Morphology and the mental lexicon

1
Morphology and the mental lexicon

• Cogs3000
• Week 6

2
Outline

• Introduction an example
• Morphological decomposition in the mental lexicon
• Psycholinguistic studies of word structure
• Cross-modal semantic priming
• Priming effects of prefixes and suffixes
• Neuroimaging studies of normal and aphasic
  morphological processing.
• An example to start with

3
The following word is ambiguous

• Unbuttonable
• Capable of being unbuttoned.
• Not able to be buttoned.
• How do you assign these respective meanings?
• By alternative morphological parsings of the
  word
• (un(button))able
• un(button(able))
• What does this example demonstrate?

4
Example demonstrates

• That words are composed of smaller elements of
  meaning (morphemes)
• That word meaning assignment is compositional
• That the way in which morphemes are combined can
  be critical for the meaning of the word
• (un(button))able un(button(able))
• Is the difference of meaning between these two
  words simply a function of differences in
  morphological constituency bracketing?

5
Not entirely due to constituent structure

• There are two distinct meanings for un-
• un1- ltnotgt e.g. unattainable
• un2- ltreverse the action ofgt e.g. unfurl, untie,
  unbutton, undo
• What are the conditions on the use of un2-?
  What does it combine with?
• Verbs whose actions are reversible.

6
Morphology the syntax of words

• The unbuttonable example demonstrates the
  compositional route to meaning, but there is also
  the direct route (non-compositional).
• Consider the compositionality of unravel.
• ltreverse the action ofgt ravel ??
• It seems that unravel cannot be treated as
  strictly compositional.
• This raises a basic question for modelling
  lexical processing.

7
Core questions for language processing at the
lexical level

• How word meanings are represented in the mental
  lexicon.
• How lexical meanings are assigned to words in the
  context of sentence processing.
• The precise nature of the items which make up the
  mental lexicon, which we have thus far identified
  as words, but have not attempted to define with
  any precision.
• Lets take the 3rd question first.

8
The question of units words and morphemes
(morphological structure).

• The units of lexical representation, typically
  referred to as lexemes by linguists (Huddleson et
  al., 2002) and as lemmas by psycholinguists
  (Levelt, 1987),
• Are smaller than words, the units conventionally
  separated by white space in printed text.
• e.g., cats and cat 2 words, 1 lexeme

9
Lexical decomposition

• The assumption that in the course of processing
  words for meaning, listeners strip inflectional
  affixes off word forms to access lexical meanings
  (Taft and Forster, 1975).
• How far does this affix stripping extend?

10
There is a cline of decomposability

• cats n plural clearly decomposable
• departed v past
• departed v adj departed guest
• departed derived noun the departed
• not clearly decomposable

11
Formal and semantic compositionality

• The question of compositionality is complicated
  by this distinction.
• Morphology is about the syntax of words.
• Lexical semantics is concerned with the
  compositionality of word meaning.
• We will deal first with the morphological
  structure of words before we discuss word meaning
  assignment.

12
The dual route model

• Mapping from the sound structure of a word to its
  meaning may be achieved compostionally (by rule),
  
• or by directly matching a word-form in the mental
  lexicon (the lexical route).
• The compositional route is more likely in the
  case of a verb like departed.
• the lexical route is the only option for a
  suppletive (completely irregular) form like
  went.
• There has been much controversy in
  psycholinguistic circles over how morphological
  inflections (such as past tense for verbs and
  plurals of nouns) are acquired and to what extent
  regular and irregular forms are processed by the
  same or different mechanisms in language
  comprehension.

13
Disagreement between Linguists and
Psycholinguists

• Over extent of morphological decomposition
• Linguists favour maximal decomposition into root
  morphemes.
• Psycholinguists usually reluctant to pursue
  lexical decomposition beyond recognizing
  inflectional morphemes and the most productive of
  the derivational morphemes.
• Language users, they argue, are not etymologists.
• How do we settle this issue?

14
Classical psycholinguistic investigations

• (Berko, 1957 Derwing, 1976) used elicitation
  techniques to test the productivity of
  morphological rules.
• This is a wug. Now there are two of them. There
  are two _____?
• More recently, repetition priming effects have
  provided evidence that morphological
  decomposition occurs (Marslen-Wilson, et al.,
  1994).

15
Extent of morphological decomposition

• Most agree The inflectional morphology of a word
  is compositional.
• English cats, stronger, rained
• Less agreement as to how far decomposition
  extends to the derivational morphology.
• government is readily decomposable into govern
  ltrulegt ment ltnoun-makergt
• argument, discernment, refinement, amazement,
  arrangement
• Can the paradigm be extended to detriment or
  department?
• How about apartment ??

16
Connectionist challenge to the dual route model.

• Rumelhart and McClelland (1986) set the
  connectionist cat among the psycholinguistic
  pigeons.
• Some pigeons held their ground. The connectionist
  simulations inadequately modelled the course of
  morphology acquisition. (Pinker and Prince,
  1988).
• Connectionists responded by building more
  sophisticated simulations.
• Pigeons again temporarily scattered.
• Today both cats and pigeons occupy the square in
  an unstable misalliance.
• Is there a key to solving this standoff?

17
Morphological regularity and frequency of word
usage.

• Low frequency words which take regular
  inflections.
• Three types of inflection may be distinguished.

What is the rationale for the regularity
frequency trade-off?
18
Rationale for the regularity frequency trade-off

• Low token frequency forms would take too long to
  retrieve
• Irregularly inflected forms are phonologically
  economical (shorter).
• Partially regular forms form families of
  phonologically related members.
• e.g. keep /kip/ - /kept/, leap, creep, sleep
• Regularly inflected forms show no family
  resemblances in terms of phonological similarity.

19
A theory of word representation and processing
needs to

• Account for the way in which the three types of
  morphological inflection are related to token and
  type frequencies.
• Can a single learning mechanism and
  representational system accommodate all three
  types?
• Or do we require a dual-route theory, perhaps
  incorporating a race model?
• Or possibly, three different learning mechanisms
  may be needed.
• Over to you.

20
Psycholinguistic studies of word structureThe
priming paradigm

• Priming the facilitatory effect that
  presentation of a word, can have on the lexical
  retrieval of a subsequent word.
• In a priming experiment, words are usually
  presented to subjects in pairs.
• The first word, called the prime, is typically
  presented shortly before some target or probe
  word.
• which may be related to the prime in some way
  phonologically, semantically, or morphologically.
• Priming effects are usually assessed by measuring
  reaction time to the probe word.
• Thus in a semantic priming experiment, lexical
  access time for the probe word author should be
  shorter, by a few tens of milliseconds when
  preceded by the related prime book, compared
  with an unrelated control probe word such as
  flower

21
Cross-modal priming task

• Subjects heard spoken prime words and respond
  with a lexical decision task to visually
  presented probes flashed onto a computer screen
• The lexical decision task was simply deciding
  whether or not the letter sequence flashed on the
  screen was a real word or a non-word. (In
  approximately half of the experimental trials the
  probe letter sequence was a phonotactically legal
  non-word letter sequence, e.g. glark).
• It was the morphological relation between the
  prime and the probe word, not any phonological
  similarity between prime and probe that was found
  to matter (Marslen-Wilson, Tyler, Waksler and
  Older, 1994)
• When subjects heard the word friendly spoken at
  the same time as they made a lexical decision to
  the visual probe friend, their responses were
  faster by 40 - 60 milliseconds than to an
  unrelated probe, but when tinsel was the
  auditory prime for the probe word tin, no priming
  effect was obtained.

22
Test conditions and morphological priming effects
Only condition 4 where there is a phonological
similarity between prime and probe but no
morphological relationship failed to yield a
priming effect.
What are the effects of semantic similarities
amongst the prime-probe pairs in this experiment?
Can they be separated from effects of
morphological similarity?
23
Morphological and Semantic relatedness priming
effects
There is no evidence here for an independent
effect of priming by virtue of a morphological
relationship between the prime and probe stimuli.
Marslen-Wilson et al. (1994) do not describe
their results this way. They say that only
morphologically related pairs which are
semantically transparent, or perceived to be
semantically related, yield priming effects.
Evidence for an independent role for
morphological structure comes from the way that
semantic transparency was found to interact with
prefixed versus suffix-derived morphological
constructions.
24
Prefixes behave differently from suffixes in
morphological priming.
Confession fails to prime confessor. But
unfasten primes refasten. Why?
25
Why do suffixed words fail to prime each other?

• The words confession and confessor are
  semantically related and this semantic similarity
  should yield a priming effect
• Some countervailing inhibitory effect must be at
  work.
• This can only come from competition between the
  two derivational suffixes -ion and -or.

26
The cohort model holds the key

• The suffixes -ion and -or will always be
  competitors as perceptual targets in word
  recognition,
• when they seek to attach to some verbal stem to
  form a derived noun.
• The target word must be confession, or
  confessor, or conceivably some other alternative.
• But it must be one of these alternatives.

27
Lateral inhibition between competitors

• Lateral inhibition between competitors is
  invoked in TRACE and the Cohort model, to resolve
  competing hypotheses about the identity of an
  input string.
• We need also to explain why prefixes do not
  mutually inhibit one another.
• A prefix typically allows numerous stem
  attachments.

28
Conclusions from the priming study.

• Morphological decomposition was seen to be
  justified, up to a point, on evidence from
  cross-modal semantic priming studies.
• Morphological decomposition is justified insofar
  as the morphological components of a word are
  semantically transparent.
• i.e., to the extent that the meaning of the whole
  word can be clearly related to the meanings of
  its component morphemes.
• e.g., indefensible ltnotgt(ltdefendgt(ltablegt))
• However, we did not provide an explicit account
  of semantic transparency, other than to appeal
  to language users intuitions about the meanings
  of words.
• That is the task of a theory of lexical semantics

29
Towards a theory of lexical semantics

• Cogs3000
• Week 8

30
Goals of a theory of lexical semantics

• A theory of lexical semantics needs not only to
  account for the meaning of individual words but
  how word meanings change in context with other
  words.
• Consider the meaning of good in the phrase good
  friend (ltloyal, reliablegt)
• Now consider the meaning of the same word in the
  phrase good lover or good meal.
• There is a chameleon-like quality to the meaning
  of some words, which needs to be accounted for.

31
Lexical semantics and Pragmatic knowledge

• The meanings of words are said to be stored in
  the mental dictionary,
• along with their phonological and
  morpho-syntactic features.
• Lexical knowledge clearly varies from speaker to
  speaker.
• Languages also vary from one another in how they
  lexicalize our knowledge and perception of the
  world
• Hence, lexical semantics though intimately
  connected with pragmatics (knowledge of the
  world) is distinct from it.
• Lexical knowledge is confined to knowledge of
  word meanings and word usage.

32
The problem of polysemy

• Dictionaries have multiple meanings listed for
  any given word and the more common the word, the
  more meanings it will typically have listed
  against it.
• Meanings of show that are found in the pocket
  Oxford.
• The contexts have been scrambled.

33
Semantic anomaly detection

• Another way we might test the adequacy of a
  theory of lexical semantics.

It is the detection of anomaly in the actual
expression which is of interest to us here,
rather than how the anomalous utterance was
produced.
34
Another test

• Ability of the system to generate acceptable
  paraphrases for individual words or phrasal
  expressions.
• To express commonalities and differences of
  meaning between verbal expressions that are
  semantically related.
• a bottle of wine versus a jug of whiskey

35
A fundamental unresolved problem on how to
approach the analysis of word and phrase meanings

• Whether semantic theory should be grounded in
  logic or in psychology
• in some formalized propositional calculus of
  truth conditions for possible worlds,
• or in perceptual experience and its internal
  cognitive representation.
• The answer to this long standing controversy may
  lie in the level of analysis adopted.
• A propositional calculus is clearly more
  applicable to sentence-level semantic analysis.
• The way we perceive objects in the world is
  clearly of critical interest for the semantic
  properties of words.

36
Semantic Networks

• One of the earliest attempts to construct a
  computational model of lexical semantics was
  undertaken by Ross Quillian (1968)
• A system ambitiously dubbed The teachable
  language comprehender(TLC)
• A program designed to be capable of being taught
  to comprehend English text.
• The TLC envisaged the lexicon as a semantic
  network encoding both world knowledge and the
  meanings of words.

37
The TLC

• Quillians semantic network is a symbolic, not a
  connectionist network.
• The nodes of the network represent concepts or
  words,
• that are linked by arcs, which represent a small
  number of relational types.
• The meaning of a word may be expressed, first by
  accessing its root node in the network and
  traversing the network elements to which this
  node is linked in a series of steps.
• A words meaning is defined by other words, whose
  meanings are defined, in turn, by other words.

38
How the TLC represents the polysemous word plant
These three meanings of plant are disjunctive in
their usage.
The basic meaning of plant (PLANT 1) and pointers
to its alternative meanings (PLANT 2 and PLANT 3)
are expressed in a plane of word memory.
39
Semantic representations in TLC

• Four basic types of arc linking nodes in a
  semantic network
• In addition to the basic logical relations of
  disjunction (OR) and conjunction (AND)
• His system recognized a subordinate-superordinate
  relation, which others have labelled the ISA
  relation and which defines a relationship of
  class membership (e.g., dog - animal).
• A third type of linkage, which may simply be
  thought of as property attribution links one
  node directly to another, declaring that the node
  to which an arrow points is a property of the
  node from which that arrow points.

40
Testing TLC