Structural Metadata Annotation of Speech Corpora: Comparing Broadcast News and Broadcast Conversations

1
Structural Metadata Annotation of Speech
Corpora Comparing Broadcast News and Broadcast
Conversations

• Jáchym Kolár Jan Švec

University of West Bohemia in Pilsen, Czech
Republic
2
Talk Overview

• Structural metadata annotation
• Speech data
• Statistics about fillers
• Statistics about edit disfluencies
• Statistics about sentence-like units
• Summary

3
Structural Metadata Extraction

• Metadata Extraction (MDE) research started as
  part of DARPA EARS program
• Metadata annotation scheme for MDE introduced
  by LDC (originally for English ? we have extended
  it to Czech)
• ULTIMATE GOAL of MDE
• Automatic conversion of raw speech recognition
  output to forms more useful to humans and
  downstream automatic processes

4
MDE Annotation Subtasks

• Boundaries of syntactic/semantic units (SUs)
• Statements, Interrogatives, Incompletes
• Coordination breaks, Clausal breaks
• Non-content words (fillers)
• Filled pauses (FPs)
• Discourse markers (DMs)
• Speech disfluencies (edits)
• Deletable regions (DelRegs), Interruption points,
• Explicit editing terms, Corrections

5
MDE Annotation Example
but I you know really pre- uh prefer this
form of of um presentation/. she Sheila
told me on Tuesday no on Wednesday/, she
didnt/. so lets move on/, because we dont
have uh dont have time/. well do you like
this this example/?
but I you know really pre- uh prefer this form
of of um presentation she Sheila told me
on Tuesday no on Wednesday she didnt so
lets move on because we dont have uh dont
have time well do you like this this example
6
Goal of This Paper

• Analyse and compare two Czech MDE corpora from
  different domains in terms of metadata statistics
• Compare Czech Broadcast News (BN) vs. Broadcast
  Conversations (BC)
• Also compare Czech and English MDE corpora
  English Broadcast News and Conversational
  Telephone Speech (CTS)

7
Czech Broadcast News Data

• News from 3 TV channels and 4 radio stations
• Both public and commercial broadcast companies
• Differing in presentation style
• 26 hours of transcribed speech
• 300 speakers
• Speech recordings and verbatim transcripts
  publicly available from LDC

8
Broadcast Conversation Data

• 52 recordings of a Czech radio talk show
  Radioforum
• 24 hours of transcribed speech
• 100 speakers
• 1-3 guests spontaneously answer questions asked
  by 1-2 interviewers
• Mostly political debates
• Currently being extended by additional 20
  recordings (10 hours)

9
Statistics about Fillers

• Filled pauses more frequent in Czech Broadcast
  Conversations (3.8 of words) than in News (0.5)
• English MDE CTS 2.2, BN 1.4
• Discourse markers also more frequent in Czech
  Conversations (1.6) than in News (0.1)
• English MDE CTS 4.4, BN 0.5

10
Statistics about Edit Disfluencies

• Deletable regions 2.8 of words in
  Conversations and 0.2 in News
• English MDE 5.4 in CTS and 1.5 in BN
• Percentage of disfluencies having a correction
  larger in News (94.6) than in Conversations
  (83.8)
• Explicit editing terms rare in both corpora
• occur just at 4 of disfluencies

11
POS Analysis of Edit Disfluencies

• Tagged the Czech corpora employing an automatic
  POS tagger
• Czech uses structured tags with 15 positions
• we only used the first position distinguishing
  10 basic POS
• Computed and compared three POS distributions
• Whole corpus
• Deletable regions only
• Corrections only

12
POS Analysis of Edit Disfluencies
13
Statistics about SUs

• Average SU length Conversations (14.5 words)
  shows longer SUs than News (13.0)
• English BN (12.5) similar to Czech, but CTS shows
  much shorter SUs (7.0) than Broadcast
  Conversations
• SU-internal breaks (clausal and coordination)
  more frequent in Conversations than in News
• (49 vs. 31 of all SU symbols)
• ? Complex and compound sentences more common in
  spontaneous conversations than in prearranged news

14
Summary

• Broadcast Conversations contain significantly
  more fillers and disfluencies than News
• Conversations also show longer SUs and contain a
  higher number of complex sentences than News
• Deletable regions and corrections in both corpora
  show different POS distributions in comparison
  with the general POS distributions
• We plan to make Czech MDE corpora publicly
  available

15
Structural Metadata Annotation of Speech
Corpora Comparing Broadcast News and Broadcast
Conversations

• Jáchym Kolár Jan Švec

University of West Bohemia in Pilsen, Czech
Republic