Semantic context effects in the recognition of acoustically unreduced and reduced words

Chao Wang, Johan Schalkwyk, Roberto Sicconi, Geoffrey Zweig, Marco van de Ven, Benjamin V. Tucker, Mirjam ErnestusInterspeech 2009

TL;DR

This study investigates how semantic context helps us understand spoken words, especially when people talk quickly and "reduce" their words (like slurring or dropping sounds). Surprisingly, semantic context seems to help more with clear, unreduced speech than with fast, reduced speech, suggesting our brains might rely on different strategies depending on speech clarity.

What Problem Does This Paper Address?

When we listen to someone speak, especially in casual conversation, they often don't pronounce every word perfectly. Sounds might be dropped, blended, or shortened – this is called "acoustic reduction." While this makes speech faster and more efficient for the speaker, it can make it harder for the listener. Previous research suggests that listeners use context (both semantic, meaning the surrounding words' meanings, and syntactic, their grammatical role) to help them figure out these reduced words.

However, there's a puzzle: if reduced words are harder to understand, shouldn't listeners rely more on context to decipher them? Some recent studies have shown that factors like word frequency, which usually help us recognize words, seem to have less of an effect on reduced words. This raises the question: does semantic context also play a less significant role when speech is reduced? This paper aims to directly compare how semantic context affects the recognition of both unreduced (clear) and reduced (fast) words.

Methods

The researchers conducted two auditory lexical decision experiments. In these experiments, participants listened to word pairs and had to decide if the second word was a real word or a made-up one. The key manipulation was the semantic relatedness between the two words in a pair, which was measured using Latent Semantic Analysis (LSA). LSA quantifies how semantically similar two words are based on their co-occurrence in large text corpora. They also controlled for word frequency and the duration of the spoken words.

  • Experiment 1: Used clear, unreduced pronunciations of words.
  • Experiment 2: Used the same words but recorded with a speaker instructed to speak as quickly as possible, leading to acoustically reduced pronunciations.

They used generalized linear mixed-effects models to analyze the data, looking at both accuracy and response times.

Datasets

  • Corpus of Contemporary American English (COCA): Used to extract 154 nouns with varying frequencies for the word materials.
  • Buckeye Corpus of American English: Mentioned as a source for understanding the prevalence of word reduction in spontaneous speech.
  • Custom Stimuli: The researchers created word pairs with varying LSA scores (from mildly to highly related) and also included unrelated filler pairs and pseudowords. The spoken stimuli were recorded by a male speaker of Canadian English.

Key Insights and Ideas

  • Semantic Context is a Lifeline for Clear Speech: For unreduced words, semantic context significantly aids recognition, especially when words are less frequent. This suggests that when speech is clear, our brains can leverage meaning to predict and understand words more easily.
  • Reduced Speech is a Different Ballgame: Surprisingly, semantic context did not show a significant effect on the recognition of reduced words in this study. This challenges the intuitive idea that listeners would lean more heavily on meaning when faced with unclear speech.
  • Frequency Still Matters, But Differently: Word frequency consistently helped participants respond faster to both reduced and unreduced words. However, in unreduced speech, this frequency effect interacted with semantic relatedness: higher frequency words were recognized faster, but this benefit was most pronounced when the words were not semantically related. This suggests a "cue trading" phenomenon – if meaning is very clear, frequency might be less critical, and vice-versa.
  • Listener-Driven Reduction Hypothesis Questioned: The findings are unexpected given the "listener-driven" account of speech reduction, which posits that speakers reduce words that are already predictable from context. If semantic context is less helpful for reduced words, this account might need revision.

Key Results

  • Experiment 1 (Unreduced):
    • Word frequency facilitated recognition (faster responses for more frequent words).
  • An interaction between word frequency and LSA was found: the benefit of word frequency was reduced for highly semantically related words, suggesting that strong semantic context might even hinder recognition in some cases (possibly due to competitor suppression).
  • Experiment 2 (Reduced):
    • Word frequency facilitated recognition.
  • No significant effect of LSA (semantic context) was found.
  • Word durations were significantly shorter than in Experiment 1.
  • Combined Analysis:
    • Participants performed worse (more errors) on reduced words than unreduced words.
  • A significant three-way interaction between word frequency, LSA, and Experiment (reduced/unreduced) confirmed the different roles of semantic context across the two conditions.

Limitations and Future Work

  • Lexical Decision Task: The study used a lexical decision task, which might not fully reflect natural listening comprehension in a sentence. Future work could investigate these effects in sentence contexts.
  • Degree of Reduction: The study used a specific type of reduction (fast speech). It would be interesting to see if these findings hold for other types of reduction or for more extreme levels of reduction.
  • Semantic Relatedness Thresholds: The study focused on LSA scores between 0.34 and 1. It's suggested that larger differences in semantic relatedness might still show an effect in reduced speech.
  • Independent Trials: The word pairs were presented in independent trials. Further research is needed to test whether similar patterns emerge when semantically related words occur within the same sentence.

Summary updated

BibTeX
@inproceedings{wang2009semantic,
  author = {Chao Wang and Johan Schalkwyk and Roberto Sicconi and Geoffrey Zweig and Marco van de Ven and Benjamin V. Tucker and Mirjam Ernestus},
  title = {{Semantic context effects in the recognition of acoustically unreduced and reduced words}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure