National Institute of Standards and Technology

1
2000 TREC-9 Spoken Document Retrieval
Track http//www.nist.gov/speech/sdr2000
John Garofolo, Jerome Lard, Ellen Voorhees
National Institute of Standards and
Technology Information Technology Laboratory
2
SDR 2000 - Overview

• SDR 2000 Track Overview, changes for TREC-9
• SDR Collection/Topics
• Technical Approaches
• Speech Recognition Metrics/Performance
• Retrieval Metrics/Performance
• Conclusions
• Future

3
Spoken Document Retrieval (SDR)

• Task
• Given a text topic, retrieve ranked list of
  relevant excerpts from collection of recorded
  speech
• Requires 2 core technologies
• Speech Recognition
• Information Retrieval

• First step towards
• multimedia information access.
• Focus is on effect of recognition accuracy on
  retrieval performance
• Domain Radio and TV Broadcast News

4
SDR Evaluation Approach

• In the TREC tradition
• Create realish but doable application task
• Increase realism (and difficulty) each year
• NIST creates
• infrastructure test collection, queries, task
  definition, relevance judgements
• task includes several different control
  conditions recognizer, boundaries, etc.
• Sites submit
• speech recognizer transcripts for benchmarking
  and sharing
• rank-ordered retrieval lists for scoring

5
Past SDR Test Collections
6
Past SDR Evaluation Conditions
7
SDR 2000 - Changes from 1999

• 2000
• evaluated on whole shows including non-news
  segments
• 50 ad-hoc topics in two forms short description
  and keyword
• 1 baseline recognizer transcript sets (NIST/BBN
  B2 from 1999)
• story boundaries unknown (SU) condition is
  required
• recognition and use of non-lexical information

• 1999
• evaluated on hand-segmented news excerpts only
• 49 ad-hoc-style topics/metrics
• 2 baseline recognizer transcript sets (NIST/BBN)
• story boundaries known (SK) focus and
  exploratory unknown (SU) conditions

8
SDR 2000 - Test Collection

• Based on the LDC TDT-2 Corpus
• 4 sources (TV ABC, CNN, Radio PRI, VOA)
• February through June 1998 subset, 902 broadcasts
  
• 557.5 hours, 21,754 stories, 6,755 filler and
  commercial segments (55 hours)
• Reference transcripts
• Human-annotated story boundaries
• Full broadcast word transcription
• News segments hand-transcribed (same as in 99)
• Commercials and non-news filler transcribed via
  NIST ROVER applied to 3 automatic recognizer
  transcript sets
• Word times provided by LIMSI forced alignment
• Automatic recognition of non-lexical information
  (commercials, repeats, gender, bandwidth,
  non-speech, signal energy, and combinations)
  provided by CU

9
Test Variables

• Collection
• Reference (R1) - transcripts created by LDC human
  annotators
• Baseline (B1) - transcripts created by NIST/BBN
  time-adaptive automatic recognizer
• Speech (S1/S2) - transcripts created by sites
  own automatic recognizers
• Cross-Recognizer (CR) - all contributed
  recognizers
• Boundaries
• Known (K) - Story boundaries provided by LDC
  annotators
• Unknown (U) - Story boundaries unknown

10
Test Variables (contd)

• Queries
• Short (S) - 1 or 2-phrase description of
  information need
• Terse (T) - keyword list
• Non-Lexical Information
• Default - Could make use of automatically-recogniz
  ed features
• None (N) - no non-lexical information (control)
• Recognition language models
• Fixed (FLM) - Fixed language model/vocabulary
  predating test epoch
• Rolling (RLM) - Time-adaptive language
  model/vocabulary using daily newswire texts

11
Test Conditions

• Primary Conditions (may use non-lexical side
  info, but must run contrast below)
• R1SU Reference Retrieval, short topics, using
  human-generated "perfect" transcripts without
  known story boundaries
• R1TU Reference Retrieval, terse topics, using
  human-generated "perfect"
• B1SU Baseline Retrieval, short topics, using
  provided recognizer transcripts without known
  story boundaries
• B1TU Baseline Retrieval, terse topics, using
  provided recognizer transcripts without known
  story boundaries
• S1SU Speech Retrieval, short topics, using own
  recognizer without known story boundaries
• S1TU Baseline Retrieval,terse topics, using
  provided recognizer transcripts without known
  story boundaries
• Optional Cross-Recognizer Condition (may use
  non-lexical side info, but must run contrast
  below)
• CRSU-ltSYS_NAMEgt Cross-Recognizer Retrieval,
  short topics, using other participants'
  recognizer transcripts without known story
  boundaries
• CRTU-ltSYS_NAMEgt Cross-Recognizer Retrieval,
  terse topics, using other participants'
  recognizer transcripts without known story
  boundaries
• Conditional No Non-Lexical Information Condition
  (required contrast if non-lexical information is
  used in other conditions)
• R1SUN Reference Retrieval, short topics, using
  human-generated "perfect" transcripts without
  known story boundaries, no non-lexical info
• R1TUN Reference Retrieval, terse topics, using
  human-generated "perfect" transcripts without
  known story boundaries, no non-lexical info
• B1SUN Baseline Retrieval, short topics, using
  provided recognizer transcripts without known
  story boundaries, no non-lexical info
• B1TUN Baseline Retrieval, terse topics, using
  provided recognizer transcripts without known
  story boundaries, no non-lexical info
• S1SUN Speech Retrieval, short topics, using own
  recognizer without known story boundaries, no
  non-lexical info
• S1TUN Speech Retrieval, terse topics, using own
  recognizer without known story boundaries, no
  non-lexical info
• S2SUN Speech Retrieval, short topics, using own
  second recognizer without known story boundaries,
  no non-lexical info
• S2TUN Speech Retrieval, terse topics, using own
  second recognizer without known story boundaries,
  no non-lexical info

12
Test Topics

• 50 topics developed by NIST Assessors using
  similar approach to TREC Ad-Hoc Task
• Short and terse forms of topics were generated
• Hard Topic 125 10 relevant stories
• Short Provide information pertaining to
  security violations withinthe U. S. intelligence
  community. (.024 average MAP)
• Terse U. S. intelligence violations (.019
  average MAP)
• Medium Topic 143 8 relevant stories
• Short How many Americans file for bankruptcy
  each year? (.505 avg MAP)
• Terse Americans bankruptcy debts (.472 average
  MAP)
• Easy Topic 127 11 relevant stories
• Short Name some countries which permit their
  citizens to commit suicide with medical
  assistance. (.887 average MAP)
• Terse assisted suicide (.938 average MAP)

13
Test Topic Relevance
14
Topic Difficulty
15
Participants

Full SDR (recognition and retrieval) Cambridge
University, UK LIMSI, France Sheffield
University, UK
16
Approaches for 2000

• Automatic Speech Recognition
• HMM, word-based - most
• NN/HMM hybrid-based, Sheffield
• Retrieval
• OKAPI Probabilistic Model - all
• Blind Relevance Feedback and parallel corpus BRF
  for query expansion
  - all
• Story boundary unknown retrieval
• passage windowing, retrieval and merging - all
• Use of automatically-recognized non-lexical
  features
• repeat and commercial detection - CU

17
ASR Metrics

• Traditional ASR Metric
• Word Error Rate (WER) and Mean Story Word Error
  Rate (SWER) using SCLITE and LDC ref transcripts
• WER word insertions word deletions word
  substitutions
• total words in reference
• LDC created 2 Hub-4 compliant 10-hour subsets for
  ASR scoring and analyses (LDC-SDR-99 and
  LDC-SDR-2000)
• Note that there is a 10.3 WER in the collection
  human (closed caption) transcripts

Note SDR recognition is not directly comparable
to Hub-4 benchmarks due to transcript quality,
test set selection method, and word mapping
method used in scoring
18
ASR Performance
19
IR Metrics

• Traditional TREC ad-hoc Metric
• Mean Average Precision (MAP) using TREC_EVAL
• Created assessment pools for each topic using top
  100 of all retrieval runs
• Mean pool size 596 (2.1 of all segments)
• Min pool size 209
• Max pool size 1309
• NIST assessors created reference relevance
  assessments from topic pools
• Somewhat artificial for boundary unknown
  conditions

20
Story Boundaries Known Condition

• Retrieval using pre-segmented news stories
• systems given index of story boundaries for
  recognition with IDs for retrieval
• excluded non-news segments
• stories are treated as documents
• systems produce rank-ordered list of Story IDs
• document-based scoring
• score as in other TREC Ad Hoc tests using
  TREC_EVAL

21
Story Boundaries Known Retrieval Condition
22
Unknown Story Boundary Condition

• Retrieval using continuous speech stream
• systems process entire broadcasts for ASR and
  retrieval with no provided segmentation
• systems output a single time marker for each
  relevant excerpt to indicate topical passages
• this task does NOT attempt to determine topic
  boundaries
• time-based scoring
• map to a story ID (dummy ID for retrieved
  non-stories and duplicates)
• score as usual using TREC_EVAL
• penalizes for duplicate retrieved stories
• story-based scoring somewhat artificial but
  expedient

23
Story Boundaries Unknown Retrieval Condition
24
SDR-2000 Cross-Recognizer Results
25
Conclusions

• ad hoc retrieval in broadcast news domain appears
  to be a solved problem
• systems perform well at finding relevant passages
  in transcripts produced by a variety of
  recognizers on full unsegmented news broadcasts
• performance on own recognizer comparable to human
  reference
• just beginning to investigate use of non-lexical
  information
• Caveat Emptor
• ASR may still pose serious problems for Question
  Answering domain where content errors are fatal

26
Future for Multi-Media Retrieval?

• SDR Track will be sunset
• Other opportunities
• TREC
• Question Answering Track
• New Video Retrieval Track
• CLEF
• Cross-language SDR
• TDT Project

27
TREC-9 SDR Results, Primary Conditions
28
TREC-9 SDR Results, Cross Recognizer Conditions