Improving syllabification models with phonotactic knowledge

Abstract

We report on a series of experiments with probabilistic context-free grammars predicting English and German syllable structure. The treebank-trained grammars are evaluated on a syllabification task. The grammar used by Müller (2002) serves as point of comparison. As she evaluates the grammar only for German, we reimplement the grammar and experiment with additional phonotactic features. Using bi-grams within the syllable, we can model the dependency from the previous consonant in the onset and coda. A 10fold cross validation procedure shows that syllabification can be improved by incorporating this type of phonotactic knowledge. Compared to the grammar of Müller (2002), syllable boundary accuracy increases from 95.8% to 97.2% for English, and from 95.9% to 97.2% for German. Moreover, our experiments with different syllable structures point out that there are dependencies between the onset on the nucleus for German but not for English. The analysis of one of our phonotactic grammars shows that interesting phonotactic constraints are learned. For instance, unvoiced consonants are the most likely first consonants and liquids and glides are preferred as second consonants in two-consonantal onsets.

On the syllabification of phonemes

“…For her language-independent PCFG-based approach, Müller (2006) reports 92.64% word accuracy on the set of 64K examples from CELEX using 10-fold cross-validation. The Learned EBG approach of van den Bosch (1997) achieves 97.78% word accuracy when training on approximately 60K examples.…”

“…With a hand-crafted grammar, Müller (2002) achieves 96.88% word accuracy on CELEX-derived syllabifications, with a training corpus of two million tokens. Without a handcrafted grammar, she reports 90.45% word accuracy (Müller, 2006). Applying a standard smoothing algorithm and fourth-order HMM, Demberg (2006) scores 98.47% word accuracy.…”

“…This grammar is then made probabilistic using counts obtained from training data. Müller (2006) attempts to make her method language-independent. Rather than hand-crafting her context-free grammar, she automatically generates all possible onsets, nuclei, and codas, based on the phonemes existing in the language.…”

Augmented words to improve a deep learning-based Indonesian syllabification

“…For instance, syllabification on Romanian using a simple Näive Bayes results in a low syllable error rate (SER) of 12.90% [14]. Other examples of data-driven approach models use conditional random fields [15], decision tree, random forest, support vector machine [14], unsupervised model [16], hidden Markov model [17], dropped-and-matched model [18], syllabification by analogy [13], and context-free grammars [19].…”

Data Augmentation Methods for Low-Resource Orthographic Syllabification

“…For example, a simple Näive Bayes produces quite low SER of around 12.90% for the Romanian language [20]. Some other statistical models use decision tree [20] [21], treebank [22], random forest [20], neural network [23] [24] [25] [26], support vector machine [20] [27], finite-state transducers [28] [29], context-free grammars [30], hidden Markov model [31], syllabification by analogy [19], dropped-and-matched model [32], n-gram [33], conditional random fields [34] [35], nearest neighbour [17], and unsupervised model [36].…”