 765 00:34:18,929 --> 00:34:21,230 in when prosodic visibility is neutral. That's 766 00:34:21,230 --> 00:34:23,389 right. The finding that just making N1 longer 767 00:34:23,389 --> 00:34:26,090 forces a boundary that completely overrides N1's 768 00:34:26,090 --> 00:34:29,010 focus. That is the defining insight of the entire 769 00:34:29,010 --> 00:34:31,889 investigation. Length dictates boundaries and 770 00:34:31,889 --> 00:34:33,929 boundaries dictate meaning. And let's connect 771 00:34:33,929 --> 00:34:35,929 this back to that reading time data, the anti 772 00:34:35,929 --> 00:34:38,530 -locality effect. Right. The faster reading after 773 00:34:38,530 --> 00:34:41,409 a long RC. It strongly suggests that implicit 774 00:34:41,409 --> 00:34:44,550 prosody isn't just affecting what the parser 775 00:34:44,550 --> 00:34:47,489 attaches to, but also how fast it processes what 776 00:34:47,489 --> 00:34:49,989 comes next. The boundary placement acts as a 777 00:34:49,989 --> 00:34:53,190 structural signpost. It signals the closure of 778 00:34:53,190 --> 00:34:55,909 a complex dependency, and that seems to ease 779 00:34:55,909 --> 00:34:58,550 the load on the parser as it transitions back 780 00:34:58,550 --> 00:35:01,030 to the main clause. It really makes that inner 781 00:35:01,030 --> 00:35:04,230 voice a core computational component of the brain, 782 00:35:04,449 --> 00:35:07,170 affecting both our interpretation and our processing 783 00:35:07,170 --> 00:35:10,500 speed. But this work... As successful as it is, 784 00:35:10,619 --> 00:35:13,400 it leaves these profound questions for future 785 00:35:13,400 --> 00:35:16,639 research about the inner voice. The unfinished 786 00:35:16,639 --> 00:35:19,539 work of prosodic parsing, what are the key areas 787 00:35:19,539 --> 00:35:21,579 we still need to investigate? We need better 788 00:35:21,579 --> 00:35:24,239 confirmation tools. I mean, while the data strongly 789 00:35:24,239 --> 00:35:26,119 suggests where these implicit boundaries are 790 00:35:26,119 --> 00:35:28,659 being placed, we don't have direct evidence of 791 00:35:28,659 --> 00:35:30,820 the reader's internal prosodic structure. We're 792 00:35:30,820 --> 00:35:33,690 inferring it from their choices. We are. So future 793 00:35:33,690 --> 00:35:35,690 studies need to develop new techniques, maybe 794 00:35:35,690 --> 00:35:38,389 something like rapid prosodic transcription tasks, 795 00:35:38,610 --> 00:35:41,590 to confirm precisely where readers actually place 796 00:35:41,590 --> 00:35:43,929 these boundaries and then correlate that data 797 00:35:43,929 --> 00:35:46,250 directly with their parsing choices. To verify 798 00:35:46,250 --> 00:35:49,250 that the binmax boundary we predict is in fact 799 00:35:49,250 --> 00:35:51,190 the boundary the reader is actually generating. 800 00:35:51,429 --> 00:35:54,269 Exactly. And then on the theory side, we have 801 00:35:54,269 --> 00:35:56,590 to address those complex nested structures we 802 00:35:56,590 --> 00:35:58,840 touched on. We need to incorporate constraints 803 00:35:58,840 --> 00:36:01,579 like strong start and equal sisters and see if 804 00:36:01,579 --> 00:36:03,880 the parser really generates recursive prosodic 805 00:36:03,880 --> 00:36:06,320 structure. That seems like a massive next step 806 00:36:06,320 --> 00:36:08,820 to determine if our inner voice is capable of 807 00:36:08,820 --> 00:36:12,000 generating fully recursive nested rhythmic structures 808 00:36:12,000 --> 00:36:15,159 just based on visual input from a page. It is. 809 00:36:15,280 --> 00:36:18,019 But the goal remains to build that robust incremental 810 00:36:18,019 --> 00:36:20,699 model that can explain the divergence between 811 00:36:20,699 --> 00:36:23,389 the immediate risk -averse prosody of the first 812 00:36:23,389 --> 00:36:26,269 pass, and the optimal balanced final prosody 813 00:36:26,269 --> 00:36:28,670 of the full sentence. The work detailed in this 814 00:36:28,670 --> 00:36:30,769 investigation confirms that the sound of silence 815 00:36:30,769 --> 00:36:32,989 is loud enough to change the entire meaning of 816 00:36:32,989 --> 00:36:35,329 the words we read. And that brings us to our 817 00:36:35,329 --> 00:36:38,329 final provocation for you, the listener. If length 818 00:36:38,329 --> 00:36:41,070 and visual structure fundamentally dictate the 819 00:36:41,070 --> 00:36:44,289 unseen rhythm in your head, consider this. The 820 00:36:44,289 --> 00:36:46,090 next time you're struggling with an ambiguous 821 00:36:46,090 --> 00:36:49,670 sentence in an email or a document, Perhaps the 822 00:36:49,670 --> 00:36:52,190 difficulty isn't syntactic at all. Maybe your 823 00:36:52,190 --> 00:36:54,710 brain is simply arguing with itself about where 824 00:36:54,710 --> 00:36:56,789 the inner pause button should have been placed. 825 00:36:57,010 --> 00:37:00,469 The parser isn't just reading words. It's constantly 826 00:37:00,469 --> 00:37:02,590 constructing a rhythm and then listening to the 827 00:37:02,590 --> 00:37:05,250 structure that it creates. Understanding this 828 00:37:05,250 --> 00:37:08,170 internal orchestra might just be the key to understanding 829 00:37:08,170 --> 00:37:09,030 comprehension itself.