TL;DR
This paper investigates how the order of tones relative to segments in a syllable affects phonotactic probability calculations. They found that while traditional n-gram models are somewhat sensitive to this ordering, modern recurrent neural networks are remarkably robust, suggesting that phonotactic probability is largely independent of tone placement within the syllable.
What Problem Does This Paper Address?
Phonotactic probability, a measure of how likely a sequence of sounds is in a language, is crucial for understanding language acquisition, word recognition, and speech production. However, for tone languages (where pitch changes meaning), it's unclear how to best incorporate tones into these probability calculations. Should a tone be treated as a segment, and if so, where should it be placed relative to consonants and vowels in the sequence? This ambiguity can lead to different models and potentially different conclusions about phonotactic constraints. The paper aims to resolve this by systematically testing various tone-segment orderings across different language models and languages.
Methods
The researchers employed two main types of language models:
- Smoothed Trigram Models: These models calculate the probability of a sequence based on the preceding two "segments" (which can include tones). They use a technique called interpolated Witten-Bell discounting for smoothing, which helps handle unseen sequences.
- Recurrent Neural Networks (RNNs): These models use a hidden state to maintain information about the entire preceding sequence, allowing them to capture longer-range dependencies than simple n-gram models. The paper used simple RNNs, noting that similar results were obtained with LSTMs.
The core of their method involved creating five distinct permutations of tone and segments for each syllable in the lexicons: tone following the coda (T/C), nucleus (T|N), medial (T|M), onset (T|O), and as the initial segment (T|#). They then trained both trigram and RNN models on these permuted lexicons and evaluated their performance using perplexity, a measure of how well the model predicts unseen data (lower perplexity is better).
Datasets
The study used syllable lexicons from four syllable-tone languages:
- Mandarin Chinese (cmn): 1,226 syllables.
- Thai (tha): 4,133 syllables.
- Vietnamese (vie): 8,128 syllables.
- Cantonese (yue): 1,884 syllables.
These lexicons were derived from existing sources and phonetized using established systems. The paper notes that these languages were chosen for convenience (available resources) and their shared syllable structure template and number of tones, while still exhibiting variations in segment-tone co-occurrence restrictions.
Key Insights and Ideas
- RNNs are remarkably robust to tone ordering: The most striking finding is that RNNs perform almost identically regardless of how tones are placed relative to segments. This suggests that their ability to capture complex dependencies within the syllable makes the specific linear ordering less critical.
- Trigram models show some sensitivity: While not as pronounced as one might expect, trigram models did show differences in perplexity based on tone ordering, particularly in languages with coda-tone restrictions.
- Tone ordering matters more when segment-tone constraints are strong: The trigram models' sensitivity was linked to specific phonotactic rules, like restrictions on tones following obstruent codas. When tone was placed before these restricted codas (e.g., T|#), the trigram model struggled more.
- Intrinsic evaluation provides a useful baseline: The study advocates for using intrinsic metrics like perplexity on the lexicon itself as a stable baseline for evaluating tone incorporation, as extrinsic measures (like human judgments) can be more variable.
Key Results
- RNNs consistently outperformed trigram models: Across all languages and permutations, RNNs achieved significantly lower perplexity, indicating a better fit to the phonotactic regularities of the lexicons.
- Minimal impact of tone ordering on RNNs: For RNNs, the numerical differences in perplexity between the best and worst tone orderings were very small, often less than 0.1.
- Trigram models showed some variation: In Mandarin, the T/C ordering yielded the lowest perplexity for trigrams. For Thai, Vietnamese, and Cantonese, the T|# ordering generally resulted in higher perplexity compared to other orderings, especially when tone was placed before a coda. The TM ordering was optimal for Thai and Cantonese trigrams.
- Table 1 and Table 2 in the paper provide detailed quantitative results showing the mean perplexity for each language, model, and ordering, along with standard deviations.
Limitations and Future Work
- Limited scope of languages: The study focused on four typologically similar syllable-tone languages. Future work could explore languages with more complex tone systems (e.g., word-tone languages, tonal specification at sub-syllabic levels) or different syllable structures.
- Representation of tone: The paper treated tone as a discrete segment. Future research could explore representations based on phonological features, which might offer theoretical insights, especially for tasks like measuring phonological distance.
- Necessity of tone in phonotactics: The paper raises the question of whether tone should be included in phonotactic probability calculations at all, acknowledging that empirical investigation is needed to determine if it improves predictions of well-formedness judgments or is relevant for specific research questions (like tone sandhi).
- Morphological tone: The study focused on lexical tone. The role of tone in morphological processes and how its ordering might affect phonotactic calculations in such contexts remains an open area.
Summary updated