TL;DR
This paper introduces a novel method for automatically generating pronunciations for unknown derived words in English by combining a morphological parser with the Combilex lexicon's existing morphological derivation capabilities. This approach promises to significantly accelerate lexicon development and provide richer phonetic information for out-of-vocabulary (OOV) words in speech technology applications, potentially serving as an alternative to traditional letter-to-sound rules.
What Problem Does This Paper Address?
Speech technology systems, like text-to-speech (TTS) and automatic speech recognition (ASR), heavily rely on accurate mappings between words and their pronunciations. Traditionally, this is achieved through large lexicons or letter-to-sound (LTS) rules. While effective for many languages, English presents a challenge due to its rich derivational morphology, where new words are formed by adding prefixes and suffixes to existing roots. Simply listing all possible derived words in a lexicon is impractical, and standard LTS rules often struggle with the complex phonetic and stress changes that accompany morphological derivation. The Combilex lexicon was designed to address this by explicitly linking derived words to their roots and encoding their pronunciations morphologically. However, the authors identify a gap: how to automatically generate pronunciations for unknown derived words that aren't yet in the lexicon.
Methods
The core of the proposed method involves using a morphological parser to analyze unknown derived words. The output of this parser, which identifies the root word and its affixes, is then fed into Combilex's existing morphological derivation component. This component, which already handles the complex rules for modifying pronunciations based on affixes (including stress shifts and phonetic changes), generates the full pronunciation string and associated linguistic information (like stress, syllable boundaries, and grapheme-phoneme alignments). The authors specifically mention using the PC-KIMMO morphological analysis tool with the Englex English morphology description.
Datasets
The primary "dataset" used for evaluation is the Combilex lexicon itself. The authors analyze the derived words already present in Combilex to assess how well their proposed method would perform on known derived words. They also augment the PC-KIMMO lexicon with over 10,000 new words (proper nouns, nouns, verbs, and adjectives) to improve the parsing success rate.
Key Insights and Ideas
- Morphological derivation is a powerful alternative to LTS: The paper argues that by leveraging morphological analysis, we can achieve richer phonetic information for OOV words than traditional LTS rules, including stress, syllabification, and grapheme-phoneme alignments, which are crucial for modern speech synthesis.
- Combilex's design is key: The explicit linking of derived words to their roots and the encoding of pronunciations morphologically within Combilex are fundamental to the proposed approach. This internal structure makes it amenable to external morphological analysis.
- English morphology is more productive than often assumed: The authors suggest that the number of derived words in English is likely much larger than non-derived words, making morphological analysis a highly relevant strategy for handling OOV words.
- Grapheme-phoneme alignment is critical: The paper highlights that Combilex's inclusion of grapheme-phoneme alignments is essential for handling complex morphological derivations that involve spelling changes and stress shifts, which are difficult for conventional LTS rules.
Key Results
The paper presents preliminary experimental results using PC-KIMMO on the derived words within Combilex.
- Initially, only 55.3% of derived words could be parsed exactly to match Combilex's internal structure.
- After adding over 10,000 new words to the PC-KIMMO lexicon and harmonizing some format differences, the exact match rate improved to 75%.
- The authors categorize reasons for parse failures into incompatible formats, differing structural choices, and multiple parses, indicating areas for future work.
Limitations and Future Work
The authors acknowledge several limitations:
- Incomplete morphological coverage: The PC-KIMMO lexicon and rules do not cover all possible roots and affixes, leading to "no parse" results.
- Inconsistent morphological structures: Differences in how Combilex and PC-KIMMO represent morphological structures (e.g., "abolition" vs. "{abolish}ion>") can lead to non-exact matches.
- Multiple parses: PC-KIMMO can return multiple valid parses for a single word, which requires user selection for lexicon building or further disambiguation for TTS.
- Handling incorrect parses: While "no parse" can be handled by falling back to LTS, incorrect parses are a more significant challenge that needs further investigation.
Future work will focus on:
- Increasing the proportion of words that can be parsed compatibly with Combilex.
- Addressing the issue of incorrect parses.
- Integrating morphological analysis more closely with Combilex's derivation component.
- Conducting a fair comparison with state-of-the-art LTS systems.
Summary updated