Lexicon Optimization Research Articles

What can children’s spelling tell us about underlying representations?

This paper examines the written productions of first and second grade Catalan children in an attempt to tap into their understanding of the underlying forms of Catalan unstressed vowels for which alternations are not present (and for which multiple inputs are thus possible). The paper also explores the differences in spelling between alternating and non-alternating [ə] and [u], in order to find out whether morphophonemic alternations are taken into consideration in children’s spellings choices, as well at which stage these start to be considered.The results for first graders show generalized misspellings in which schwa and [u] are systematically associated with the graphemes a and u, regardless these vowels alternate with vowels in stressed position or do not. This behavior clearly supports the Lexicon Optimization Hypothesis, according to which there is a first stage in language acquisition in which, from all possible input candidates (Richness of the Base), the learner selects the one that matches the adult output representation as the optimal input (Smolensky 1996).The results for second graders show a noticeable decline in misspellings for alternating schwa and [u]. This leads to think that morphophonemic alternations are progressively incorporated and taken into account in the process of UR construction. Of course, this better performance in the spelling of alternating vowels might also be a consequence of a better familiarity with conventional spelling. Importantly, though, the decline in misspellings for alternating schwa and [u] coincides with an increase in spelling mistakes for non-alternating vowels. In our view, this undescores the influence of conventional orthography in children’s spelling choices, points to the influence of phonology, and more specifically to a stage of vacillation with respect to the UR of non-alternating forms, with apparent overgeneralizations. This behavior, thus, may support the free-ride approach to morphophonemic learning (McCarthy 2005).The results for both first graders and second graders show differences in the spelling between back vowels (with fewer mistakes) and non-back vowels (with more mistakes), and we argued this might be related to the fact that the children have to deal with a more reduced typology of output-input alternations for back vowels (two-to-one) than for non-back vowels (three-to-one). This is evidence, again, for the role of morphophonemic alternations (i.e. morphology) in children’s spelling choices. The results for both graders also reveal that more productive alternations, such as the ones found between a base and the diminutive, are more transparent to the kids than others: whereas diminutives were generally spelled correctly, non-diminutives were not, giving support again to the influence of morphophonemic alternations, and their transparency, in children’s spelling choices.

A study of common behavior in /s/-irregular and /h/-final stems of Korean

The main purpose of this article is to bring to light the common behavior of the so-called ‘/s/-irregular’ and ‘/h/-final’ stems in Korean, focusing on the similarly opaque phonological behavior of those stems when they encounter vowel-initial suffixes. It is proposed that their phonological similarity should be due to the lack of place feature in their input specifications of the stem-final consonants. The proposal follows from the discussion that combining the notion of Richness of the Base with Lexicon Optimization in optimality-theoretic framework, language learners should consult surface alternations and choose input forms that map onto outputs in the least offensive way with respect to the grammar of ranked constraints. Adopting placeless specifications as their input for both /s/-irregular and /h/-final stems, their apparent opaque alternations are adequately dealt with employing independently motivated universal constraints. The irregular status of /s/-irregular stems in Korean is still maintained but the source of their peculiar behavior turns out to be the lack of place feature just like /h/. Under the proposed ranking of constraints, it is also best explained that the stem-final /s/ in /s/-irregular stems gets their empty place node filled through the help of spreading from the following obstruent while that in /s/-regular stems is simply brought into the purview of place assimilation. I conclude that all relevant cases of /s/- and /h/-final stems can be dealt with successfully in a framework that has recourse to input specifications. (Sookmyung Women’s University)

Does SAE have /flap/? Evidence from Canadian Raising and Vowel Durations

In American English, /t/ and /d/ neutralize to flaps intervocalically: write ∼ writer and ride ∼ rider. There also exist words which contain ambiguous surface flaps that do not alternate, for example, the name Ryder. Does the language user treat this flap like an allophone of /t/ or /d/ (as predicted by the Free Ride Principle), or as an independent segment (as predicted by Lexicon Optimization)? To investigate this question, acoustic data from elicited (Experiment 1) and spontaneous (Experiment 2) speech are analyzed. Voiceless consonants cause the vowel preceding them to be shorter and trigger Canadian Raising. Therefore, if language users treat non-alternating flaps as allophones of /t/, the vowel durations and F1 trajectories of the vowels in these two environments will be similar to each other and different from before flapped /d/, and vice versa if language users treat non-alternating flaps as allophones of /d/. If language users treat non-alternating flaps as different than both the flap allophones of /t/ and /d/, vowel durations and F1 will be different in all three environments. Results show that the durations and F1 trajectories of the vowels in the three environments are all different from each other, providing evidence for Lexicon Optimization.

Multilayer structure based lexicon optimization for language modeling

Multilayer structure based lexicon optimization for language modeling

Selective lexicon optimization

Phonological generalizations that are not absolute, but rather statistic, are a traditional problem for formal phonological theory. This paper gives an example of such a phenomenon and then proposes how such a skewed lexical distribution can be understood in terms of language learning in optimality theory. If lexicon optimization is extended minimally so as to incorporate the fact that the learner is confronted with (random, phonetic) variation, individual words may change, in order to start satisfying lower ranked markedness constraints. If sufficient words have undergone such a change, this may eventually lead to constraint reranking and therefore grammatical change.

Lexicon optimization based on discriminative learning for automatic speech recognition of agglutinative language

For automatic speech recognition (ASR) of agglutinative languages, selection of a lexical unit is not obvious. The morpheme unit is usually adopted to ensure sufficient coverage, but many morphemes are short, resulting in weak constraints and possible confusion. We propose a discriminative approach for lexicon optimization that directly contributes to ASR error reduction by taking into account not only linguistic constraints but also acoustic–phonetic confusability. It is based on an evaluation function for each word defined by a set of features and their weights, which are optimized by the difference in word error rates (WERs) between ASR hypotheses obtained by the morpheme-based model and those by the word-based model. Then, word or sub-word entries with higher evaluation scores are selected to be added to the lexicon. We investigate several discriminative models to realize this approach. Specifically, we implement it with support vector machines (SVM), logistic regression (LR) model as well as the simple perceptron algorithm. This approach was successfully applied to an Uyghur large-vocabulary continuous speech recognition system, resulting in a significant reduction of WER with a modest lexicon size and a small out-of-vocabulary rate. The use of SVM for a sub-word lexicon results in the best performance, outperforming the word-based model as well as conventional statistical concatenation approaches. The proposed learning approach is realized in an unsupervised manner because it does not require correct transcription for training data.

Interlanguage Segmental Mapping as Evidence for the Nature of Lexical Representation

AbstractThe traditional view of phonological representation assumes that lexical representation is economical and free of redundancy and that a phoneme is represented as a combination of contrastive phonetic features. The traditional view, however, is challenged by the recent developments in phonological theories. In Optimality Theory, the distinction between contrastive vs. non‐contrastive aspects of pronunciation is made not by the lexical representation but by the constraint ranking and the commonly adopted principle of Lexicon Optimization requires lexical representation to be as close to the surface representation as possible. An exemplar model of phonology also assumes that predictable and redundant phonetic properties are also specified in the lexical representation and that each contextual variant of a phoneme forms a separate phonetic category. The current paper reviews recent studies on segmental mapping in interlanguage phonology, examines the nature of the L1 phonological representation that the mapping crucially refers to, and discusses how the interlanguage data bear on the debate on the nature of the lexical representation.

Has in Hungarian

The paper discusses the possible analyses of the behaviour of [h] and [x] in Hungarian. It argues that in a derivational, rule-based framework two types of analyses are possible: one that assumes two separate underlying segments, /x/ and /h/, and thus misses the generalisation that the two segments are in complementary distribution, a typical characteristic of allophones. The second kind of approach argues that [h] and [x] come from the same underlying segment; this type of analysis can be further divided into two subtypes. According to one of these, the underlying segment is /h/. To be able to derive the attested output forms, three separate strengthening rules must be posited, an obvious disadvantage. The other possible approach, on the other hand, argues that the underlying segment is always /x/ weakened into a [h] in onsets and deleted in a group of lexically marked words by a minor rule. Besides, we also consider the behaviour of H-type segments in voice assimilation: they trigger but do not undergo that process. Siptár and Törkenczy (2000) suggest that if a filter disallowing surface voiced dorsal fricatives is proposed, then the desired result is obtained. While such a filter is an ad hoc device in rule-based theories, it is an organic part of a solution in Optimality Theory (OT), which argues that both /h/ and /x/ may occur in the input and the constraint hierarchy must be such that they should always select well-formed output candidates as optimal regardless of the input. As a result of this and Lexicon Optimization (LO), non-alternating forms will have /h/ or /x/ in their underlying representation depending on the output forms while alternating forms may have an underlying /x/ or /h/ as a result of the alternation sensitive LO (Inkelas 1994). Finally, we will show that the treatment of the behaviour of /x/ or /h/ in voice assimilation is simple in OT if we assume the constraint proposed by Siptár and Törkenczy (2000), prohibiting voiced dorsal fricatives, which, interacting with the ones suggested by Petrova et al. (2001), will be able to select the actual surface form as optimal in all cases.