Whole-word Morphology

Dr. Will Styler - LIGN 120

Today's Plan

• The Storage of Linguistic Experience

• Emergence

• Whole-Word Storage

• The Weaknesses of Whole-Word Storage

Previously, on LIGN 120

How are words being built?

• "Let's see, I need to talk about gluing this thing, again, in the past. That's 'glue', plus 're-', and '-ed'."

• "Let's see, I need to talk about gluing this thing, again, in the past. Looks like 'reglued' is the word I need!"

Let's talk about our reading

Bybee 1998

"In the traditional view, the lexicon is a storage area for all and only the content words or morphemes of a language. The lexicon is relatively static compared to the grammar, which contains all the moving parts of sentence generation: in the metaphor of a dictionary, lexical items are just passive items on a list which wait to be recruited into syntactic structures."

Bybee 1998 Continued

"Rather than arguing that a lexicon of this type does not exist, I am going to argue that if such a lexicon does exist it is because it is emergent from the storage of linguistic experience, which is of a very different nature than the traditional conception of a lexicon would suggest."

"Emergent from the Storage of Linguistic Experience"

• This is the key to understanding these approaches

The Storage of Linguistic Experience

Is storage expensive?

Morpheme-based Lexicons store only maximally analyzed stems and affixes

• Generative approaches to grammar create all and only the attested forms

• Morpheme-based lexicons are great for efficiently handling regular patterns

• They struggle with irregularity and hard-to-memorize forms

• Computation is cheap, but storage is expensive!

In this view, we store the minimum required to do assembly

• Phonemes, which build morphemes, which build words, which build sentences

• Very little storage, lots of computation, language is generated

What about the other end of things?

What if storage was cheap and easy?

We have evidence that we're storing a lot of things

• What was the route from your childhood home to the grocery store?

• The color of your parents first car?

• Your first teacher in grade school's name?

Storing Linguistic Experience

What is Linguistic experience?

• Words, phrases, and sentences

• Individual phonetic realizations of words

• Every single word ever spoken to you

  • Potentially with metadata

Are we storing individual speech tokens?

• Lightfoot 1989 found listeners retained familiarity with unknown talkers' voices after 10 days

• Lightfoot, N. (1990). Effects of familiarity on serial recall for spoken word lists. The Journal of the Acoustical Society of America, 87(S1):S70–S71.

• Goldinger & Azuma 2004 showed that individual tokens affect listener production 6 days after hearing them

  • Goldinger, S. D. and Azuma, T. (2004). Episodic memory reflected in printed word naming. Psychon Bull Rev, 11(4):716–722.

• What does your BFF sound like?

• So, it's not insane to think that we're storing a lot of data

The proposed unit of storage could be acoustic representations of words

Bye bye, Phonemes and Morphemes!

• Words are stored as monolithic acoustic chunks

• We are storing the surface forms

  • We don't need phonemes nor phonological processing

• The smallest unit of storage is the acoustic word, stored with its semantic reference to the real world

'Word' doesn't just mean 'thing bounded by spaces'

• Anything that happens frequently and without variation could be stored

• "Mai Tai", or "For all intents and purposes", "Inflectional Morphology"

• Set phrases could be stored as 'words'

I know what some of you are thinking

"How can such a model account for..."

• The 'felt' reality of phonemes and morphemes?

• Phonological rules?

• Morphological productivity?

• Morphological creativity?

• ... and, uh, Language?

Emergent Grammar

Emergence

When an entity is observed to have properties that its parts do not have on their own

How would you describe this complex pattern?

Langton's Ant

At a white square, turn 90° right, flip the color of the square, move forward one unit. At a black square, turn 90° left, flip the color of the square, move forward one unit

Complexity can emerge despite simple rules

• We don't need elaborate rules to be a part of the system for the system to have elaborate patterns!

Emergence in Language

Let's say we've memorized a lot of words

• We know acoustic forms, and their mappings to semantic meanings

• We are sitting on a huge pile of forms with meanings

• How does morphology work?

We grab the word we need in the moment from the pile

• ... and plug it into the sentence(s)

• No phonology or morphology needed!

• Frequency effects come for free!

  • More on this Wednesday!

... but what about creativity and productivity!

• "How are you going to handle the wugs?"

Analogy

• There are lots of form-meaning correspondences in the words we've memorized

• We can take advantage of those to build new, understandable forms!

Sub-word units don't exist, they emerge

• 'able' is just that thing that is shared among all those words having to do with possibility

• '-en' doesn't occur in many plural forms, that can't be a thing

• Maybe 'cran-' is a type of berry?

• If 'affixed' forms occur often enough, and their 'bare stems' occur often enough, we can come to 'see' the affix as an independent entity

Analogy allows productivity

• "Wug. That's like drug, or hug, or dug, or rug. When I talk about two of those things, it seems to always end with this -z gesture. So, [wʌgz]?"

• Phonological and morphological rules are just formalizations of patterns visible across attested surface forms

  • (which is relatively uncontroversial)

• Grammar comes from statistical patterns in language

  • These are quantifiable and testable!

• This is exactly how reanalysis works, even in a generative world!

So, wrapping back...

"Rather than arguing that a lexicon of this type does not exist, I am going to argue that if such a lexicon does exist it is because it is emergent from the storage of linguistic experience, which is of a very different nature than the traditional conception of a lexicon would suggest."

Usage-based or 'Exemplar' method strengths

• Accounts easily for many phonetic phenomena

• Phonology comes 'for free'

• Frequency-based effects are built in to the model

• Easily quantifiable and testable, given sufficient data

• Handles notions of morphology with less complexity!

• It's complicated, but it works!

Weaknesses of the Whole-Word Lexicon

Reality of Bulk-Storage

This depends on a big assumption about storage

• If we can't store this much, these theories are DOA

• ... but it sure seems like we can

Morphemes sure seem real

Phonological rules can seem to privelege morphemes

• "This is legal across morphemes, but not within"

• "This can only be found in this one suffix"

Morphological rules can seem based on derivation

• Dutch "ge-" only works if there are no other derivational affixes

  • gesproken, but *gebesproken

• Other rules depend on the presence of absence of certain morphemes

  • "Attach this as a prefix on a bare stem, but use this suffix if there's other morphology there"

Morphologically Complex Languages

How do we handle agglutinating languages?

What about polysynthetic languages?

Is the 'word' a sane level of storage when it is often the largest linguistic unit?

• If there's a limited number of affixes, storage is tractable

• Imagine you could add a single noun to a single verb in a language...

  • Your number of stored items increases by [#InflectedVerbForms]*[#Nouns]

'Word' remains tricky

When we store whole words, what is that?

• How do we differentiate words from sentences?

• Idioms?

• Sub-parts?

• 'Word' is tough, here, too.

This is tough for actual analysis

Usage-based methods require huge amounts of data

• Saying anything meaningful about data requires statistical generalizations across that data

• "How often does this form occur relative to that one?"

• "How regularly is this affix attested across the language?"

• Usage-based Linguistics depends on easy access to huge amounts of data

  • This is intractable for language documentation

• You can't 'describe a language' using usage-based methods

  • Just individual phenomena!

'Analogy' turns into generative rules quickly

• Are emergent morphemes so different from stored ones?

• If we compute the grammar as needed, and make reference to that grammar, aren't we back where we started?

• It's very hard to describe big patterns without some sort of generalizations!

Generative linguistics is like picking up broken glass in carpet

• Disposing of the big chunks is easy and effective

• You're gonna spend a long time chasing down those last slivers

• ... and you're still gonna need some tweezers from time to time

Usage-based Linguistics is like picking up a dumptruck full of needles on a smooth concrete floor

• The task is incredibly simple, and 100% effective

• Once you're done with an area, you're done

• ... but it's going to take you forever

• ... and you'll probably want something more efficient for a big room!

Wrapping Up

• Whole-Word storage privileges computation over storage

• It replaces a great deal of complexity with bags of words

• Analogy allows novel forms and productivity

• It has some weaknesses, both with the seeming reality of sub-word chunks, in complex morphology, and in practical description

For Next Time

• We'll talk about frequency as a force in morphology!

• ... which absolutely bears on this discussion

Thank you!