Predicting Prosodic Boundaries for Children's Texts

Mansi Dhamne, Sneha Raman, Preeti RaoEMNLP 2025 · OpenReview

TL;DR

This paper introduces a novel text-based model to predict natural pause locations (prosodic boundaries) in children's reading, addressing the unique challenges of their developing prosody. By combining linguistic features with a new "breath duration" feature, the model achieves strong performance, paving the way for better educational tools and more natural text-to-speech systems for children.

What Problem Does This Paper Address?

Reading aloud isn't just about decoding words; it's also about natural phrasing and rhythm, known as prosody. While adults tend to pause at grammatical boundaries like punctuation, children aged 8-13 often struggle with this, inserting awkward pauses due to limited breath control and developing prosodic awareness. This can hinder comprehension and make their reading sound unnatural. Existing methods for predicting these pauses often rely on punctuation alone or acoustic cues, which aren't available in plain text. This paper tackles the challenge of predicting cognitively appropriate pause locations in children's reading material using only text-based features, aiming to model both natural pauses and "forbidden" (inappropriate) ones.

Methods

The researchers developed a text-based model that leverages a rich set of features. Key among these is a novel breath duration feature. This feature quantifies the number of syllables uttered since the last major prosodic boundary, directly inspired by the physiological limitations of children's breath capacity (around 5-7 syllables per breath for 8-13 year olds, compared to 14-16 for adults).

Beyond this novel feature, the model incorporates:

  • Lexical Features: Word form, capitalization, punctuation, word length, and BERT embeddings.
  • Syntactic and POS Features: Part-of-speech tags, dependency relations, and constituency parse information (like constituent depth and label).
  • Semantic Features: Contextualized word embeddings reduced via PCA and clustered to represent semantic similarity.
  • Positional Features: Word position within the sentence (absolute and normalized).

The model was trained using the LightGBM framework, employing both regression for predicting boundary strength (based on annotator agreement) and classification for binary tasks (presence/absence of boundaries and forbidden pauses).

Datasets

The study introduces a new, curated dataset comprising 54 leveled English stories specifically designed for children aged 8-13. These stories were annotated for prosodic boundaries by 21 fluent English speakers. The dataset contains 8662 words and 701 sentences, with annotations reflecting cognitively appropriate pause locations for child readers. The data was collected with a focus on ensuring inter-annotator agreement, achieving an average Fleiss' Kappa of 0.68.

Key Insights and Ideas

  • Punctuation is Not Enough: A significant finding is that nearly 30% of natural pauses occur at non-punctuation locations. This strongly suggests that relying solely on punctuation for prosodic boundary prediction is insufficient, especially for children's texts.
  • Breath Matters: The introduction of the "breath duration" feature, grounded in physiological research on children's breath capacity, proves crucial. This highlights the importance of incorporating developmental and physiological constraints into prosody modeling.
  • Continuous Boundary Strength: Treating prosodic boundary strength as a continuous variable (based on annotator votes) rather than a binary decision allows for more nuanced modeling, reflecting the inherent ambiguity in some pause locations.
  • Grade-Level Trends: The study observed that as grade level increases, the average number of syllables between pauses also increases. This indicates that prosodic phrasing naturally adapts to growing syntactic complexity and cognitive load in older children.

Key Results

The model achieved strong performance across various tasks:

  • Regression: The full model achieved an R² of 0.87, RMSE of 0.94, and MAE of 0.57 for predicting prosodic boundary strength. Performance was best at the extremes (unanimous agreement for pause or no pause) and weakest in ambiguous mid-ranges.
  • Classification:
    • Prosodic Boundary Detection (PBD): Achieved an F1-score of 0.89 and an accuracy of 0.96.
  • Forbidden Pause Detection (FPD): Achieved an F1-score of 0.87 and an accuracy of 0.84.

Feature importance analysis revealed that syntactic and dependency features carried the most weight, followed by lexical information and syllable counts, supporting the hypothesis that grammatical structure is a primary driver of phrasing, influenced by cognitive and physiological factors.

Limitations and Future Work

The authors acknowledge several limitations:

  • Text-Only Simplification: The model relies solely on text features, potentially missing nuances influenced by speaker intent, emphasis, or pragmatic context, which can affect prosody.
  • Generalizability: The model assumes prosodic patterns generalize across readers and linguistic backgrounds, but dialectal variations and individual reading habits might impact performance.
  • Annotation Subjectivity: Human annotations, while aiming for consistency, inherently introduce subjectivity and potential inconsistencies.
  • Limited Scope: The model does not explicitly model multi-level prosodic features like boundary tones or pitch accents, which could enhance naturalness.

Future work could involve expanding the dataset to include more diverse texts and languages (e.g., Hindi and Marathi), incorporating acoustic measurements for validation, and exploring more advanced prosodic features.

Summary updated

Abstract

Reading fluency in any language requires accurate word decoding but also natural prosodic phrasing i.e the grouping of words into rhythmically and syntactically coherent units. This holds for, both, reading aloud and silent reading. While adults pause meaningfully at clause or punctuation boundaries, children aged 8-13 often insert inappropriate pauses due to limited breath control and underdeveloped prosodic awareness. We present a text-based model to predict cognitively appropriate pause locations in children’s reading material. Using a curated dataset of 54 leveled English stories annotated for potential pauses, or prosodic boundaries, by 21 fluent speakers, we find that nearly 30% of pauses occur at non-punctuation locations of the text, highlighting the limitations of using only punctuation-based cues. Our model combines lexical, syntactic, and contextual features with a novel breath duration feature that captures syllable load since the last major boundary. This cognitively motivated approach can model both allowed and “forbidden” pauses. The proposed framework supports applications such as child-directed TTS and oral reading fluency assessment where the proper grouping of words is considered critical to reading comprehension.
BibTeX
@inproceedings{dhamne2025predicting,
  author = {Mansi Dhamne and Sneha Raman and Preeti Rao},
  title = {{Predicting Prosodic Boundaries for Children's Texts}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=949qHohQTH},
}
Figure