Debate quantification: How MAD did they get?

Cynthia McLemore noted on Facebook:

The last couple of debates (including this one) Trump has started off with a very narrow pitch range, using a monotone. I guess he's been coached not to sound 'emotional'. The only question is when the lid comes off…

Here's a simple attempt at quantification of this insight.

I pitch-tracked each candidate's (nominally 2-minute-long) answers to each of the first three questions. (This covers roughly the first half of the debate — though I used only the three two-minute answers, because the rest of the discussion gets pretty messy and I don't have time today to do a careful transcript of all the back-and-forth.)

Then I transformed the f0 estimates into differences in semitones from the median value for each answer, and took the median of the absolute value of those differences.

I did the same thing for Chris Wallace's introduction and questions. In the plot below, the MAD ("median absolute deviation") from the median f0 of Trump's first three answers are plotted as T1, T2,and T3; the corresponding values for Clinton are plotted as C1, C2, and C3; and the corresponding value for Wallace's contributions taken as a whole is plotted as W:

So it's true — Clinton remains fairly level, at about the same MAD value as Wallace, but Trump starts out much more monotone, and ends up more prosodically animated.

Trump's median f0 values also increased (from 112 to 144 Hz), as did his pitch range  measured as the 90th percentile minus 10th percentile f0 values, which increase from 26.6 Hz to 83.3 Hz).

It's an extra step to the explanation offered by the New Yorker cartoon that McLemore added as a comment:

…but a plausible one.

He's still waiting

From Francois Lang:

Attached is a photo of a sign in the washroom at Heckman's Deli in Bethesda, MD

I kept waiting for all the employees to wash my hands. I even asked. But nothing. Maybe it was something I said?

In another example of fun with pronouns and implicit subjects, my university's front page currently features this captioned image:

A few of the many LLOG posts on the implicit-subjects question, subcategory "dangling participles":

"Dangling etiquette", 12/14/2003
"Stunning inept modifier manners", 3/10/2005
"Dangling modifier in The Declaration of Independence", 7/4/2005
"The Fellowship of the Predicative Adjunct", 5/12/2005
"A dangler in The Economist", 10/8/2009
"Does it really matter if it dangles?", 11/20/2010

More old names for Singapore

We have already studied an old name for Singapore on the back of an envelope dating to 1901:

Now, Ruben de Jong, relying on the works of Dutch scholars, has discovered several others.

The first comes from Gustav Schlegel, ‘Chinese loanwords in the Malay language’, T’oung Pao 1.5 (1891):  391-405, 393.

Here's a screen shot of the relevant entry, viz. 寔辣:

The pronunciation Sít-lát is based on Hokkien, as in the previous blog post.  However, we have not seen these specific characters in the previous post nor in the added example in the comments by Kirinputra.

Besides the historic predominance of Hokkien in Singapore and surrounding region (at least among the Chinese), finding this transcription in this particular topolect makes sense because Schlegel, like other Dutch Sinologists in the nineteenth century, mainly studied fāngyán 方言 like Hokkien and Hakka. They did so because these were the Sinitic languages spoken in the Dutch East Indies (today’s Indonesia), and their training as Sinologists was based on a future career as civil servants (interpreters and translators) in the colony (for more, see the excellent recent PhD dissertation of Pieter Nicolaas Kuiper on the topic).

Schlegel cites a certain Pijnappel, i.e., his Malay-Dutch dictionary, and from this work Schlegel might have gotten some of his information:  e.g., Schlegel connects the Hokkien name to the Malay word for “strait” (as we did in the previous post) .

Besides Pijnappel's dictionary, Schlegel’s own multivolume Dutch-Hokkien dictionary (i.e., in Zhangzhou topolect, literary reading) might also be consulted. This work has been partly digitized, although the volume most likely to contain the name for Singapore is missing (see here).

In the name 寔辣, 寔 is a variant of shí 實 / 实 ("real; actual; solid; true"). See this zdic entry.

The second item comes from P.A. van de Stadt, Hakka-woordenboek (Batavia: Landsdrukkerij, 1912).  As the title indicates, this is a dictionary for Hakka and was based on the topolect as spoken in Bangka and Belitung, Indonesia.

Here is p. 409:

It shows two names for Singapore.  Once again, the first character of the second name (Sit lát), 息 ("interest; rest; breath; news"), has not been mentioned before in our discussions on this subject. A nice aside: the Dutch text in the right-hand column informs the reader that often similar sounding characters are used.

The third name comes from J. J. C. Francken and C. F. M. de Grijs, Chineesch-Hollandsch woordenboek van het Emoi dialekt (Batavia: Landsdrukkerij, 1882).  In spite of its title indicating that this is a dictionary for the Emoi (Amoy) dialect, this work – like that of Schlegel – is also in part based on Zhangzhou Hokkien (cf. p. 404 of the PhD diss. by Kuiper, mentioned above).

Here is p. 537:

We see the same character as in Van de Stadt’s work, but with a mouth radical added on the left, 口+息.

The upshot of all this is that, in studying the Sinitic transcriptions of foreign names, we must take into account topolectal variation and the secondary nature of graphic representation:  what matters most are the sounds, not the shapes.

Don't let 'bigly' catch on

Scott Adams, the Dilbert cartoon creator and diehard Trump promoter, has taken to the semi-jocular practice of adopting the mishearing of Trump's much-loved adjunct big-league, and using bigly as if it were a real adverb ("I just watched the debate on replay. Trump won bigly. This one wasn't close"). Adams is kidding, I think, but the mishearing is very common: by May 5, bigly was getting over 70,000 hits in the Google News index. I'm worried it may catch on, and we'll wake up some morning not only with the orange-quiffed sexist boor in the White House but with bigly added to the stock of adverbs in standard English.

Believe it or not, there is potential theoretical relevance. If the formation of adverbs like carefully from adjective stems like careful is a matter of lexical word formation (in other words, if it is derivational morphology, hence appropriately discussed in Chapter 18 of The Cambridge Grammar of the English Language), then we should expect to find arbitrary lexical exceptions, and the absence of lexemes such as *bigly, *littlely, *bluely, *oldly, *youngly, etc., is to be expected. Lexical word-formation is riddled with such gaps: take a look at the table I gave in this post, comparing the strikingly irregular pattern seen in horror / horrify / horrific / horrid with terror / terrify / terrific / *terrid and stupor / stupefy / *stupific / stupid and fervor / *fervify / *fervific / fervid and so on.

But if adjective and adverb are the same category and -ly is just an inflectional affix that the words in that category have to take in some syntactic contexts (for example, when they modify verbs), then we should expect much broader application of the suffix. Inflectional affixes show a strong tendency to apply very widely across items in a given category, not picking and choosing at random the way that -ify and -ic and -id seem to do.

In short, every new case of an adjective taking the -ly suffix to form an adverb is a data point that might undercut the idea that -ly is a derivational affix that forms adverbs (not an inflectional affix sometimes required on adjectives).

Of course, the adjectives of English already happen to have a core subset that take inflections that most other adjectives reject. Compare big / bigger / biggest, small / smaller / smallest, cool / cooler / coolest, etc., with intelligent / *intelligenter / *intelligentest, fascinating / *fascinatinger / *fascinatingest, important / *importanter / *importantest, cantankerous / *cantankerouser / *cantankerousest, etc. So a defender of the view that -ly is an inflection would simply say that adjectives are generally pickier than verbs or nouns when it comes to acceptance of inflectional suffixes, but we should expect a word like big to move over time into the long list of adjectives that take the -ly inflection in addition to the -er and -est inflections.

The differentiators and the collapsers can both cover the facts. We are still looking for a clincher. It may surprise you (it surprises me) that 400 years of work on describing the grammar of standard English, including the past 50 years of quite intensive investigation, have not managed to settle an issue as basic as whether a distinct lexical category of adverbs should be posited in English.

Tradition favors the differentiators. It is traditional to assume that the two categories are distinct, as a glance at almost any grammar at all will confirm. But that doesn't mean all those grammars are right; on identifying the members of the class of prepositions, for example, I claim they are all wrong.

However, I currently think that positing a separate category of adverbs is not wrong. I'm a differentiator. A long paper in 2010 entitled "The distribution and category status of adjectives and adverbs" (Word Structure 3(1), 31–81) argues that the traditional view on this point is indeed correct, and the -ly adverb-forming suffix is derivational. I'm sure you will trust me when I tell you that the paper is truly magisterial, a triumph of human intellect, the definitive work on this topic. Though when later you discover that I am one of its authors (it is by John Payne, Rodney Huddleston, and me), that may undermine your trust a little ("He would say that, wouldn't he?").

The opposing view — that in some Indo-European languages the adverbs should be collapsed with the adjectives into a single category — has been advocated by a number of linguists. Jerzy Kurylowicz proposed it for French (in 1936); Gerard Moignet developed the case further (in 1963); Joseph Emonds championed it for English (in 1976); Andrew Radford pushed for it too (in 1988); so did several other linguists. (The relevant references can be found in the bibliography of the Payne/Huddleston/Pullum paper.)

Recently the collapsers have attracted support through a number of new arguments presented in a 2012 paper by Heinz Giegerich: "The morphology of -ly and the categorial status of 'adverbs' in English" (English Language & Linguistics 16(3), 341–359). Yes, Giegerich has the effrontery not just to disagree with Payne, Huddleston, and me, but in addition to put the word "adverb" in scare quotes in his title as if it were disreputable! Scoundrel.

Naturally I would like to see Giegerich's paper reduced to a laughing-stock, retracted by the publishers as worthless, fraudulent, and a disservice to science. But it's not as easy as that. Personally, I face a real problem. Heinz Giegerich works just down the hall in my building. He's an important member of my department, and one of my favorite colleagues. His work is respected. Dammit, I like the guy. And even more relevantly, his arguments are perfectly intelligible; I don't disagree with any of his facts. He just sees different implications emerging from the facts. So now what?

The Payne/Huddleston/Pullum paper and the Giegerich paper added together occupy 68 pages of learned journal space. Other respectable and crucially relevant articles on the topic that I know about take up far more than that (and there will be yet more contributions that I still haven't found out about). The arguments are complex and both sides make sound and persuasive points.

The truth is that linguists cannot at present settle the question of whether the (traditionally so-called) adverbs of English should be segregated from the adjectives in a separate category, or whether there should be one syntactic category that includes them all, -ly being merely a suffix that some of them have to bear in some contexts. It's just one indication that grammatical analysis is really quite a bit harder than you might think.

Human parity in conversational speech recognition

Today at ISCSLP2016, Xuedong Huang announced a striking result from Microsoft Research. A paper documenting it is up on — W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer, A. Stolcke, D. Yu, G. Zweig, "Achieving Human Parity in Conversational Speech Recognition":

Conversational speech recognition has served as a flagship speech recognition task since the release of the DARPA Switchboard corpus in the 1990s. In this paper, we measure the human error rate on the widely used NIST 2000 test set, and find that our latest automated system has reached human parity. The error rate of professional transcriptionists is 5.9% for the Switchboard portion of the data, in which newly acquainted pairs of people discuss an assigned topic, and 11.3% for the CallHome portion where friends and family members have open-ended conversations. In both cases, our automated system establishes a new state-of-the-art, and edges past the human benchmark. This marks the first time that human parity has been reported for conversational speech. The key to our system's performance is the systematic use of convolutional and LSTM neural networks, combined with a novel spatial smoothing method and lattice-free MMI acoustic training.

30 years ago, things were quite different — Philippe, Jeanrenaud, Ellen Eide, U. Chaudhari, J. McDonough, Kenney Ng, M. Siu, and Herbert Gish, "Reducing word error rate on conversational speech from the Switchboard corpus", IEEE ICASSP 1995:

Speech recognition of conversational speech is a difficult task. The performance levels on the Switchboard corpus had been in the vicinity of 70% word error rate. In this paper, we describe the results of applying a variety of modifications to our speech recognition system and we show their impact on improving the performance on conversational speech. These modifications include the use of more complex models, trigram language models, and cross-word triphone models. We also show the effect of using additional acoustic training on the recognition performance. Finally, we present an approach to dealing with the abundance of short words, and examine how the variable speaking rate found in conversational speech impacts on the performance. Currently, the level of performance is at the vicinity of 50% error, a significant improvement over recent levels.

And even a few months ago, reported error rates on the same task were substantially higher — Tom Sercu,  Christian Puhrsch, Brian Kingsbury, and Yann LeCun, "Very deep multilingual convolutional neural networks for LVCSR", IEEE ICASSP 2016:

Convolutional neural networks (CNNs) are a standard component of many current state-of-the-art Large Vocabulary Continuous Speech Recognition (LVCSR) systems. However, CNNs in LVCSR have not kept pace with recent advances in other domains where deeper neural networks provide superior performance. In this paper we propose a number of architectural advances in CNNs for LVCSR. First, we introduce a very deep convolutional network architecture with up to 14 weight layers. There are multiple convolutional layers before each pooling layer, with small 3×3 kernels, inspired by the VGG Imagenet 2014 architecture. Then, we introduce multilingual CNNs with multiple untied layers. Finally, we introduce multi-scale input features aimed at exploiting more context at negligible computational cost. We evaluate the improvements first on a Babel task for low resource speech recognition, obtaining an absolute 5.77% WER improvement over the baseline PLP DNN by training our CNN on the combined data of six different languages. We then evaluate the very deep CNNs on the Hub5'00 benchmark (using the 262 hours of SWB-1 training data) achieving a word error rate of 11.8% after cross-entropy training, a 1.4% WER improvement (10.6% relative) over the best published CNN result so far.

This is not the end of the speech-recognition story. There are harder tasks where there's plenty of room for improvement. And perhaps it's time to go beyond simple evaluation in terms of overall word error rate, since some errors are more consequential than others. But still, this is an important and impressive milestone.

[See also the MS blog entry "Historic Achievement: Microsoft researchers reach human parity in conversational speech recognition".]



Knife and fork

Nathan Hopson came across a marvelous Japanese word from the interwar period the other day:  naihoku ナイホク.

Nathan first saw naihoku ナイホク, a portmanteau (or just contraction?) of "knife and fork" (naifu [to] fōku ナイフ[と]フォーク) on p. 10, l. 7 of a delightful 1929 guide to the famous eateries of Tokyo (Tōkyō meibutsu tabe aru ki 東京名物食べある記):

Daibu konde kitanode gururi to mimawasu to, Ichōgaeshi, Shimada nado nado / tō tō, ikina konshun watari no nēsan-tachi ga gokumi, rokkumi, nareta tetsuki de naihoku o tsukatte iru no wa hoka ni mirarenai zu de aru.


Roughly: "It had gotten pretty crowded. When I looked around, there was the unique sight of five or six groups of fashionable young women [in kimonos? and with their hair formally done up] using knives and forks with a practiced hand."

"Naihoku ナイホク" seems to be older, though, appearing in #5 of Akutagawa Ryūnosuke's (1892-1927) brief series of urban vignettes, Tokai de — Aruiwa sen kyū hyaku roku nen no Tōkyō — 都会で — 或は千九百十六年の東京 — (In the city — Or, Tokyo in the year 1916).  Interestingly, it is both glossed with a note (that takes up a full line of the two-line section!), and is already being used as a kind of workplace slang to mean something more than simply "knives and forks":

Aru jokyū no kotoba.―― Iyada wa. Kon'ya wa naihoku nandesu mono. Chū: Naihoku wa naifu dano fōku dano o arau ban ni ataru koto de aru.


The words of a waitress: "Yuck. It's naihoku tonight."
Note: "Naihoku" refers to it being her turn to wash the knives and forks.

The term "naihoku ナイホク" is not in any of the reference materials at "Japan Knowledge," even with a full-text search. That's a list that includes:

Dejitaru Daijisen デジタル大辞泉 (Digital Daijisen)
Dejitaru Daijisen purasu デジタル大辞泉プラス (Digital Daijisen plus)
Nihon kokugo daijiten 日本国語大辞典 (Dictionary of the Japanese National Language)
Nihon dai hyakkazensho 日本大百科全書 (Encyclopedia Nipponica)

And a Google search gets only about 14 hits. A few are for Akutakagawa's story and the rest are irrelevant (internet handle names, etc.).  Bing and Yahoo! are similarly at a loss.

So far as I know, we don't have something analogous to "naihokuナイホク" in Chinese.  Perhaps the closest we get is dāochā 刀叉 ("knife [and] fork"), but that's pure Chinese for these Western implements.

It's curious that just a few days ago, in reference to David Moser forgetting how to write the right side of chú 厨 ("kitchen"), someone suggested drawing in a knife or fork instead of the misremembered 寸.  See "Pinyin in the kitchen" (10/16/16).

[Thanks to Cecilia Segawa Seigle and Miki Morita]

AND Trump's rhetorical style again

Listening to Donald Trump's 10/14/2016 speech in Charlotte NC, I noticed something that I hadn't noticed in listening to his earlier speeches. He often  uses a loud isolated monosyllable as a way of transitioning between phrases — and perhaps also as a substitute for the filled pauses that he almost never uses. Some of these transitional syllables are particles like and, but, so,yet; some of them are subject pronouns, especially we. These are all words that are usually "cliticized", that is, merged phonologically with a following word — and Trump sometimes pronounces them that way. But here's a sample of his isolated ANDs from the Charlotte speech:

Here's an example in context:

And here's a more normal cliticized and:

I don't recall having heard another public speaker who does this so frequently.

RBG: THOUGHT-raising and r-vocalization

Katy Steinmetz, "How Ruth Bader Ginsburg found her voice", Time Magazine:

For three years, NYU linguistics professor emeritus John Victor Singler, along with researchers Nathan LaFave and Allison Shapp, pored over hours of audio of Ginsburg’s remarks at the Supreme Court. They used computer programs to analyze thousands of vowel and consonant utterances during her time arguing cases in the 1970s, and then from the early ’90s onward, after she returned to the court in robes. While one can hear flecks of classic New York features in Lawyer Ginsburg’s remarks—like the pursed, closed-mouthed vowels—her Brooklyn roots are more obvious in the speech of Justice Ginsburg, they found.

Their theory, reported here for the first time, is that “conscious or not,” the lawyer was doing something everyone does, what is known in linguistics as accommodation: adapting our ways of communicating depending on who we’re talking to. Accommodating can be done through word choice, pronunciation, even gestures. A common example would be when someone returns to the town where they grew up and their accent comes roaring back as they talk to friends and family who sound that way, too.

This is the first time that I can recall having seen embedded Soundcloud audio clips in a publication of this kind.

Steinmetz's Time article is very well done, in my opinion, but I'm not sure what to make of the claim that the theory is "reported here for the first time" — you can read more about this work in Allison Shapp, Nathan LaFave & John Victor Singler, "Ginsburg v. Ginsburg: A Longitudinal Study of Regional Features in a Supreme Court Justice’s Speech", University of Pennsylvania Working Papers in Linguistics 2014. Their abstract:

This study investigates the variable use of New York City (NYC) dialect features by Brooklyn-born Ruth Bader Ginsburg at the Supreme Court, both from her time as a lawyer arguing cases before the Court in the 1970s and as a Justice hearing cases from the bench from 1993 onward. Our data comes from digitized recordings of Supreme Court cases available at The Oyez Project ( The immensity of the Oyez Project’s corpus and its public availability provide us with tokens all along Ginsburg’s timeline at the Court. We look at THOUGHT vowels (N=556) and postvocalic /r/ (N=3304) with reference to their NYC variants, i.e., THOUGHT-raising and r-vocalization. While Ginsburg moved to Washington from NYC in 1980 and has remained there, her data at the endpoint of our study (2011–2012) shows a greater use of NYC vernacular features than was true of the data at the beginning (1972). Mixed-effects regression models using both linguistic and social predictors would seem to point to the importance of chronology for both features: for THOUGHT-raising, the best-fit model makes a binary temporal distinction, between the “Lawyer” years of the 1970’s and the “Justice” years from the 1993 to the 2011 terms. We refer to Communication Accommodation Theory (Giles, N. Coupland and J. Coupland 1991; Giles and Gasiorek 2013) to frame our explanation for what we see as Ginsburg’s reduced use of raised thought in the 1970’s. For r-vocalization, there is again a fundamentally binary distinction, with the year 2000 as the point of division. The forces that motivate this greater use of vocalized-r after 2000 are much less obvious than those behind the Lawyer v. Justice opposition that we propose for THOUGHT-raising. We weigh competing and somewhat contradictory explanations for Ginsburg’s increased use of r-vocalization.

[h/t Cynthia M.]


Pinyin in the kitchen

[This is a guest post by David Moser]

We're in the midst of moving to a new apartment.  Yuck.  So I'm packing boxes with our ayi, who is from Anhui province, and has been helping us with cooking and cleaning house for a few years now.  I think she has at least a middle school education, but probably high school as well.

So we're in the kitchen, and want to mark the cardboard boxes with the kitchen utensils in it.  I hand her the magic marker and ask her to label the boxes with the word chu2fang2.  I notice her holding the marker, waving her hand a bit… and yep, 提笔忘字, she couldn't remember how to write the character.  She slaps her forehead in embarrassment, and sheepishly hands the marker to me.  "Lai," she says, "Ni kending hui xie."  So I grab the marker, and then realize I only remember a few of the components for chu2, "Let's see, a 厂 on top, a 豆 is in there, right?  And something to the right of it…."  Can't quite retrieve it.  So I'm reaching for my cell phone to consult Pleco, but then it occurs to me: Can ayi write it in pinyin?  So I ask her to write it in pinyin instead.  At first she says "Wo de pinyin bu hao!", but insist that she try.  She repeats the first character a few times "chu2…chu2…chu2", and then writes out "chu fng", missing the 'a' in "fang".  When I say "You're missing a letter!" she quickly fills in the missing 'a'.

Typical, I know.  But just add it into the category of "What kind of a writing system would flummox a native speaker of Chinese, and a foreign Ph.D. in Chinese Studies who has been learning for 30 years, in trying to write an incredibly common word like 'kitchen'???"

Given that both us could both read — and write — pinyin, why in the world were we even struggling with the Chinese characters to begin with?


Notes [VHM]:

āyí 阿姨 ("aunt; auntie; aunty; nanny; maid", etc.)

chúfáng 厨房 ("kitchen")

tí bǐ wàng zì 提笔忘字 ("pick up the pen but forget the character[s]"); some people use this for "character amnesia"

"Lai," she says, "Ni kending hui xie." ("Here / C'mon, you certainly can write it".)

Cf. "Dumpling ingredients and character amnesia" (10/18/14).

Bob Dylan's poetry and the Nobel Prize

A. E. STALLINGS says: "At the news that Bob Dylan had won the Nobel Prize in Literature, poets, at least judging from my Facebook feed, were either very much pro- or very much con- (often along generational lines), delighted or outraged…"

I found I fell into neither camp. At first, I was pleased to hear the news, and judged the Nobel committee's view of Dylan to be exactly right: although his early recordings suggest he could hardly win prizes as a singer, guitarist, or harmonica player (don't confuse being strikingly different and new with being highly skilled), he did deserve to be considered seriously as a significant 20th-century poet. So I started with no negative feelings at all about the decision.

And then I looked at some of his lyrics in written form to see if I could find good evidence to cite for this, and found that even my favorite songs looked truly feeble on the page. I responded to some of them when they were originally sung; but looking at them now, I couldn't find anything of high poetic quality at all. And mentally putting them back in their musical context didn't help.

This isn't a strong reaction, and it isn't determined by the generation I belong to: I'm neither too old to have felt the political outrage of the 1960s nor too young to have memories of it. But I would say that, restricting ourselves to the living in accord with Nobel rules, Paul McCartney has written better poetry. Leonard Cohen has written better poetry. Ralph McTell has written better poetry. Don Henley (in collaboration with the late Glenn Frey, who is now far beyond the pain) has written better poetry.

And in particular, Paul Simon has written much better poetry.

So I found (somewhat to my disappointment, because I actually wanted to be pleased by the decision) that I didn't mind at all having the Nobel literature committee remind us that some 20th-century songwriters deserve to be remembered as very fine poets, but, perhaps laying too much emphasis on Dylan's influence on later musicians and political impact on the 1960s and too little on literary skill, they had chosen the wrong one.

Comments on this topic are open below Mark Liberman's post on this topic, so feel free to continue the discussion there.