Modeling liaison in French by using decision trees

Josafá de Jesus Aguiar Pontes, Sadaoki FuruiInterspeech 2010

TL;DR

This paper tackles the challenging problem of French "liaison" (pronouncing a silent final consonant before a vowel-initial word) for text-to-speech systems. By using C4.5 decision trees trained on a manually curated database of linguistic and phonetic features, the authors achieve state-of-the-art performance, outperforming most existing systems.

What Problem Does This Paper Address?

French pronunciation is notoriously complex, especially at word boundaries. One significant hurdle for text-to-speech (TTS) systems is "liaison," where a normally silent final consonant of a word is pronounced when the following word begins with a vowel, a silent 'h', or certain glides. For example, "mes amis" (my friends) is pronounced with a [z] sound at the boundary, while "mes doigts" (my fingers) is not. Existing rule-based systems for handling liaison are often large, complex, and difficult to maintain, leading to a significant number of pronunciation errors in TTS output. This paper aims to improve the accuracy and manageability of French liaison modeling.

Methods

The core of this paper's contribution lies in its novel approach to modeling liaison using C4.5 decision trees. Instead of manually crafting intricate rule sets, the authors propose an automated method:

  1. Database Creation: A comprehensive database is built by manually clustering liaison cases from linguistic literature. This involves:
  2. Data Selection and Clustering: Approximately 1500 liaison contexts are collected and grouped into sets (Ri) based on shared contextual features (e.g., part-of-speech of words, initial phoneme of the next word). This clustering helps manage complexity and allows for prioritized handling of different liaison types (Forbidden > Obligatory > Optional). Special attention is given to fixed expressions by creating separate rule sets for different lengths.
  3. Attribute Specification: Relevant linguistic and phonetic features (attributes) are identified and categorized into discrete labels (e.g., "next word starting phoneme" can be "vowel," "consonant," "mute_h"). An "other" category is crucial for forcing decision tree splits.
  4. Database Specification: For each rule set Ri, a training table Ti is created by computing the Cartesian product of all attribute values. Each tuple in the table includes the required information for liaison handling (index, phoneme change function, liaison phoneme, etc.) or a "skip" or "ordinary" label. Linguistically invalid combinations are removed.
  5. Decision Tree Training: Each table Ti is used to train a corresponding C4.5 decision tree (DTi). These trees are specifically designed to handle the particular contexts covered by their respective rule sets. The trees are trained without pruning to ensure all valid paths are explored and to guarantee perfect performance on the training data.
  6. Prediction Algorithm: The trained decision trees are integrated into the post-lexical analysis routine of a G2P converter. When processing text, the system analyzes word contexts and passes them through the relevant decision trees (Modules). If a tree identifies a liaison context, the specified phonetic changes are applied; otherwise, the "ordinary" pronunciation is used, or the task is passed to the next module ("skip").

Datasets

  • Training Data: A database of approximately 1500 French liaison contexts, manually compiled from linguistic literature. This database is then processed to create around 50 tables, totaling about 30,000 tuples (reduced to 7,000 after removing invalid entries).
  • Evaluation Data: A French text corpus extracted from "Le Monde" newspaper (January 1987), containing about 26,000 words/tokens and 2,000 sentences. This corpus includes approximately 1,500 liaison candidates (600 obligatory) and manually prepared phonetic transcriptions with pronunciation variations.

Key Insights and Ideas

  • Automating Rule Generation: The paper demonstrates that complex linguistic phenomena like liaison can be effectively modeled by automating rule generation using machine learning (decision trees) rather than relying solely on manual rule coding.
  • Hierarchical and Prioritized Processing: Clustering rules into prioritized sets (Forbidden > Obligatory > Optional) and handling fixed expressions separately significantly reduces complexity and improves prediction accuracy.
  • The Power of "Skip" and "Other" Labels: The inclusion of "skip" labels in the training data is crucial for guiding the decision tree search when a context isn't handled by a particular tree. The "other" attribute category ensures that decision trees explore all possibilities and don't prematurely stop splitting.
  • Generalization through Categorization: Decision trees, by operating on categorized features (like PoS tags or phoneme groups), can generalize predictions to unseen words that share similar linguistic characteristics.

Key Results

The prototype system, built using the decision tree approach, was evaluated against six state-of-the-art G2P converters on a French text corpus. The results show:

  • The proposed method outperforms most of the current converters.
  • It is equivalent to the second-rank system in terms of liaison error reduction.
  • Specifically, the prototype produced only 15 liaison errors, compared to systems B (111 errors), C (123 errors), D (76 errors), F (49 errors), and G (38 errors).

Limitations and Future Work

  • Limitations:
    • The creation of the initial database is a manual process requiring human intervention.
  • The paper focuses primarily on Obligatory and Forbidden liaison contexts, with less emphasis on Optional liaisons.
  • Information regarding the specific methods used by the compared systems is limited, making direct comparison of techniques challenging.
  • Future Work:
    • Enlargement of the database used by the prototype.
  • Investigation of Optional liaison contexts using annotated corpora.
  • Research into other aspects of G2P conversions, such as proper nouns.

Summary updated

BibTeX
@inproceedings{pontes2010modeling,
  author = {Josafá de Jesus Aguiar Pontes and Sadaoki Furui},
  title = {{Modeling liaison in French by using decision trees}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure