TL;DR
This paper introduces Rhapsodie, a project aiming to create a richly annotated corpus of spoken French, focusing on the interplay between syntax and prosody. They propose a novel annotation strategy that embraces, rather than removes, the disfluencies and reformulations characteristic of spontaneous speech, enabling a more comprehensive analysis of oral language.
What Problem Does This Paper Address?
Analyzing spoken language presents unique challenges compared to written text. Spontaneous speech is characterized by disfluencies (hesitations, repetitions, false starts), reformulations, and a less rigid syntactic structure. Traditional linguistic analysis often "cleans up" these features, treating them as noise to be removed before applying standard parsing techniques developed for written language. This approach, however, risks losing valuable information about the natural flow and structure of oral communication. The authors highlight the lack of freely available, richly annotated spoken French corpora, particularly those that integrate syntactic and prosodic information while preserving the natural characteristics of speech.
Methods
The Rhapsodie project tackles this by developing a multi-level annotation strategy that integrates automatic and manual processing.
- Embracing Disfluencies: Instead of removing them, the project annotates disfluencies and reformulations as "entassements paradigmatiques" (paradigmatic stackings). This approach, inspired by the "école d'Aix," treats these phenomena as integral parts of the utterance's structure, not as errors.
- Segmentation: The primary unit of analysis is the "Segment Rectionnel Elargi" (SRE), which approximates the intuitive notion of a sentence. Crucially, SREs can span across speaker turns, reflecting collaborative utterance construction.
- Multi-Level Annotation: The process involves several stages:
- Level 000 (Brute Transcription): Raw orthographic transcription of the audio, including speaker changes and para-linguistic information.
- Level 00 (Simple Automatic Preprocessing): Automatic identification of trivial disfluencies and discourse markers, which are then manually corrected.
- Level 0 (Macro-syntactic Tagging): Manual annotation of macro-syntactic units like entassements and incises, using a custom tool called Pilepilot.
- Level 0+ (Syntactic Analyzer Input): The output of Level 0 is used to segment the text into pieces that can be processed by existing written-language syntactic analyzers.
- Level 1- (Syntactic Analyzer Output): The raw output from automatic analyzers, which will require translation and merging.
- Level 1 (Corrected Syntactic Annotation): Manual correction of the automatic analysis, aiming for a complete syntactic analysis (micro and macro-syntax).
- Tool Development: The project utilizes and develops tools like Pilepilot for efficient manual annotation, supporting an "analysis in grid" presentation and direct audio playback.
Datasets
The paper does not explicitly name a specific dataset used for training or evaluation. Instead, it describes the creation of a new corpus, Rhapsodie, which is intended to be a "reference corpus of spoken French." This corpus is being built from existing recordings and will be freely distributed with specific prosodic and syntactic annotations.
Key Insights and Ideas
- The Value of "Noise": The most significant insight is the reframing of disfluencies and reformulations not as errors but as crucial structural elements of spoken language. This allows for a deeper understanding of how speakers construct meaning in real-time.
- Bridging Oral and Written Analysis: By using existing written-language parsers on pre-processed and tagged oral data, the project offers a pragmatic approach to leveraging powerful existing tools for spoken language analysis without needing to build entirely new parsers from scratch.
- Prosody-Syntax Interplay: The project's core goal is to study the interface between syntax and prosody. The proposed annotation scheme aims to provide the necessary data to explore how prosodic contours align with or diverge from syntactic structures.
- Collaborative Utterances: The concept of SREs spanning speaker turns highlights the interactive and collaborative nature of spoken discourse.
Key Results
The paper focuses on the methodology and annotation strategy rather than presenting quantitative results from a fully annotated corpus. The key "result" is the proposed annotation framework itself, which aims to:
- Enable a comprehensive syntactic analysis of spoken French, including disfluencies and reformulations.
- Provide a basis for studying the relationship between syntax and prosody in spoken language.
- Contribute to the development of analyzers for oral language by creating a richly annotated corpus.
Limitations and Future Work
- Limitations:
- The paper acknowledges that automatic analyzers still make errors (estimated at 30% of dependencies needing correction) and do not use the same tagset as Rhapsodie.
- The current strategy of duplicating segments for analysis means that some text fragments will be analyzed multiple times, requiring subsequent merging and potential manual intervention if analyses differ.
- The naming of annotation levels is noted as being "unconventional" due to historical reasons.
- Future Work:
- The authors plan to further enrich the corpus with discourse-level annotations, including coreference chains and discourse relations, to build a more connected textual structure.
- They aim to confirm their hypothesis that prosodic constituents derive from topological structure by correlating topological constituent boundaries with prosodic boundaries.
- The project's ultimate goal is to contribute to the development of analyzers for spoken language.
Summary updated