Medical bills

Starting about two years ago, I got an unexpected medical bill in the mail. The amount wasn’t very high, but I was quite frustrated and annoyed. First, this was from a local College of Dentistry, where most procedures are free for the insured (and probably not insured too); there was no “explanation of benefits” that explained this was a co-pay, or that my insurance only covered some portion. Secondly, I hadn’t been to the College of Dentistry in quite a while, so I had no idea which of the various procedures this was or even what day I received the billed service. Third, there was no way to get more information: the absolute worst thing about this provider is that the administrative staff are some of the most overloaded and overworked people I have ever seen, and I have witnessed them just let the phone ring because they’re dealing with a huge line of in-person patients (some of whom are bleeding from their mouth). So I didn’t pay it. After a while though, the bills continued and I started to worry. Was I wasting paper for no reason? Would this harm my credit score? So I put about an hour into finding a way to actually get in touch with the billing office: turns out this was a Google Form buried somewhere on a website, and if you fill it out, a someone calls you (in my case, within the hour!), looks up your chart, and can tell you the date of service and why you were billed. Why they didn’t just include this in the bill in the first place? I have to imagine this makes it ever harder for the College to actually collect on these debts.

Representation vs. explanation?

I have often wondered whether detailed representational formalism is somehow in conflict with genuine explanation in linguistics. I have been tangentially involved in the cottage industry that is applying the Tolerance Principle (Yang 2005, 2016) to linguistic phenomena, most notably morphological defectivity. In our paper on the subject (Gorman & Yang 2019), we are admittedly somewhat nonchalant about the representations in question, a nonchalance which is, frankly, characteristic of this microgenre.

In my opinion, however, our treatment of Polish defectivity is representationally elegant. (See here for a summary of the data.) In this language, fused case/number suffixes show suppletion based on the gender—in the masculine, animacy—of the stem, and there is lexically conditioned suppletion between -a and -u, the two allomorphs of the gen.sg. for masculine inanimate nouns. To derive defectivity, all we need to show is that Tolerance predicts that, in the masculine inanimate, there is no default suffix to realize the gen.sg. If there are two realization rules in competition, we can implement this by making both of them lexically conditioned, and leaving nouns which are defective in the gen.sg. off both lexical “lists”. We can even imagine, in theories with late insertion, that the grammatical crash is the result of uninterpretable gen.sg. features which are, in defective nouns, still present at LF.1

It is useful to contrast this with our less-elegatn treatment of Spanish defectivity in the same paper. (See here for a summary of the data.) There we assume that there is some kind of grammatical competition for verbal stems between the rules that might be summarized as “diphthongize a stem vowel when stresssed” and “do not change”. We group the two types of diphthongization (o to ue [we] and e to ie [je]) as a single change, even though it is not trivial to make these into a single change.2 This much at least has a venerable precedent, but what does it mean to treat diphthongization as a rule in the first place? The same tradition tends to treat the propensity to diphthongize as a phonological (i.e., perhaps via underspecification or prespecification, à la Harris 1985) or morphophonological property of the stem (a lexical diacritic à la Harris 1969, or competition between pseudo-suppletive stems à la Bermúdez-Otero 2013), and the phonological contents of a stem is presumably stored in the lexicon, and not generated by any sort of rule.3 Rather, our Tolerance analysis seems to imply we have thrown in our lots with Albright and colleagues (Albright et al. 2001, Albright 2003) and Bybee & Pardo (1981), who analyze diphthongization as a purely phonological rule depending solely on the surface shape of the stem. This is despite the fact that we are bitterly critical of these authors for other reasons4 and I would have preferred—aesthetically at least—to adopt an analysis where diphthongization is a latent property of particular stems.

At this point, I could say, perhaps, that the data—combined with our theoretical conception of the stem inventory portion of the lexicon as a non-generative system—is trying to tell me something about Spanish diphthongization, namely that Albright, Bybee, and colleagues are onto something, representationally speaking. But, compared with our analysis of Polish, it is not clear how these surface-oriented theories of diphthongization might generate grammatical crash. Abstracting from the details, Albright (2003) imagines that there are a series of competing rules for diphthongization, whose “strength” derives from the number of exemplars they cover. In his theory, the “best” rule can fail to apply if its strength is too low, but he does not propose any particular threshold and as we show in our paper, his notion of strength is poorly correlated with the actual gaps. Is it possible our analysis is onto something if Albright, Bybee, and colleagues are wrong about the representational basis for Spanish diphthongization?

Endnotes

  1. This case may still be a problem for Optimality Theory-style approaches to morphology, since Gen must produce some surface form.
  2. I don’t have the citation in front of me right now, but I believe J. Harris originally proposed that the two forms of diphthongization can be united insofar as both of them can be modeled as insertion of e triggering glide formation of the preceding mid vowel.
  3. For the same reason, I don’t understand what morpheme structure constraints are supposed to do exactly. Imagine, fancifully, that you had a mini-stroke and the lesion it caused damaged your grammar’s morpheme structure rule #3. How would anyone know? Presumably, you don’t have any lexical entries which violate MSC #3, and adults generally does not make up new lexical entries for the heck of it.
  4. These have to do with what we perceive as the poor quality of their experimental evidence, to be fair, not their analyses.

References

Albright, A., Andrade, A., and Hayes, B. 2001. Segmental environments of Spanish diphthongization. UCLA Working Papers in Linguistics 7: 117-151.
Albright, A. 2003. A quantitative study of Spanish paradigm gaps. In Proceedings of the 22th West Coast Conference on Formal Linguistics, pages 1-14.
Bermúdez-Otero, R. The Spanish lexicon stores stems with theme vowels, not roots with inflectional class features. Probus 25: 3-103.
Bybee, J. L. and Pardo, E. 1981. On lexical and morphological conditioning of alternations: a nonce-prob e experiment with Spanish verbs. Linguistics 19: 937-968.
Gorman,. K. and Yang, C. 2019. When nobody wins. In F. Rainer, F. Gardani, H. C. Luschützky and W. U. Dressler (ed.), Competition in Inflection and Word Formation, pages 169-193. Springer.
Harris, J. W. 1969. Spanish Phonology. MIT Press.
Harris, J. W. 1985. Spanish diphthongisation and stress: a paradox resolved. Phonology 2: 31-45.

Automatic batch sizing

Yoyodyne is my lab’s sequence-to-sequence library, intended to be a replacement for Fairseq, which is (essentially) abandonware. One matter of urgency for me in building Yoyodyne was to enable automatic hyperparameter tuning. This was accomplished by logging results to Weights & Biases (W&B). We can perform a random or Bayesian hyperparameter sweep using a “grid” specified via a YAML file, monitor progress on the W&B website, or even hit the API to grab the best hyperparameters. One issue that kept coming up, however, is that it is easy to hit out-of-memory (OOM) errors during this process. Here’s what we did about it:

OOMs are not purely due to model size: the model, batch, and gradients all need to fit into the same VRAM. PyTorch Lightning, which is a key part of the Yoyodyne backend, provides a function for automatically determining the maximum batch size that will not trigger an OOM. Basically, it works by starting with a low batch size (by default, 2), randomly drawing three batches of that size, and then attempting training (but in fact caching parameters so that no real training occurs). If this does not trigger an OOM, it doubles the batch size, and so on.1,2 You can enable this approach in Yoyodyne using the flag --find_batch_size max. You’d want to use this if you believe that a giant batch size is fine and you just want to fully saturate your GPU.

A slightly more sophisticated version of this, useful when you actually want to tune batch size, is enabled with the flag --find_batch_size opt. This again begins by doubling the size of randomly drawn batches as well, but here it halts once the doubling exceeds the value of the --batch_sizeflag. If the max batch size is larger than the requested size, it is used as is; thus this acts as a soft check against OOMs. If, however, the max batch size is smaller than --batch_size it instead solves for a new batch size, the largest batch size which is smaller than the max and which is a divisor of --batch_size`. It then enables multiple rounds of gradient accumulation per update,3 thus perfectly-losslessly simulating the desired batch size while using as much of VRAM as possible. I can assure you this is a killer feature for neural network tuning.

Endnotes

  1. This is a little imprecise, and one can refine it by doing a binary search, but in practice it’s not worth the effort when working with ragged data.
  2. Whatever batch size was requested with the --batch_size flag is ignored.
  3. More formally, given desired batch size $b$ and a max batch size $n’$, it finds $a, n$ such that $a$ is the smallest integer, and $n$ is the largest integer, where $an = b$. This is computed via brute force; my implementation of an elegant solution based on the prime factorization was a bit slower.

An interesting semantic change: “raw dogging”

The term raw-dogging is a slightly-obscene, slangy term for engaging in unprotected sex, often used to celebrate that occasionally-risky behavior. However, this term has undergone an interesting semantic change in the last five or so years. I think the actuator of this chain of events is prolific Twitter user @jaboukie:

This is a straightforward, jocular, semantic extension, generalizing the sense of danger associated with unprotected sex to life itself. In its wake (it was a very popular tweet), I also saw a tweet about “raw dogging” to refer to riding the subway without headphones or sunglasses. Years later, I read a blind item about a US senator flying commercially from the States to Israel; apparently, according to his seat mate, during the long flight, he didn’t listen to music or podcasts, read, check email, nap, or watch a movie, he just…sat there, for hours and hours, like an absolute maniac. I haven’t been able to find this story, and I don’t remember whether it referred to raw-dogging, but I have since seen several stories discussing raw-dogging flights (e.g., this recent one in GQ). Discussions of raw-dogging in the commercial aviation sense largely recognize the act’s covert prestige: it is recognized as a curious and difficult task, one associated with macho and/or maleness. The GQ article also quotes individuals who refer to stimulation-free commercial flying as barebacking, which traditionally refers to unprotected anal sex between men. (In contrast raw-dogging in its original sense does not specify the specific sex act beyond some form of genital-genital penetration, nor does it specify the gender or sexual orientation of the participants.)

“Indic” considered harmful

Indic is an adjective referring to the Indo-Aryan languages such as Hindi-Urdu or Bengali. These languages are spoken mostly in the northern parts of India, as well as in Bangladesh, Pakistan, Sri Lanka, Nepal, and the Maldives. This term can be confusing, because hundreds of millions of people in the Indian subcontinent (and nearby island nations) speak non-Indic first languages: over 250 million people, particularly in the south of India and the north of Sri Lanka, speak Dravidian languages, which include Malayalam, Tamil, and Telugu. Austronesian, Tibeto-Burman, and Tai-Kadai languages, and many language isolates, are also spoken in the India and the other nations of subcontinent, as is English (and French, and Portuguese). Unfortunately, there is now a trend to use Indic to mean ‘languages of the subcontinent’. See here for a prominent example. This is a new sense for Indic, and while there is probably a need for such a lexeme to express the notion (language of India or subcontinental language would work), reusing Indic, which already has a distinct and well-established sense, just adds unnecessary confusion.

A minor syntactic innovation in English: “BE crazy”

I recently became aware of an English syntactic construction I hadn’t noticed before. It involves the predicate BE crazy, which itself is nothing new, but here the subject of that predicate is, essentially, quoted speech from a second party. I myself am apparently a user of this variant. For example, a friend told me of someone who describes themselves (on an online dating platform) as someone who …likes travel and darts, and I responded, simply, Likes darts is crazy. That is to say, I am making some kind of assertion that the description “likes darts”, or perhaps the speech act of describing oneself as such, is itself a bit odd. Now in this case, the subject is simply the quotation (with the travel and part elided), and while this forms a constituent, a tensed VP, we don’t normally accept them as the subject of predicates. And I suspect constituenthood is not even required. So this is distinct from the ordinary use of BE crazy with a nominal subject.

I suspect, though I do not have the means to prove, this is a relatively recent innovation; I hear it from my peers (i.e., those of similar age, not my colleagues at work, who may be older) and students, but not often elsewhere. I also initially thought it might be associated with the Mid-Atlantic but I am no longer so sure.

Your thoughts are welcome.

Vibe check: EACL 2024

I was honored to be able to attend EACL 2024 in Malta last month. The following is a brief, opinionated “vibe check” on NLP based on my experiences there. I had never been to an EACL, but it appealed to me because I’ve always respected the European speech & language processing community’s greater interest in multilingualism compared to what I’m familiar with in the US. And, because when or why else would I get to see Malta? The scale of EACL is a little more manageable than what I’m used to, and I was able to take in nearly every session and keynote. Beyond that, there wasn’t much difference. Here are some trends I noticed.

We’re doing prompt engineering, but we’re not happy about it

It’s hard to get a research paper out of prompt engineering. There really isn’t much to report, except the prompts used and the evaluation results. And, there doesn’t seem to be the slightest theory about how one ought to design a prompt, suggesting that the engineering part of the term is doing a lot of work. So, while I did see some papers (to be fair, mostly student posters) about prompt engineering, the interesting ones actually compared prompting against a custom-built solution.

There’s plenty of headroom for older technologies

I was struck by one of the demonstration papers, which was using fine-tuned BERT for the actual user-facing behaviors, but an SVM or some other type of simple linear model trained on the same data to provide “explanability”. I was also struck by the many papers I saw in which fine-tuned BERT or some other kind of custom-built solution outperformed prompting.

Architectural engineering is dead for now

I really enjoy learning about new “architectures”, i.e., ways to frame speech and language processing problems as a neural network. Unfortunately, I didn’t learn about any new ones this year. I honestly think the way forward, in the long term, will be to identify and eliminate the less-principled parts of our modeling strategies, and replace them with “neat”, perhaps even proof-theoretic, solutions, but I’m sad to say this is not a robust area.

Massive multilingualism needs new application areas

In the first half of Hinrich Schütze’s keynote, he discussed a massively multilingual study covering 1,500 languages in all. That itself is quite impressive. However, I was less impressed with the tasks targeted. One was an LM-based task (predicting the next word, or perhaps a masked word), evaluated with “pseudo-perplexity”. I’m not sure what pseudo-perplexity is but real perplexity isn’t good for much. The other task was predicting, for each verse from the Bible, the appropriate topic code; these topics are things like “recommendation”, “sin”, “grace”, or “violence”. Doing some kind of semantic prediction, at the verse/sentence level, at such scale might be interesting, but this particular instantiation seems to me to be of no use to anyone, and as I understand it, the labels were projected from those given by English annotators, which makes the task less interesting. Let me be clear, I am not calling out Prof. Schütze, for whom I have great respect—and the second half of his talk was very impressive—but I challenge researchers working at massively multilingual scale to think of tasks really worth doing!

We’ve always been at war with Eurasia

I saw at least two pro-Ukraine papers, both focused on the media environment (e.g., propaganda detection). I also saw a paper about media laws in Taiwan that raised some ethical concerns for me. It seems this may be one of those countries where truth is not a defense against charges of libel, and the application was helping the police enforce that illiberal policy. However, I am not at all knowledgeable about the political situation there and found their task explanation somewhat hard to follow, presumably because of my Taiwanese political illiteracy.

My papers

Adam Wiemerslage presented a paper coauthored with me and Katharina von der Wense in which we propose model-agnostic metrics for measuring hyperparameter sensitivity, the first of their kind. We then use these metrics to show that, at least for the character-scale transduction problems we study (e.g., grapheme-to-phoneme conversion and morphological generation), LSTMs really are less hyperparameter-sensitive than transformers, not to mention more accurate when properly tuned. (Our tuned LSTMs turn in SOTA performance on most of the languages and tasks.) I thought this was a very neat paper, but it didn’t get much burn from the audience either.

I presented a paper coauthored with Cyril Allauzen describing a new algorithm for shortest-string decoding that makes fewer assumptions. Indeed, it allows one for the first time to efficiently decode traditional weighted finite automata trained with expectation maximization (EM). This was exciting to me because this is a problem that has bedeviled me for over 15 years now when I first noticed the conceptual gap. <whine>The experience getting this to press was a great frustration to me, however. It was first desk-rejected at a conference on grammatical inference (i.e., people who study things like formal language learning) on the grounds that it was too applied. On the other hand, the editors at TACL desk-rejected a draft of the paper on the grounds that no one does EM anymore, and didn’t respond when I pointed out that there were in fact two papers in the ACL 2023 main session about EM. So we submitted it to ARR. The first round of reviews were not much more encouraging. It was clear that these reviewers did not understand the important distinction between the shortest path and shortest string, even though the paper was almost completely self-contained, and were perhaps annoyed at being asked to read mathematics (even if it’s all basic algebra).  One reviewer even dared to asked why one would bother, as we do, to prove that our algorithm is correct! To the area chair’s credit, they found better reviewers for the second round, and to those reviewers’ credits, they helped us improve the quality of the paper. However, the first question I got in the talk was basically a heckler asking why I’d bother to submit this kind of work to an ACL venue. Seriously though, where else should I have submitted it? It’s sound work.</whine>

“Segmented languages”

In a recent paper (Gorman & Sproat 2023), we complain about conflation of writing systems with the languages they are used to write, highlighting the nonsense underlying common expressions like “right-to-left language”, “syllabic language” or “ideographic” language found in the literature. Thus we were surprised to find the following:

Four segmented languages (Mandarin, Japanese, Korean and Thai) report character error rate (CER), instead of WER… (Gemini Team 2024:18)

Since the most salient feature of the writing systems used to write Mandarin, Japanese, Korean, and Thai is the absence of segmentation information (e.g., whitespace used to indicate word boundaries), presumably the authors mean to say that the data they are using has already been pre-segmented (by some unspecified means). But this is not a property of these languages, but rather of the available data.

[h/t: Richard Sproat]

References

Gemini Team. 2023. Gemini: A family of highly capable multimodal models. arXiv preprint 2312.11805. URL: https://arxiv.org/abs/2312.11805.

Gorman, K. and Sproat, R.. 2023. Myths about writing systems in speech & language technology. In Proceedings of the Workshop on Computation and Written Language, pages 1-5.