Modeling the Latin honor analogy with a computational learner

Adam Albright
MIT Dept. of Linguistics and Philosophy


Related files:



Paradigm leveling has long been considered a sporadic and irregular type of language change that can be described by general tendencies (Kurylowicz 1947, Manczak 1958, Hock 1991), but can not be predicted or quantified. Explicit constraints on Paradigm Uniformity or Uniform Exponence in OT (Kenstowicz 1995, Steriade 1996, Kenstowicz 1999) have brought phonological theory one step closer to a formal description of paradigm leveling, but do little to attack the basic problem of predicting when leveling will occur, which alternations will be leveled, and in which direction. Consider, for example, the leveling in pre-classical Latin of s~r alternations in noun paradigms:

(1) Latin hono:s > honor 'honor'

case pre-leveling post-leveling
nom. [hono:s] [honor]
gen. [hono:ris] [hono:ris]
dat. [hono:ri:] [hono:ri:]
acc. [hono:rem] [hono:rem]
abl. [hono:re] [hono:re]

Kenstowicz (1995) proposes that a Uniform Exponence constraint for nouns, together with a markedness constraint against intervocalic [s], pushed Latin towards uniform paradigms with [r] throughout. The change was somewhat more complicated than this, however, and a uniform exponence analysis leaves many questions unanswered. First, leveling was restricted primarily to masculine and feminine polysyllabic nouns (Hock 1991; Kiparsky 1997); however, it is not obvious why uniform exponence should have been limited in this way. In addition, the older s~r alternation was replaced by a vowel length alternation ([o] ~ [o:]) (Kiparsky 1997; Hale, Kissock and Reiss 1998; Baldi 1999), and morphophonemic alternations in many other words were left intact (e.g., [ars] ~ [artis] 'art', [urps] ~ [urbis] 'city'). Thus, if [hono:s] changed to [honor] to satisfy a uniform exponence constraint, this constraint must have cared specifically about uniformity of [s]'s and [r]'s, and not vowel length, stop voicing, or the presence or absence of [t]. Finally, and most importantly, the uniform exponence analysis does not explain why the change took place in the way it did. Why did a uniform exponence constraint suddenly get promoted to outrank markedness and IO-faithfulness constraints? What allowed the nominative to get rebuilt in Latin, contrary to the usual direction of leveling? And why do languages sometimes move in the opposite direction, extending alternations and making paradigms less uniform?

In this paper, I suggest that paradigm leveling is motivated by more than simply a Humboldtian preference for non-alternation within the phonology. I propose instead that leveling has its roots in the way that paradigms are projected morphologically, and results when non-alternation is the strongly dominant ("default") morphological pattern in the language. I present an computationally implemented model of morphological acquisition that chooses base forms and learns to project paradigms by means of stochastic morphological rules. The base selection mechanism allows us to make predictions about which forms in the paradigm should serve as the pivot for analogical changes. The model also provides estimates of the strength of different patterns in the language, allowing us to make quantitative predictions about which words should come under pressure to change, and which morphological patterns should be extended.

The morphological acquisition model employed here proceeds in two stages: first, a base is chosen by evaluating each slot in the paradigm for how "effective" it is in predicting the remainder of the paradigm. A form is considered an effective base when it allows the remaining forms to be derived confidently and accurately, and when the resulting grammar does not generate large numbers of competing possibilities. Conversely, a form is a bad base when it leaves open many possibilities for derived forms, or when its predictions are wrong or uncertain. For example, in the hypothetical language in (2), it is difficult to predict the genitive form given the nominative:

(2) Neutralization in the nominative

nom.   gen.
[gluptus] ~ [glupti:]
[nokus] ~ [noki:]
[reptus] ~ [reptoris]
[kortus] ~ [kortoris]

In this language, the novel word [tulpus] has a 50% chance of having the genitive form [tulpi], and a 50% chance of [tulporis]. On the other hand, it is easy to predict the nominative given the genitive: new genitives [tulpi] and [pulkoris] could only have nominatives [tulpus] and [pulkus], respectively, with 100% certainty. In general, slots in the paradigm that suffer from neutralizations make bad bases, because one must guess about unpredictable information in order to project the remainder of the paradigm. In the case of Latin, the nominative suffers from more neutralizations than any other form in the paradigm, and is the worst choice of base according to several different criteria (avg. score = 1 out of 6). Among the oblique forms, the ablative is overall the best choice of base (avg. score = 5.4) , following by the genitive (avg. score = 4.6).

Once a base has been identified, the model constructs a grammar of stochastic rules to produce the rest of the paradigm, using the minimal generalization algorithm developed by Albright and Hayes (1998). This algorithm explores the lexicon, keeping track of which morphological patterns are statistically most reliable in different phonological environments. The model was trained on a database of 1,700 Latin nouns in their pre-leveling forms, and was then used to produce nominative forms for all nouns with [-o(:)r] in the genitive. Among the outputs, three qualitatively distinct patterns could be distinguished: for agentives (all non-neuter, polysyllabic, with stem-final [t] or [s]) -or nominatives were favored correctly over -o:s nominatives (.98 vs. .23). For other non-neuter polysyllables, such as [honoris], -or was somewhat favored over -o:s (.75 vs. .60), despite the presence of -o:s forms in the training set. In other words, the model predicts leveling for this particular class of words. Finally, -o:s remained favored to a moderate extent for polysyllabic neuters (.66 vs .39), and strongly for monosyllables (.71 vs. .132).

A surprising aspect of this model is that it predicts paradigm leveling without any explicit notion of paradigm uniformity to motivate the change. The preference for [r] forms comes solely from the fact that non-alternation was the dominant pattern in the Latin lexicon, and words like [hono:s] were brought into line with the rest of the vocabulary. An important question for this model, then, is why there should be a universal tendency for paradigms to become more uniform, if such changes are driven only by lexical statistics. I suggest that the reason for this is that morphophonemic alternations generally affect only a small subset of the phonemes in a language; therefore, even before any leveling takes place, paradigmatic alternations tend to affect only a minority of lexical items. A prediction of this model, however, is that if an alternation does happen to affect the majority of lexical items, then the alternation should be extended just as easily as uniformity is extended in other cases. In fact, this prediction does seem to be true; as an example, I offer evidence from an analogical change that is currently underway in Korean, in which a [t¬] ~ [s] alternation is being generalized to words with less radical alternations (such as [t˺] ~ [t], [t˺] ~ [tʰ] and [t˺] ~ [tʃʰ]).

N.B. The IPA symbols at the end of the last paragraph use Unicode! If you do not see a schwa here: [ə], your browser is not configured correctly to use a unicode font that can display IPA characters. For more info, click here.

Last updated 9/24/01