SWDA Alignment

Alignments for Switchboard Dialog Act (SWDA) tags with Mississippi State (MS-State) word-level transcripts with timing information.

Overview

This project aligns the dialog act annotations from the Switchboard Dialog Act Corpus with the word-level time-aligned transcripts from the Mississippi State Switchboard annotations. The result is a dataset where each utterance has:

• Dialog act tags (from SWDA)
• Word-level timing information (from MS-State transcripts)
• Speaker identification
• Aligned transcripts from both sources

Data Sources

• SWDA (Switchboard Dialog Act Corpus): Dialog act annotations for Switchboard conversations

  • 1,155 conversations with DA tags
  • Utterance-level annotations with act tags (statement, question, backchannel, etc.)

• MS-State Transcripts: Word-level time-aligned transcripts

  • 2,438 conversations available
  • Word-level start/end timestamps
  • Located in swb_ms98_transcriptions/

Pipeline

1. Word-Level Alignment (align_swda_mstrans.py)

The main alignment script that:

• Loads SWDA annotations and MS-State transcripts
• Normalizes text for better matching (handles variants like "uh-huh", "um-hum", etc.)
• Performs sequence alignment using difflib.SequenceMatcher
• Prioritizes longer consecutive match sequences over isolated single-word matches
• Automatically detects and corrects speaker channel flips (84 conversations had flipped channels)
• Outputs word-level alignments to aligned_words/

Key features:

• Text normalization for common backchannel variations
• Lookahead logic to skip short matches in favor of longer sequences
• Speaker flip detection based on alignment quality (match rate < 50% threshold)
• Processes all 1,155 conversations with both SWDA and transcript data

Output: aligned_words/aligned_conv_{conv_num}.csv containing:

• speaker: A or B
• alignment_type: match, mismatch, insertion, or deletion
• swda_word: word from SWDA transcript
• trans_word: word from MS-State transcript
• utterance_index, subutterance_index: SWDA utterance IDs
• act_tag: Dialog act tag
• conv_no: Conversation number
• turn_id: MS-State turn ID
• start, end: Word timing in seconds

2. Turn-Level Aggregation (process_aligned.py)

Post-processes word-level alignments to create turn-level utterances:

• Drops deletion alignments (words only in SWDA, not in transcript)
• Sorts by timestamp
• Fills missing act tags using forward/backward/nearest neighbor passes
• Groups words by utterance to reconstruct full turns
• Outputs turn-level data to aligned_turns/

Output: aligned_turns/aligned_turns_{conv_num}.csv containing:

• speaker: A or B
• utterance_index, subutterance_index: SWDA utterance IDs
• start, end: Utterance timing in seconds
• turn_id: MS-State turn ID
• transcript_swda: Full utterance text from SWDA
• transcript_ms: Full utterance text from MS-State
• act_tag: Dialog act tag (simplified DAMSL tags)

3. Coarse Tag Conversion (convert_to_coarse_tags.py)

Converts fine-grained dialog act tags into 4 coarse categories for simplified analysis/question-answer pair turn extractions:

• Handles continuation markers: when act_tag is "+", replaces it with the previous turn's tag from the same speaker
• Maps fine-grained tags to coarse categories:
  • question: tags starting with "q" (e.g., qy, qw, qh)
  • statement: tags starting with "s" (e.g., sd, sv, sv^d)
  • answer: tags starting with "a", "n", or "b" (e.g., aa, ny, bh)
  • other: all remaining tags (e.g., h, fp, x)
• Outputs coarse-tagged data to coarse_tags/

Output: coarse_tags/aligned_turns_{conv_num}.csv containing all columns from turn-level files plus:

• act_tag_merge: Coarse dialog act category (question, answer, statement, other)

Usage

Setup

# Clone and setup SWDA data
bash download.sh

# Install dependencies
pip install -r requirements.txt

Run Alignment

# Step 1: Word-level alignment
python align_swda_mstrans.py

# Step 2: Turn-level aggregation
python process_aligned.py

# Step 3: Convert to coarse tags (optional)
python convert_to_coarse_tags.py

Processing Time

• Word-level alignment: ~5 minutes on Intel Core Ultra 7 155U
• Turn-level aggregation: ~3 minutes on Intel Core Ultra 7 155U
• Coarse tag conversion: ~1 minute on Intel Core Ulttra 7 155U

Results

• Aligned all 1,155 conversations from SWDA
• 84 conversations had speaker channels flipped (automatically corrected)
• ~99.7% word coverage on average (missing act_tags tracked in logs)

Alignment Quality

Most conversations have >99% of words successfully aligned with act tags. Conversations with most unfilled act_tags:

• 2768: 46/2121 words (2.2%)
• 2884: 45/2377 words (1.9%)
• 2386: 42/2215 words (1.9%)

See logs.txt for complete processing details.

Files

• align_swda_mstrans.py: Main word-level alignment script
• process_aligned.py: Turn-level aggregation script
• convert_to_coarse_tags.py: Coarse tag conversion script
• compare_alignment.py: Comparison of alignment algorithms (difflib's Ratcliff-Obershelp algorithm vs. Needleman–Wunsch algorithm)
• download.sh: Setup script for SWDA data
• flipped_channels.txt: List of conversations with flipped speaker channels
• logs.txt: Processing logs and statistics
• aligned_words/: Word-level alignment outputs (1,155 files)
• aligned_turns/: Turn-level aggregated outputs (1,155 files)
• coarse_tags/: Coarse-tagged turn-level outputs (1,155 files)

License

See LICENSE file for details.

References

• Switchboard Dialog Act Corpus
• DAMSL Annotation Manual
• Mississippi State Switchboard resegmentation project
• My old code - a monstrosity that would probably cause most people to cringe, but it was surprisingly useful in giving me better, more focused ideas to prompt Claude with and reminding me of pitfalls I had run into previously.