TL;DR
This paper introduces a system for automatically tagging spoken English with grammatical information, using prosodic features and a general-purpose dictionary to overcome the challenges of spontaneous speech. The proposed method focuses on identifying grammatical phrases by searching from right to left within "tone units," a novel approach that aims for efficiency and linguistic insight.
What Problem Does This Paper Address?
Accurately tagging spoken language with grammatical information is a significant challenge. Unlike carefully constructed written text, spoken English is characterized by hesitations, disfluencies, and a less rigid structure. Existing computational linguistics approaches often struggle with this variability, typically focusing on restricted "universes of discourse" or relying heavily on manual annotation. This paper aims to bridge this gap by developing a system that can process the complexities of spontaneous spoken English, making it more amenable to linguistic analysis and computational processing. The authors highlight the need for a system that can extract information directly from perceptual strategies used by human listeners.
Methods
The proposed system for grammatical tagging of spoken English is built upon several key methodological choices:
- Tone Units as the Basic Unit: The analysis is centered around "tone units," which are segments of speech marked by prosodic features like pitch changes, stress, and pauses. This choice is motivated by previous research showing a strong correlation between tone units and grammatical units (groups, phrases, and clauses).
- Surface-Level Analysis: The initial tagging relies on "surface properties" – directly observable features of the speech signal and orthographic transcription. This approach aims to capture information that is readily available to human listeners during decoding.
- Interactive, Semi-Manual Analysis: Recognizing that a fully automatic system for grammatical tagging does not yet exist, the authors propose an interactive, semi-manual mode of analysis. This allows for human intervention to correct or refine automated decisions.
- Word Class Tagging: The system utilizes a general-purpose dictionary to provide initial word class tags for each word.
- Phrase Recognition (Right-to-Left Search): A core innovation is the search for grammatical phrases from right to left within each tone unit. This unorthodox approach is justified by the typical structure of English phrases, where the "head" of the phrase (e.g., the verb in a verb phrase, the noun in a noun phrase) is often to the right. This strategy is believed to be computationally economical and linguistically rewarding, potentially simplifying the handling of left-recursion structures and reducing ambiguity compared to left-to-right methods.
- Hierarchical Phrase Recognition: The phrase recognition rules are applied in a specific order: Verb Phrases (VPH), Adverb Phrases (APH), Adjective Phrases (JPH), Noun Phrases (NPH), and Prepositional Phrases (PPH).
Datasets
The primary dataset used in this work is the London-Lund Corpus (LLC) of spoken English. This corpus consists of approximately 450,000 words of transcribed spoken English, which includes orthographic transcription and detailed prosodic marking (tone units, nuclei, stresses, pauses, etc.). The material is available in both machine-readable and printed forms. The paper also mentions the use of word classes from the Longman Dictionary of Contemporary English for initial tagging.
Key Insights and Ideas
- Prosody as a Grammatical Cue: The paper strongly emphasizes the role of prosodic features (tone units, nuclei, stresses) in signaling grammatical structure in spoken English. This suggests that a deeper integration of prosody can significantly improve grammatical analysis.
- Right-to-Left Phrase Search Efficiency: The unconventional right-to-left search strategy for identifying grammatical phrases is a key idea. The authors hypothesize that this approach is more efficient and linguistically grounded for English phrase structures, potentially simplifying parsing.
- Tone Units as Grammatical Units: The finding that tone units often align with grammatical units (groups, phrases, clauses) provides a strong foundation for using tone units as the primary analytical unit.
- Bridging Spoken and Written Language Analysis: While specifically designed for spoken English, the authors suggest that the proposed analysis should be largely applicable to written English, indicating a potential for a unified approach.
Key Results
The paper outlines the proposed system and its methodology rather than presenting extensive quantitative results. However, it highlights the following:
- System Outline: A detailed outline of a system for grammatical tagging of spoken English is presented.
- Interactive Mode: The system is designed to operate in an interactive, semi-manual mode, acknowledging the current limitations of full automation.
- Phrase Recognition Order: A specific order for recognizing different types of grammatical phrases is established.
- Linguistic and Computational Benefits: The right-to-left search is expected to yield both linguistic insights and computational economies.
Limitations and Future Work
- Semi-Manual Nature: The current system is semi-manual, meaning full automation has not yet been achieved. This is acknowledged as a limitation, with the goal of increasing automation being a key future direction.
- Applicability to Written English: While the authors suggest broad applicability, the primary focus and validation are on spoken English. Further work would be needed to fully assess its performance on written text.
- Ambiguity Resolution: The paper touches upon disambiguation of tokens but does not detail the specific mechanisms or their success rates. This remains an area for further development.
- Increasing Automation: The authors explicitly state an aim to "see how far it is possible to carry out syntactic analysis based on graphic words and prosody... and word class tags." This implies that increasing the level of automatic analysis is a primary goal for future work.
Summary updated