Slu Eval

slu-eval

ESPnet-SLU: Advancing Spoken Language Understanding through ESPnet — Siddhant Arora et al. (2021) (arXiv:2111.14706, 2021)

What this evaluates

Evaluates spoken language understanding capabilities across multiple tasks including intent classification, slot filling, emotion recognition, and dialogue act classification. It probes a model's ability to map raw audio inputs to semantic labels, test robustness to noise and low-resource settings, and assess the utility of pretrained ASR/NLU feature extractors in end-to-end speech processing pipelines.

Datasets

• FSC (Fluent Speech Commands) — total ?; splits: test (-1)
• Snips — total ?; splits: train (-1), test (-1)
• SLURP — total ?; splits: train (-1), test (-1)
• IEMOCAP — total ?; splits: train (-1), test (-1)
• Switchboard (NXT-format) — total ?; splits: test (-1)
• Grabo — total ?; splits: test (-1)
• CAT-SLU MAP — total ?; splits: test (-1)
• Google Speech Commands — total ?; splits: test (-1)
• HarperValleyBank — total ?; splits: test (-1)

Metrics

• Intent Classification Accuracy (primary) — range: percent
  • Percentage of correctly predicted intent labels out of total test instances. F1 is also reported in some tables as a macro-averaged harmonic mean of precision and recall.
• Slot Filling F1 — range: [0, 1]
  • Macro-averaged F1 score for entity and slot label prediction.
• Emotion Recognition Accuracy — range: percent
  • Percentage of correctly predicted emotion classes (neutral, happy, sad, angry) out of total instances.
• Dialogue Act Accuracy — range: percent
  • Percentage of correctly predicted dialogue act classes out of total instances.

Input / output format

Input: Raw audio waveform files, optionally concatenated with transcripts or acoustic signals from preceding utterances for context. Noisy variants include added real-world noise.

Output: Single token or word representing the predicted intent, emotion, or dialogue act class. For slot filling: intent token followed by entity label and lexical filler tokens, separated by special separator tokens.

Scoring recipe

def compute_metric(predictions, gold_labels, metric_type='accuracy'):
    if metric_type == 'accuracy':
        return sum(p == g for p, g in zip(predictions, gold_labels)) / len(gold_labels) * 100
    elif metric_type == 'f1':
        from sklearn.metrics import f1_score
        return f1_score(gold_labels, predictions, average='macro') * 100

Common pitfalls

• Splits are not standardized across datasets; authors use original splits but adopt alternative random splits for Snips, complicating direct comparison.
• Pretrained ASR/NLU feature extractors do not consistently improve performance; on SLURP, raw FBANK features outperformed HuBERT/Wav2vec2.
• Vocabulary mismatch between ASR transcripts and intent/slot labels can degrade performance if not mitigated (e.g., via BPE).

Evidence (verbatim from paper)

Table 5. Intent Classification accuracy on the SLURP Dataset [4] where we perform comparison between different pretrained ASR and NLU systems as feature extractors.

Citation

@misc{arora2021espnetslu,
  title={ESPnet-SLU: Advancing Spoken Language Understanding through ESPnet},
  author={Siddhant Arora et al. (2021)},
  year={2021},
  note={arXiv:2111.14706}
}

• arXiv: 2111.14706