Prosody cues for classification of the discourse particle "hã" in hindi

Sankalan Prasad, Kalika BaliInterspeech 2010

TL;DR

This paper investigates how prosodic features like pitch, duration, and energy can be used to distinguish the different discourse functions of the Hindi particle "hã" (similar to "yes" or "uh-huh"). The authors found that while prosody alone isn't perfect for classification, it offers valuable cues, especially when combined with other linguistic information in the future.

What Problem Does This Paper Address?

The Hindi particle "hã" is a versatile word that can function as a discourse particle (like an acknowledgement, backchannel, or agreement) or as a direct lexical equivalent of "yes." Disambiguating these different functions is crucial for building more natural and effective spoken dialogue systems. While previous research has explored prosody for disambiguating similar particles in other languages, there's a notable gap in understanding how prosodic cues specifically help classify the diverse meanings of "hã" in Hindi, a language with many dialects and a high frequency of this particle's use. The challenge lies in identifying which specific prosodic features are most indicative of each function and how well they perform in isolation.

Methods

The researchers collected a corpus of spontaneous Hindi telephonic conversations and manually extracted instances of "hã." They then analyzed these utterances using prosodic features such as pitch contour, power (energy), and duration. To capture pitch variations, they used a set of notations similar to ToBI, defining specific pitch profiles for different functions. For classification, they employed a k-means clustering algorithm. Before clustering, they performed feature selection using a regression tree on a subset of the data to identify the most relevant features out of an initial set of 42.

Datasets

The study utilized a custom corpus of 35 spontaneous telephonic conversations in colloquial Hindi, featuring 70 unique speakers (49 male, 21 female) representing five different dialects. A total of 1322 "hã" tokens were extracted. These tokens were manually annotated into three main categories: questions (133 utterances), agreement/acknowledgement/yes (550 utterances), and backchannels (639 utterances).

Key Insights and Ideas

  • Prosody as a Signal: The core idea is that different intentions behind "hã" manifest as distinct prosodic patterns. The researchers hypothesized that by identifying these patterns, they could classify the function of the particle.
  • Distinct but Overlapping Profiles: While each function (question, backchannel, agreement) tends to have a characteristic prosodic profile (e.g., rising pitch for questions, flatter for backchannels), there's significant overlap, especially between backchannels and agreements.
  • Duration and Power as Differentiators: For ambiguous cases, longer duration and higher power were found to be more indicative of agreement, while shorter duration and lower power suggested backchannels.
  • Speaker Dependency: The study acknowledges that prosodic patterns can be speaker-specific, which can impact classification accuracy.
  • Limited Context, Limited Success: The experiment focused on isolated utterances, meaning contextual and lexical information was excluded. This limitation is highlighted as a reason for the observed classification entropy.

Key Results

  • Feature Importance: Out of an initial 42 prosodic features, a regression tree identified 23 as most relevant for classification.
  • Classification Performance: Using k-means clustering on the selected features, the average entropy of classification was 0.4184, indicating that approximately 41% of utterances were misclassified.
  • Precision vs. Recall: Identification of backchannels achieved the highest precision (65% of classified backchannels were indeed backchannels), while agreements had the highest recall (61% of all agreements were positively identified).
  • Comparison to Other Languages: The misclassification percentage, while high, is noted as being comparable to similar studies in other languages.

Limitations and Future Work

  • Isolated Utterances: The primary limitation is the classification of isolated "hã" utterances without any lexical or contextual information. This significantly impacts accuracy.
  • Speaker and Dialect Variation: The inherent variability in prosody across speakers and the five represented dialects contributed to the high misclassification rate.
  • Prosody's Multifunctionality: Prosody is not a single-purpose signal; factors like focus, stress, and emotion can also influence patterns, making it hard to isolate the function of "hã" solely based on prosody.

The authors suggest that future work should focus on combining prosody with lexical and acoustic information. Additionally, incorporating contextual information from the surrounding discourse (i.e., utterances from the other interlocutor) is proposed as a way to achieve stronger disambiguation cues.

Summary updated

BibTeX
@inproceedings{prasad2010prosody,
  author = {Sankalan Prasad and Kalika Bali},
  title = {{Prosody cues for classification of the discourse particle "hã" in hindi}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure