A preliminary study of Mandarin filled pauses

Yuan Zhao, Dan JurafskyDiSS 2005

TL;DR

This paper offers a preliminary look at Mandarin filled pauses (FPs), revealing that Mandarin speakers use demonstratives like "zhege" (this) and "nage" (that) as FPs, in addition to the more common "uh" and "mm." These demonstrative FPs appear to signal more complex planning issues, particularly in nominal searches, and their usage is influenced by regional dialects.

What Problem Does This Paper Address?

Filled pauses (FPs) are a common feature of spontaneous speech, acting as verbal placeholders that can significantly complicate automatic speech recognition (ASR) systems. While extensive research exists on FPs in English and other languages, there's a notable gap in empirical studies on Mandarin FPs. Previous work on Mandarin has primarily focused on repairs and repetitions, leaving the distinct forms, acoustic properties, and discourse functions of Mandarin FPs largely unexplored. This lack of understanding leads to confusion in transcription and hinders the development of robust Mandarin ASR. This paper aims to fill this gap by identifying Mandarin FPs, investigating their acoustic characteristics, and analyzing their distribution and potential discourse functions.

Methods

The study employed a multi-corpus approach to identify Mandarin FPs. The primary dataset was the LDC 98-HUB5 Mandarin corpus of telephone conversations, which included hand-labeled "uh" and "mm" FPs. To identify demonstrative FPs (zhege, nage), the first author manually labeled them across this corpus and two additional LDC corpora: 2003-HUB5 and 96-CallHome. Acoustic features, specifically duration and pitch movement, were analyzed from a sample of 100 FPs produced by male speakers. The distribution of FPs across various syntactic units (clauses, predicates, complements, within NPs) was also examined. Finally, the study investigated the influence of demographic factors (gender, age, education, dialect) on FP usage.

Datasets

  • LDC 98-HUB5 Mandarin corpus of telephone conversations: Used for analyzing "uh" and "mm" FPs, which were hand-labeled by LDC. This corpus contains 37 conversations with identifiable speaker information (gender, dialect, age, education).
  • LDC 2003-HUB5 and 96-CallHome telephone conversation corpora: Used in conjunction with the 98-HUB5 corpus to identify demonstrative FPs (zhege, nage), which were hand-labeled by the first author.

Key Insights and Ideas

  • Mandarin's FP Repertoire: Mandarin speakers utilize not only the common "uh" and "mm" but also demonstratives like "zhege" (this) and "nage" (that) as filled pauses.
  • Demonstratives as Planning Signals: Demonstrative FPs are significantly longer than their non-FP counterparts and are more likely to be accompanied by other disfluencies (repetitions, false starts). This suggests they signal more serious planning problems, potentially related to lexical retrieval, especially in nominal contexts.
  • Syntactic Distribution Differences: "Uh" and "mm" tend to occur at the beginning of clauses, while demonstrative FPs are more prevalent before complements, particularly noun phrases, hinting at a role in nominal search.
  • Dialectal Influence: Regional dialects play a role, with southern Chinese speakers using significantly more FPs, particularly "uh," compared to northern speakers.
  • Acoustic Distinctions: "Mm" FPs are acoustically longer than "uh" FPs, and Mandarin FPs, in general, tend to be prolonged rather than followed by short pauses, suggesting a strategy to fill longer pauses.

Key Results

  • FP Frequency: The total FP frequency in Mandarin (10.69 per 1000 words) appears lower than in English corpora.
  • Demonstrative FP Dominance: Demonstratives ("zhege" and "nage") constitute the largest FP category in Mandarin, with a combined frequency of 6.68 per 1000 words, compared to 4.01 per 1000 words for "uh" and "mm."
  • Duration Differences: "Mm" FPs are over 1.5 times longer than "uh" FPs. Demonstrative FPs ("zhege" and "nage") are approximately twice as long as their non-FP uses.
  • Syntactic Placement: Demonstrative FPs are most likely to occur before complements (34% for "nage," 28.4% for "zhege") and noun heads (18.4% for "nage," 23.5% for "zhege"), totaling over 50% in nominal environments. "Uh" and "mm" are primarily clause-initial.
  • Disfluency Association: Demonstrative FPs are significantly more likely to be accompanied by repetitions and other disfluencies (34.3% for "zhege," 29.3% for "nage") than "uh" and "mm" (14.8% for "uh," 10.7% for "mm").
  • Dialectal Impact: Southern speakers use significantly more FPs (especially "uh") than northern speakers.
  • Demographic Factors: Gender, age, and education levels did not show a significant effect on FP usage, though a trend suggested more educated speakers use fewer FPs. An interaction between sex and region was significant, with southern female speakers using the most FPs and northern male speakers the least.

Limitations and Future Work

  • Preliminary Nature: The authors emphasize that this is a preliminary study, and much more work is needed to fully understand Mandarin FPs.
  • Manual Labeling: The identification of demonstrative FPs relied on manual labeling, which can be time-consuming and subjective. Future work could explore automated methods.
  • Limited Corpora: While three corpora were used, expanding to more diverse datasets and conversational styles would provide a more comprehensive picture.
  • Sociolinguistic Variables: While dialect showed a significant effect, further investigation into the interaction between demographic factors and FP usage is warranted.
  • Discourse Functions: The study suggests hypotheses about the discourse functions of demonstrative FPs (e.g., nominal search), but these require more in-depth qualitative and quantitative analysis.
  • Acoustic Nuances: While duration and pitch were analyzed, other acoustic features and their relation to FP function could be explored.

Summary updated

BibTeX
@inproceedings{zhao2005preliminary,
  author = {Yuan Zhao and Dan Jurafsky},
  title = {{A preliminary study of Mandarin filled pauses}},
  booktitle = {DiSS},
  year = {2005},
}

Loading related content...

Figure