TL;DR
This paper introduces KORAIS, a comprehensive Greek-English dictionary enhanced with advanced speech and language processing tools. It aims to significantly improve language learning by integrating features like text-to-speech synthesis, lemmatization, and grapheme-to-phoneme conversion into a user-friendly electronic dictionary.
What Problem Does This Paper Address?
Traditional dictionaries, even in electronic form, often lack the dynamic capabilities needed to fully support language learners. The challenge lies in bridging the gap between static lexical information and the practical, interactive needs of users acquiring a new language. Specifically, this paper addresses the need for a more intelligent and interactive Greek-English dictionary that leverages modern speech and language technologies to provide richer learning experiences, going beyond simple lookups to offer pronunciation guidance, grammatical analysis, and more.
Methods
The KORAIS dictionary is built upon a large Greek-English lexicon and integrates several key speech and language processing tools:
- Text-to-Speech (TTS) Synthesis: The paper utilizes the Festival speech synthesis framework, implementing diphone-based TTS for both Greek and English. This approach aims for reliable and computationally efficient synthesis, allowing for user modifications to rhythm and volume.
- Lemmatization: A language-independent lemmatizer is employed, which uses string distance metrics (Levenshtein Distance and Dice coefficient) to automatically identify the base form (lemma) of inflected words. This process is enhanced by an iterative inflectional suffix removal procedure.
- Grapheme-to-Phoneme (G2P) Conversion: A rule-based G2P converter for Greek is developed, based on the Computer Phonetic Alphabet (CPA). It handles context-dependent rules, particularly for "iotacism" (multiple spellings for the /i/ sound), diphthongs, and double consonants.
- Syllabification System: A two-step system is used for Greek syllabification. A "naïve" function initially separates words based on vowel presence, followed by a "strict" function that refines the segmentation using grammatical rules to improve pronunciation aid.
- Human-Machine Interface (HMI): A user-friendly Graphical User Interface (GUI) is designed with distinct panels for searching, word listing, results display, and options. The interface is adaptable to the user's operating system theme and offers customizable display settings. The underlying data is stored in an SQL database to facilitate efficient and flexible searching.
Datasets
- Greek and English TTS: Prosodic models for English were trained on the Boston University FM Radio corpus. For Greek, the WCL-1 database was used.
- Lemmatization Evaluation: Two test files, one for Modern Greek and one for English, each containing 3,000 regular and mildly irregular words, were used for experimental evaluation.
- Grapheme-to-Phoneme Evaluation: Performance was tested on Greek newspaper texts (10,558 different words from a corpus of 100,000 words), as well as on town names, capital names, and Christian names.
Key Insights and Ideas
- Integrated Learning Environment: The core idea is that combining a rich lexicon with advanced linguistic tools creates a more powerful and effective learning environment than isolated components.
- Efficiency through Diphones: Using diphone-based TTS synthesis offers a good balance between implementation complexity, computational demand, and speech quality for both languages.
- Robust Lemmatization: The combination of Levenshtein Distance and Dice coefficient, along with suffix removal, provides a robust method for automatically identifying word lemmas, achieving high accuracy.
- Rule-Based G2P for Greek Complexity: The G2P system's success in handling Greek's "iotacism" and other complexities highlights the effectiveness of well-defined rule sets for specific linguistic challenges.
- Syllabification for Pronunciation: The two-step syllabification process is crucial for Greek, where syllable structure closely relates to pronunciation, aiding learners significantly.
Key Results
- Lemmatizers: Achieved over 95% accuracy for both Modern Greek and English, with performance significantly improved by inflectional suffix removal.
- Grapheme-to-Phoneme Converter: Demonstrated an overall accuracy of 98.7% on Greek newspaper texts, with further improvements pushing it above 99% after additional rule incorporation. Accuracy on town names and capitals was around 97.3%, and 93.5% for Christian names.
- Syllabification System: Achieved accuracy higher than 99.5% on a test set of 2000 different words.
- Dictionary Size: The dictionary contains 81,515 entries, 192,592 translations into English, and 50,106 usage examples with translations.
Limitations and Future Work
- TTS Quality: While listening tests indicated good intelligibility and acceptable naturalness for the Greek TTS system, the authors note that work is ongoing to further improve its quality.
- Lemmatizer Performance: Although high accuracy was achieved, the paper mentions that the lemmatizer performs well for "regular and mildly irregular words." The handling of highly irregular words might be a point for future exploration.
- Further Tool Integration: The paper lists "other tools" as being included, but does not detail them. Future work could involve expanding the suite of integrated linguistic tools.
- User Feedback: While the interface is described as user-friendly, incorporating user feedback and conducting more extensive user studies could further refine the system.
Summary updated