Chapter 7 Summary: Stem Changes vs. Affixes

Chapter 7: Are Stem Changes as Natural as Affixes?

1. One Processing Type or Two?

• Two types of processing have been proposed for morphologically complex words:

  • Traditional Generative Account: Uses symbolic rules in a module separate from the lexicon to act on lexically represented morphemes (Chomsky and Halle 1968; Saddock 1985).

  • Pattern Network Associators: Various versions propose that irregular past-tense verbs are stored in the lexicon and organized into networks (Bybee & Slobin 1982; Bybee & Moder 1983; Bybee 1985, 1988).

• Bybee (1985, 1988) suggests that even regular morphological patterns can be accounted for by pattern network associators, eliminating the need for a separate module for morphological derivation.

• Productivity of morphological processes is determined by type frequency.

• Connectionist models rely on networks of associated patterns (Rumelhart & McClelland 1986; MacWhinney et al. 1989).

• Another proposal suggests that both types of processing are utilized, with irregular formations stored in the lexicon and regular processes handled by symbolic rules (Pinker 1991; Marcus et al. 1992; Prasada and Pinker 1993).

• Regular processes are defined as productive, easily extended to new items.

• Regular processes are often affixal (and agglutinative), while irregular ones involve stem changes or fusion.

• Affixation may be better handled by symbolic rules, and stem changes by lexical connections.

Natural Morphology

• Natural morphology theory (Dressler) posits that affixation is more "natural" than internal stem changes.

• Criteria for "more natural" include:

  1. More common in world languages.

  2. More common within a single language.

  3. Easier to acquire, thus acquired earlier.

  4. Later lost in aphasia.

  5. Easier to process in general (Dressler 1985).

• If affixation correlates with regularity, its naturalness could be due to processing by symbolic rules being preferred over lexical pattern networks (Marcus et al. 1993).

• A model where all morphology is handled in the lexicon via associative pattern networks does not distinguish affixation from stem change.

• Affixes are analyzed via lexical connections, just as stem changes are, as proposed by Bybee (1985, 1988).

• Example:

  • (1a) regular suffix: hap → hapin, hop → hopin, plæn → plænin

  • (1b) semiproductive pattern: string, strung; sting, stung; cling, clung; sling, slung.

• If this model is correct, there should be no preference for affixation over stem change in acquisition or processing.

• However, studies show suffixation is mastered before stem changes (Clark & Berman 1987).

• Stem changes obscure the lexeme's shape and tend to be restricted to lexically arbitrary classes.

• The study in this article attempts to determine whether stem changes per se are more difficult to acquire than affixes.

• The hypothesis is that stem changes are no more difficult than affixes to store and process, but their difficulty arises from being lexically arbitrary and of low type frequency.

Manipulating Artificial Languages

• To determine whether it is stem change itself or other factors that delay acquisition, artificial languages are used where stem change can be made independent of lexical arbitrariness and type frequency.

2. The Diachronic Perspective

• Affixation is common across languages because of the diachronic phenomenon of grammaticization.

• Grammaticization is the process by which lexical morphemes or phrases develop into grammatical morphemes and affixes (Heine and Reh 1984; Bybee 1985; Heine, Claudi, and Hünnemeyer 1991; Traugott and Heine 1991; Bybee, Perkins, and Pagliuca 1994).

• Grammatical affixes primarily originate from free words or phrases.

• Semitic stem changes could have been conditioned by the presence of affixes that are no longer present.

• Grammaticization is ongoing in all languages. Examples include the development of:

  • English regular past-tense suffix -ed (OE -ede) from the preterite form of the verb don 'to do', dyde.

  • The productive affix -ly from the OE word liç meaning 'body'.

• The development of internal stem changes is a related phenomenon that depends entirely on the prior existence of affixes.

• Stem changes tend to develop very late in the grammaticization process.

• The longer an affix is attached to a stem, the greater the phonological fusion between them.

• Affixes can condition changes in the stem. For instance, an old Germanic noun plural suffix -i caused fronting of stem vowels.

• When an affix conditions a change in the stem and is then deleted, the stem change signals the morphological category.

• Example: Development of an English Stem Change for Plural

  • TABLE 7.1. The Development of an English Stem
    Change for Plural

  • Singular

  • Plural

  • Proto-Germanic

  • Proto-Germanic

  • *mūs

  • *mūsi

  • Old English

  • *mūs

  • mūs

  • *mysǝ

  • Middle English

  • mūs

  • mys [müs]

  • Modem English

  • mouse [maus]

  • mice [mais]

• Since grammaticization is continuous and new morphemes constantly arise, most affixes do not last long enough to produce stem changes.

• Grammaticization explains why affixes are common, but stem changes are rare.

• During grammaticization, a free grammatical morpheme can become an affix, involving changes:

  1. The morpheme becomes fixed in its position.

  2. No open-class items can come between the grammatical morpheme and the lexical morpheme.

  3. The combination may be reanalyzed as a single unit (a word) if it constitutes a coherent conceptual unit (Bybee 1985).

  4. Phonological reduction occurs in the grammaticizing morpheme.

  5. Phonological fusion between stem and affix takes place.

Romance Languages

• The development of the future tense in Romance languages (e.g., Spanish, French) illustrates this.

• Latin periphrastic construction:

  • (2) amare habeo

  • love inf aux + 1s

  • 'I have to love, I am to love'

• In Old Spanish:

  • (3) amar he

  • love inf aux + 1s

  • 'I will/shall love'

  • Example:

  • (4) amar lo he

  • love inf him aux + ls

  • 'I will love him'

• Later, the auxiliary becomes a suffix:

  • (5) lo amaré

  • 'I will love him'

• Affixation results from frequent co-occurrence and conceptual coherence, leading speakers to store units together.

• If affixes are represented lexically, there is no qualitative difference between affixes and stem changes.

3. The Experiment

• Affixes are typically used in productive patterns with high type frequency, while stem changes occur in lexically arbitrary classes with low type frequency.

• Artificial languages are created to manipulate variables independently to determine if there is an acquisition or processing advantage to affixes.

• Artificial languages help study linguistic strategies speakers bring to tasks without specific language properties.

3.1. Background

• A study by Bybee, Fox, and Klinck (1987) examined whether affixes were more "natural" than stem changes.

• Natural was defined as the frequency subjects generalized stem changes or suffixes to novel words.

• Subjects were presented with sixteen singular and plural noun pairs in an artificial language.

• Singular forms were single syllables with a CVC form.

• Half used a suffix schwa in the plural, and the other half had an internal vowel change to schwa.

• Both suffix and stem change had equal type frequency and regularity.

• Subjects underwent three learning trials and, by the last trial, performed at 90% accuracy.

Generalization Task

• Subjects were tested on a generalization task with six novel words in singular form.

• Results showed no preference for affixation over stem change.

• Of 120 possible responses, 63 were affixes (52.5%).

• Eight subjects had more affix responses, six had an equal number, and six had more stem-change responses.

• These data suggest no obvious advantage for affixation.

• The experiment did not systematically address the possibility that type frequency determines generalization.

• Both stem change and suffix were regular with equal type frequency.

• The following experiment addresses this issue by manipulating type frequency and type of morphological process while keeping token frequency constant.

Controlled Variables

1. Affixation versus stem change.

2. Phonetic simplicity (Clark and Berman 1987).

3. Lexical arbitrariness.

4. Type frequency.

5. Token frequency.

• Suffix type (regular or irregular) and stem-change type (regular or irregular) were crossed in a 2 × 2 design.

• Subjects received one of four possible combinations:

  1. Suffix regular, stem change regular.

  2. Suffix irregular, stem change irregular.

  3. Suffix irregular, stem change regular.

  4. Suffix regular, stem change irregular.

• This design assesses the effects of regularity and determines if there is an overall preference for one morphological process.

• Subjects were randomly assigned to conditions.

• Dependent variables were percent choices of plural forms.

Predictions of Hypotheses

• The "naturalness" hypothesis predicts subjects should produce suffixes more frequently regardless of type frequencies.

• If type frequency is important, generalization rates should vary: high type frequency results in more generalization, and vice versa.

• Stem changes are expected to be equivalent to suffixes.

4. Method

4.1. Subjects

• Forty students at the University of New Mexico were subjects.

• Subjects were native English speakers.

• The use of English speakers biases the experiment against the result expected.

4.2. Materials

• Two sets of materials matched the learning and generalization phases.

• Sixteen novel words (CVC form) were generated for the learning phase.

• Vowels were used with equal frequency.

• Words were presented as the "singular" form with a line drawing.

• Subjects learned singular and plural forms.

• A generalization task presented ten novel words in singular form.

• The set of words was divided in half, with eight words assigned to the affix condition and eight to the stem-change condition.

• Regularity: regular suffixes and stem changes had only one form of the plural.

• This meant either the addition of [$\partial$] as a suffix or the replacement of the vowel with [$\partial$] for the stem change, for all eight instances of each condition (see table 7.2)

• In irregular conditions, four different plural forms were used.

• For the irregular suffix condition, the four forms used for the affix were [$\partial$] and three other suffixes: [$\partial$m], [$\partial$t], and [ni].

• The irregular plural stem changed to [$\partial$], [u], [1], and [ou].

• Stimuli Example:

  • TABLE 7.2. Example Stimuli for Stem-Change and
    Affixation Conditions

  • Singular

  • Plural

  • Туре

  • Picture

  • faut

  • ->

  • fǝt

  • stem change

  • girl(s)

  • dit

  • ->

  • ditǝ

  • suffix

  • flower(s)

• Example Stimuli Used:

  • TABLE 7.3. Stimuli Used in the Learning
    Phase of the Experiment

  • Plural form

  • Regular

  • Irregular

  • Stem change

  • waiš

  • wǝš

  • wǝš

  • zib

  • zǝb

  • zǝb

  • raik

  • rǝk

  • ruk

  • sauf

  • sǝf

  • suf

  • vaum

  • vǝm

  • faut

  • fǝt

  • voum

  • fout

  • jik

  • jǝk

  • JIK

  • naid

  • nǝd

  • nid

  • Affix

  • hauk

  • haukǝ

  • haukǝ

  • mir

  • mirə

  • mirǝ

  • sail

  • sailǝ

  • nauz

  • jaib

  • dit

  • ditǝ

  • nauzǝ

  • jaibǝ

  • sailǝm

  • nauzǝm

  • jaibni

  • ditni

  • vaud

  • faiš

  • vaudǝ

  • faišǝ

  • vaudǝt

  • faušǝt

Language Master

• Words were recorded by a female native English speaker using a Bell and Howell Language Master.

• Pictures of referents were copied onto the card.

• Items were recorded on separate cards to shuffle stimuli.

• Ten additional novel words were generated for generalization:

  • (6) Novel items used in the generalization phase of the experiment

  • biš

  • waub

  • zaud

  • šaim

  • lauj

  • kaij

  • tis

  • zaif

  • vin

  • raus

• There were no consistent cues from the structure of the CVC.

4.3. Procedure

4.3.1. Learning phase

• Subjects learned singular and plural forms from a novel language.

• The set of sixteen words was randomized and divided into four sets of four items each.

• On the initial trial, subjects heard singular and plural forms with pictures.

• They were then presented with the singular form only and asked to generate the plural form.

• Responses were recorded and corrected.

Practice Lists

• This process was repeated for three recall trials.

• Following the third iteration, the subject was presented with the second set of four items (in both singular and plural form), and the same procedure, with three recall trials, was followed.

• The initial phase of the learning task was repeated for the remaining two sets of four items in the list.

• For purposes of analysis, each of the three recall trials is referred to as a practice list (numbered from 1 to 3), and the results will be reported across all sixteen words for each list.

• Subjects were presented with the full list of sixteen singular items and instructed to generate the correct plural for each.

• Feedback was given. The final learning trial consisted of the same sixteen singular items.

• Final learning with no feedback

4.3.2. Generalization phase

• Subjects were given the new list of ten novel words with pictures.

• They supplied what they considered appropriate plural forms without feedback.

4.4. Results

4.4.1. Learning phase

• Analysis was conducted on the mean percent correct recall across conditions.

• The recall data were examined separately for the three practice lists and the two final, full lists.

• The four conditions differed in initial learnability (see figure 7.1).

Observations

• The regular conditions led to higher initial learnability.

• The irregular conditions led to the lowest initial performance.

• The effect of regularity was present for both suffixes and stem changes.

• There was a significant main effect of suffix regularity, $F(1,36) = 5.941$, $p < .02$, $MSe = 2065$

• A marginally significant main effect of stem-change regularity was determined, $F(1,36) = 4.054$, $p < .052$, $MSe = 1409$

• The interaction of suffix regularity stem-change regularity was not significant, $F < 1$.

• The initial difference was a significant difference between condition 1 (suffix and stem change both regular) and condition 2 (suffix and stem change both irregular).

• Condition 2 led to considerably fewer correct responses (condition 2: 46%, condition 1: 72%), two-tailed $t(18) = 3.31$, $p < .004$.

• Condition 1 was also marginally different from condition 3, suffix irregular, stem regular (59%): two-tailed $t(18) = 2.03$, $p < .058$.

• Performance on the remaining practice lists quickly reached near ceiling levels.

• There was a marginally significant effect of stem-change regularity on practice list 2, $F(1,36) = 4.020$, $p < .053$, $MSe = 612$.

• No other effects approached significance for either practice list 2 or 3.

• Mean percent correct on practice list 3, collapsed across conditions, was 94%.

Full List Analaysis

• Analyses were also conducted on the percent correct scores for the two full lists.

• The analyses on the data for full list 1 revealed that there was a marginally significant interaction between suffix and stem-change regularity, $F(1,36) = 3.239$, $p<.08$, $MSe = 712$

• A marginally significant effect of suffix regularity, $F(1,36) = 3.732$, $P < .061$, $MSe = 821$

• Irregular suffixes in the context of a regular stem change lowered relative performance considerably more than did irregular stem changes in the context of regular suffixes

• The suffix-regularity effect is shown in the graph by the fact that the mean of the two suffix-regular conditions (both regular, and suffix regular, stem irregular) is higher than the mean of the two suffix irregular conditions (both irregular, and suffix irregular, stem regular).

• Performance on full list 2 was quite high (82%).

• There were no significant differences in the rate of correct responses across the four conditions: suffix regularity × stem-change regularity, $F<1$

• Suffix regularity, $F<1$

• Stem-change regularity, $F(1,36) = 1.656$, $p < .206$, $MSe = 389$.

• The learning data indicate that there is no inherent difference in difficulty in acquiring stem changes rather than suffixes.

• Although there were initial differences in percent correct as a function of regularity for both suffixes and stem changes, these quickly disappeared.

4.4.2. Generalization data

• The rate of suffix generalization was 48.75%, and stem-change generalization was 50.25%.

• Other responses included the use of an English suffix (.25%) and no change to the singular form (.75%).

• Overall, stem changes and suffixes are equally learnable.

• Regularity was manipulated to explore its effects.

• The regularity of stem-change had no effect on the use of suffixes or stem changes.

• Suffix regularity did have an effect.

• The rate of suffix use was higher than stem-change use when the suffix input in the learning phase was regular, but lower than stem-change use when the suffix input was irregular.

• Results are detailed in figure 7.3.

• A 2 (suffix regularity) × 2 (stem-change regularity) analysis of variance was conducted on the mean percent suffix choices.

• There was a main effect of suffix regularity, $F(1,36) = 5.009$, $p< .031$, $MSe = 1562$.

• Neither the main effect of stem-change regularity nor the interaction was significant, both $Fs < 1$.

• Suffix regularity determines the rate of suffix and stem-change generalization.

• Acquisition of a stem change is not inherently more difficult than adding a suffix.

• In two conditions with irregular suffixes, subjects chose the stem change more frequently.

• If affixation were more “natural,” there would be a greater advantage for suffixes.

5. Discussion

• Stem changes can be learned as easily as affixes.

• Stem changes seem difficult because they are lexically arbitrary and have low type frequency.

• The results showed no overall advantage of suffixes over stem changes but there was an effect of suffix irregularity or type frequency.

• Irregularity in both affix and stem-change conditions led to initial difficulties.

• Regularity of suffixes led to greater use of suffixes, while irregularity of suffixes led to greater use of stem changes.

• Regularity affects the ease of generalization of the suffix.

• Under the right circumstances, stem changes can be acquired as easily as, or even more easily than, suffixes.

• Subjects were basically involved in a rote-learning task, setting up lexical representations for this fragment of an artificial language.

More notes

• For the conditions in which both processes were regular or both were irregular, responses would be approximately evenly divided among suffixes and stem changes.

• When both processes were regular, subjects used significantly more suffixes, but when they were both irregular, subjects chose significantly more stem changes.

• There is no overall suffixation preference.

• Prior experience with natural language(s), explains the fact stem change to be associated with irregularity and for affixes to be associated with regularity.

• English speakers should have been predisposed to use suffixes.

• One possible explanation of this result is that they were speakers of English.

• The best way to avoid this inherent bias would be to use as subjects speakers of an isolating language such as Mandarin Chinese.

• It is clear that there should be no overall advantage for suffixes over stem changes as, in our view, neither form is more "natural" than the other.

• Even in such a model, it would still be reasonable to suppose that morphemes that consist of continuous phonological material, rather than interrupted material (as in sing, sang, sung, where the root is s-ng), have an acquisitional or processing advantage (Dressler 1985).

• Our experiments turned up no evidence of such a tendency.

• The suggested association of concatenation with regularity and stem change with irregularity (Marcus et al. 1993) was not supported in our experiment, where we demonstrated that the English speakers behave as though these are separate variables.

• Only certain avenues of development are available, and these lead to certain structures.

• It may only mean that these structures arise more easily and more frequently than others

• The fact that affixes are more widespread and common than stem changes has a straight-forward diachronic explanation.