(QRTR) Specification for Chinese Broadcast Data
Contents
1. A Mm A 13 7 1 3 Numbers EA 4 13 7 1 4 Proper Nou A 13 7 1 5 Contractions and Acronyms 000000nnnn 13 7 1 66 Spokeg ttters M A 13 7 1 7 Talli 14 7 2 Disfluent Sp 00 00 0000 14 7 2 1 Filled Pauses and Hesitation SOUNCS ccseeeeeeeeeeeeeeeeeeeeees 14 7 2 2 ar 1S NMI i a ea eiee aaea 15 7 23 Mispronounced WOrdGS cccccccceeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeaaeeeeees 15 7 2 4 Idiosyncratic VVOPS s nins ns na ns s n h nsd na g s n bindi sa 15 7 3 Speaker Errors and Non standard Usage ccccccccceeeeeeeeeeeeeeeeeeeees 15 7 4 Foreign Languages and Dialects 0000eeeeeeeeeeeeeeeeerrrrrrrrrrr 16 7 4 1 is 16 7 4 2 Dialects ia tr ssi a aid ku kraka fey Terre ter peo ye ee 16 7 5 Background and Speaker Noise ccccceceeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeees 17 7 6 Hard to Understand Regions 000 0000 eeen 17 ak minal POMS seon ra a SSR ROS ROS A 17 Appendix 1 Recommended Strategy cccccccccccecceceeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeees 18 LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 2 of 18 Version 2 August 30 2006 1 Introduction and Overview The goal of quick rich transcription QRTR for broadcast news and broadcast conversation is to produce a verbatim time aligned transcript with minimal but useful markup QRTR also identifi2. Occasionally a speaker will make up a new word on the spot These are not the same as slang words but rather are words that are unique to the speaker in that conversation If annotators encounter an idiosyncratic word they should transcribe it to the best of their ability and mark it with an asterisk This is extremely rare in Chinese 7 3 Speaker Errors and Non standard Usage Annotators should not correct grammatical errors e g seen him for saw him The words must be transcribed as spoken The same goes for non standard usage or mis used words e g qe pas If the speaker says Hi A 3E7K but means HH 7K KA transcribe it as HARE ZK Annotators should transcribe exactly what is spoken not what they expect to hear or what they consider correct speech LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 15 of 18 Version 2 August 30 2006 7 4 Foreign Languages and Dialects 7 4 1 Foreign Languages Portions of speech in any language other than the target language are annotated using the lt language gt text lt language gt convention to indicate the language and to transcribe the words that are spoken in that language Note that southern Chinese dialects are treated as foreign languages These include Cantonese Wu Xiang Gan Minnan Hakka For instance FREST ATA lt English gt notebook lt English gt lit lt Cantonese gt 4 lt Cantonese gt If the annotator does not know the nam
3. 1 4 Recognizing SU Boundaries It can sometimes be difficult to determine where a sentence unit boundary exists and when to place two clauses within the same SU Annotators should rely primarily on the meaning conveyed by the utterance and apply SU breaks in accordance with the rules described in these guidelines However annotators may sometimes rely on prosodic features like sentence intonation or pauses to determine where to place an SU boundary In practice SU boundaries tend to occur at the ends of fragments simple sentences and complex sentences Speaker1 Ya AMEN AE AN EY EAN EMI R Speaker1 f vi UIE t pees I eK A RAR R SIAR HK Speaker2 KRIE Speaker1 fh SoA Speaker1 Fe 1435 SEN f lu I AE BF TH BAR N Ee Y 18 ER Wee ER AA CL ate a a PERE In Chinese we frequently see a subject introduced in the first sentence of a narrative and then dropped repeatedly from subsequent sentences In such cases annotators should rely on the context of the utterance to make a decision about where to put SU boundaries For instance Speaker1 Hse Jn RBB EA aE EA AIA LI M WH Zeta AWLP UR GELS Speaker1 JLIU VES ATS Bin BRT SP A Eh BEJM Speaker1 tit A it 2 A FREE H R Below are some additional rules of thumb annotators can follow in deciding where to put SU boundaries 1 Rely on temporal expressions If two clauses share a su
4. a speaker interrupts him herself and then restructures the utterance and continues speaking on the same topic an incomplete SU exists In other cases the speaker may trail off at the end of his her turn and abandons the utterance completely without restructuring it or continuing along the same lines For instance Speaker1 AI MRAIRS Aa FRR ZINE PLY Speaker1 AIXA Speaker1 IUR RAIA EN fi 5038 BH SE MK The other frequent case of incomplete SU occurs when one speaker s turn is cut short by an interruption from the other speaker as in the following Speaker2 HIS kk fg Kt Speaker2 M471 jm HIS PEE s HAR 95 FP Speaker2 4 ZK Jm Ht VC aL ES ES SE EAA I G Ah Speaker1 HW Kt IAF IK SARA EIT LE T J RA hth Speaker2 IN ie RHE ek E N FEY Speaker1 Xf QRTR punctuation guidelines require annotators to use the double dash at the ends of incomplete SUs LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 8 of 18 Version 2 August 30 2006 Speaker1 ER Speaker1 tbh ihr fh EA T Be careful not to confuse incomplete SUs with sentence fragments that express a complete thought for instance a response to a question that is expressed as a phrase rather than a complete sentence Sentence fragments that express a complete thought and show no signs of being caused by an interruption or by the speaker simply trailing off should be labeled as statement SUs 4
5. agony Wa ai2 emphasis 2 RII ai disagreement denying I lt aid regret IR ai4 disappointment i ha1 triumphant My hatha1 triumphant I hai1 sadness regret i heil drawing attention w i hel surprise I hng5 _ dissatisfaction I WE n2 ng2 query IR n3 ng3 out of expectation nal n4 ng4 answer 3 H E 01 understanding Jam O1yo1 surprise o2 half belief half doubt RAE 04 understanding IR pei1 discarding scolding IEE wa1 wa3 surprise a yot slight surprise men 7 2 Disfluent Speech Regions of disfluent speech are particularly difficult to transcribe Speakers may stumble over their words repeat themselves utter partial words restart phrases or sentences and use hesitation sounds For purposes of QRTR annotators should not spend too much time trying to precisely capture difficult sections of disfluent speech but should make their best effort to transcribe what they hear after listening to the segment once or twice then move on 7 2 1 Filled Pauses and Hesitation Sounds Filled pauses are non word sounds that speakers employ to indicate hesitation or to maintain control of a conversation while thinking of what to say next The spelling of filled pauses is not altered to reflect how the speaker pronounces the LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 14 of 18 Version 2 A
6. be marked as mispronunciation For example Speaker says zil dao4 transcription should be kept as i LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 16 of 18 Version 2 August 30 2006 Speaker says lan2fang1 transcription should be kept as PA 77 Speaker says feng4fang2 transcription should be kept as J 7 5 Background and Speaker Noise Transcribers are not required to specially label background noise or sound effects Note however the convention for indicating long periods of non speech within or outside an SU segment Section 3 3 Speaker produced noise is identified with one of the following four tags laugh cough sneeze lipsmack 7 6 Hard to Understand Regions Sometimes an audio file will contain a section of speech that is difficult or impossible to understand In these cases annotators should use double parentheses to mark the region of difficulty t may be possible to take a guess about the speaker s words In these cases annotators transcribe what they think they hear and surround the area of uncertain transcription with double parentheses Speakerl IARA O If an annotator is truly mystified and can t at all make out what the speaker is saying s he uses empty double parentheses to surround the untranscribed region For example Speakerl 0 Do not skip the region 7 7 Final Pointers 1 Transcribe what you hear not what you think is correct 2 Do not add i
7. broadcast audio the end of one segment is also the beginning of the next Segments are also classified by type and subtype We identify three kinds of segments in the QRTR task Sections Turns and Sentence Units These are arranged hierarchically sections contain turns turns contain sentences It is suggested that annotators begin segmentation by identifying the most fine grained segment type sentence units SUs SU boundaries frequently occur at natural boundaries in the audio pauses breaths speaker turns which makes segmentation easier This is not always the case especially for complex or atypical SUs and annotators will need to fine tune some SU boundaries once they have completed transcription As segments are created XTrans will prompt the annotator to supply SpeakerlD information and the annotator will also indicate section storyand commercial boundaries as encounter them The sections that followprovide detailed information about each step of the process Annotators should note that segmentation in XTrans can be done with the keyboard only with the mouse only or with a combination of both After you ve become familiar with basic XTrans functionality you will find that using only the keyboard is both faster and more intuitive than using the mouse 3 2 Timestamping the Audio Timestamps are required for all segments In XTrans annotators create a timestamped segment simply by marking the appropriate region of audio in
8. the waveform display then inserting the selected segment Timestamps are designated in seconds rounded to the nearest thousandth of a second Note that while XTrans does not show start end timestamps within the transcript display the waveform display includes a color coded horizontal bar representing each segment along with its start time end time and duration Because broadcast speech recordings use a single audio channel segments occur one right after the other in direct succession and typically without 1 Detailed instructions for using the XTrans toolkit are available in Using XTrans for Broadcast Transcription A User Manual distributed with the XTrans package and available from LDC s transcription website http www ldc upenn edu Projects Transcription LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 4 of 18 Version 2 August 30 2006 intervening periods of unsegmented audio silence Small gaps in the succession of segments should indicate an untranscribed event like a commercial music sound effects or background noise All speech and other material to be transcribed must be segmented Timestamps should always be placed in between words not inside of them or at the very edges of words where speech sounds could be truncated Good places to insert timestamps are during pauses breaths or other non speech events which typically occur at sentence unit SU boundaries Finally it is critical that
9. the time and the audio event are properly aligned so that the words transcribed within each segment match the speech associated with thatsegment 3 3 What to Segment All broadcast speech must be segmented and classified into sections news reports conversational segments or non news News reports and conversational segments must also be segmented into SUs with speakerlDs added Non news sections like commercials should not be segmented into smaller units or labeled for speakerlD and they should not be transcribed Very brief under 0 5 seconds periods of silence music background noise or other types of non speech that occur while someone is speaking should simply be included within that SU segment or split between two adjoining speaker SU segments No other treatment is necessary Lengthy segments of non speech like sound effects that interrupt a speaker s turn or that come in between speaker turns should be separated out and left unsegmented Note that annotators should make an effort to leave SU segments intact that is avoid splitting a single SU into multiple segments even when it includes a lengthy pause 3 4 Segmenting Overlapping and Simultaneous Speech In broadcast audio overlapping speech from two or more speakers is a relatively frequent occurrence Although broadcast files contain a single audio channel within XTrans each unique speaker in a file is assigned a separate virtual channel Transcribers can simply create
10. Quick Rich Transcription QRTR Specification for Chinese Broadcast Data XTrans Format Version Version 2 August 30 2006 Linguistic Data Consortium http www idc upenn edu GALE Transcription LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 1 of 18 Version 2 August 30 2006 1 should SA A A SA Ai SAGA ga 3 2 SE sasssa a a i A A Rg 3 3 Segmentation Task innan 4 3 1 MA ea a bala nn 4 32 Timestamping the AU rr ee eee 4 eo SMVINAL IO SOOMMICME ace ete ec cecl eed cuales cenit ara rr 5 3 4 Segmenting Overlapping and Simultaneous Speech 5 4 Sentence Units SU cececececececeeeeeeeeeneeeeeeeeeeeeeeeneeeeeeeneneeeeeseeeeeeenenseene 6 4 1 1 Statement S US vcccsncreccesnesssencuscadsustereuvaczesenaeanertaeeaheseseueaereset Ree eneee Ret 6 4 1 2 Q eston SS assist 7 4 1 3 Incomplete SUS AM 8 4 1 4 Recognizing SU Boundaries fli 9 5 Identifying Section Boundaries eee ann ean ae 10 6 Speaker Identification ge eee aaae aenar nenea 11 6 1 Speaker Type cccccccccccceees MT E 11 6 2 Names and ldentifiers gffiaag a rr MED ceeeeeeeeeeees 11 6 3 Native and Non native Speakers a ads ranna 12 T Transcription enesenn N a 12 7 1 Orthography and Spelling gp D 0000 12 7 1 1 Spelling ooe 12 7 1 2 Punctuation
11. aker information dialog box use tab amp arrow keys to select options 8 create next segment repeat 4 7 To create segment for same speaker first select speaker in speaker panel then repeat steps 4 6 9 assign section boundary Ctrl I Ctrl s 10 assign SU type Ctrl I Ctrl U Ctrl 11 transcribe the segment 12 save your work frequently Alt F Alt S 13 repeat steps 4 12 14 save and exit 8 Some transcribers prefer to fully segment the file the go back and transcribe it while others prefer to transcribe as they segment LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 18 of 18 Version 2 August 30 2006
12. and or syntactic content LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 6 of 18 Version 2 August 30 2006 Speakerl FRM AA AA Ber A PE Speakerl APEI H A r F RI Speakerl Will AE 3x lt English gt A a FHFEW ASS AI jil Speakerl MER Bie H E Speaker2 St HEX 4 1 1 1 Backchannel SUs A backchannel is a word or phrase that provides feedback to the dominant speaker indicating that the non dominant speaker is still paying attention to the conversation In QRTR backchannels are treated as statement SUs When a speaker chains together several backchannels in succession annotators tag them as a single statement SU For instance Speakerl PTLABU ZEAE eit EF RMAF ANAT Speaker2 HA Speakert FY 7 RILA JJL al Speaker1 Fl 2 EHER Speaker1 tk P LEEK AEZUAY a ERES He IT BY Speaker2 Wiki Speaker2 XT XTX Long statements with multiple verbs are very common in Chinese In these cases annotators should use their judgment about whether the verb change warrants a new statement SU See Section 4 1 4 for additional guidelines on determining SU boundaries 4 1 2 Question SUs The question label should be used for a complete sentence that functions as an interrogative The expected end of sentence punctuation for a question is a question mark In Chinese one of several sentence final words may indicate a question SU su
13. bject but have different temporal information they should be treated as separate SUs LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 9 of 18 Version 2 August 30 2006 2 When the subject changes from one clause to the next create a new SU 3 If subjects are overtly expressed in two adjacent clauses treat the clauses as separate SUs 4 For quotations there is one SU break after the quoted material However if the quotation itself contains multiple sentences break up the quote into separate SUs 5 Identifying Section Boundaries The QRTR task also calls for identification of section boundaries A section is a topically contiguous segment of the broadcast Sections begin at SU boundaries At the beginning of each new section annotators simply insert the appropriate section label Consecutive sections of the same type should receive separate section boundary labels except in the case of consecutive commercials and other untranscribed segments which should be grouped together as a single untranscribed section All audio in a speech file must be assigned to a section We recognize three section types e Reports include typical talking head news broadcast with an anchor reading the news This may also include broadcasts from reporters in the field News reports may be of any length as long as they constitute a complete cohesive news report on a particular topic Notethat single news stories may discuss mo
14. ch as 15 We IE If these interrogative particles are not present other words or phrases may signal the presence of a question SU for instance tt WE YE ATTA etc A tag question is a phrase added to the end of an utterance that invites the listener to give feedback Tag questions usually do not stand alone as a question but rather form a complete question with the previous utterance Rhetorical questions should also receive a Question SU label Speakert HIS if EVIE VR AS eh AE IK HE A YTS LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 7 of 18 Version 2 August 30 2006 Speaker2 fk is A r LEMMAS A KT Speaker2 fk r A E NE PELE WEI He NR Speaker3 Al Ay fk BLA RAN DST IL AN DST DL TE HE Speaker3 f J A w Speaker4 RAE MAX Speakerd iX M Ne lk r ERAT PE ARE At The question SU label should only be used when the utterance is clearly asking a question or functioning as a tag or rhetorical question If you are unsure whether the SU is functioning as a statement or a question you should label it as a statement 4 1 3 Incomplete SUs When an utterance does not constitute a grammatically complete sentence and does not express a complete thought it is labeled as an incomplete sentence unit In standard writing this kind of incomplete SU might be followed by double dashes or ellipses Incomplete SUs frequently occur in two situations When
15. cter yl 1 A yi bai 100 F shi 10 Ses er shiyi 21 7 1 4 Proper Nouns No special markup is required for proper nouns Note however that spelling of names should be consistent within the transcript and should match the spelling of the name n withinsthe assigned speaker ID instance if the speaker ID uses the transliteration Hu Jintao th transcript should use H45 4 when that name is spoken not iHn ZE or some other form 7 1 5 Contractions and Acronyms These phenomena rarely if ever occur in Chinese and no special guidelines apply 7 1 6 Spoken Letters Transcribers may come across English spoken letters Letters pronounced as a sequence of individual letters should be written in English as they are pronounced marked with a tilde with no space between the letters RS ARG IRN TAA I got an A in today s quiz MERJA EN Af BUSH His name is Bush B U S H LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 13 of 18 Version 2 August 30 2006 FBI CEO 7 1 7 Interjections The following standardized spellings are used to transcribe interjections Transcribers may findalot of variations among speakers The list below is for reference Interjections do not require any special markup bal a1 surprise praise bil a2 questioning bal a3__ disbelief bal a4 answer surprise praise I WE ail dissatisfaction IEF ailyal surprise complaining Mt ailyol surprise
16. e of the language or what is being said they should use the tag lt foreign gt instead of the language name lt foreigm Ih AR PUL lt foreign 7 4 2 Dialects Annotators will frequently encounter non Putonghua dialect in the broadcast conversation programs such as Henan Hua Shandong Hua Sichuan Hua Yunnan Hua Portions of speech in northern Mandarin dialect should be surrounded with a special non Putonghua marker lt non PTH gt text lt non PTH gt to indicate the use of colloquial dialect The words should be transcribed as spoken using standard Chinese orthographic conventions If the conversationsswitches back and forth between Putonghua and dialect mark the dialectportions using the convention described above as in PTH text lt non PTH gt text lt non PTH gt PTH text SU segmentation is unaffected by the presence of non PTH speech A single SU segment may containall PTH all non PTH or a mix of both In the following example the speaker is speaking Sichuan and Wuhan dialect respectively lt non PTH gt ik MIR te A BE HL lt non PTH gt lt non PTH gt NP AWk lt non PTH gt Putonghua pronounced with a dialectal accent should still be treated as Putonghua Do not transcribe any accent features for example missing retroflex and confusion between lateral and nasal etc but rather use the standard orthography No special mark up is needed and dialect pronunciation should not
17. es for the various transcription tasks http Awww idc upenn edu Projects Transcription 2 Data These guidelines pertain to data in the following genres e Broadcast News BN consisting of talking head style news broadcasts from radio and or television networks e Broadcast Conversation BC consisting of talk shows interviews roundtable discussions and other interactive style broadcasts from radio and or television networks Data is divided into files which typically correspond to a recording of one broadcast from a single program Files are typically 30 to 60 minutes in duration though they may be of any length Files come from a range of radio television satellite and web broadcast sources from around the world Each show is pre LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 3 of 18 Version 2 August 30 2006 designated as BN or BC based on its characteristic content Note however that BN shows can sometimes contain stories that are conversational while BC shows can include hard news reports 3 Segmentation Task 3 1 Introduction Transcription begins with segmentation During the segmentation task annotators virtually chop an audio recording into smaller units that correspond to certain features of the broadcast for instance sentence units or speaker turns Each segment must be timestamped that is time aligned with the audio to identify where the segment starts and ends In most cases in
18. es some salient structural features of the broadcast and provides speaker identification The elements of a quick rich transcript include e verbatim transcription e time aligned section boundaries speaker turns and sentences segmentation e section and sentence type identification e speaker identification e standard treatment of common spoken phenomena Transcription begins with audio segmentation This involves timestamping structural boundaries including sections i e story transitions speaker turns and sentence units SUs Speakers are identified by name where possible or by a unique identifier and other speaker traits like sex are noted Once audio has been virtually segmented into smaller units annotators transcribe the content of each segment Special conventionsyare used to flag certain speech phenomena like disfluencies and mispronounced words Quality control checks verify the completeness and accuracy of segmentation and transcription QRTR differs from Quick Transcription QTR in that each sentence unit is timestamped and labeled for its type QRTR differs from careful transcription CTR in the amount of detail contained in the transcript markup the number of features identified the degree of accuracy and completeness of the transcript the amount of time taken to complete the file and the number of quality checks that are performed on the finished product Please see LDC s transcription website for links to guidelin
19. ors before they are considered complete 7 1 Orthography and Spelling 7 1 1 Spelling Transcribers should use standard Chinese orthography All files must be checked for typos after transcription is complete When in doubt about the orthography of a character or a proper name annotators should consult a standard reference like an online or paper dictionary world atlas or news website LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 12 of 18 Version 2 August 30 2006 7 1 2 Punctuation Annotators should include standard punctuation for ease of transcription and reading Acceptable punctuation is limited to the following Type Usage Symbol period end of sentence markup for Statement SUs question mark end of sentence markup for Question SUs double dash end of sentence markup for Incomplete SUs comma sentence internal used to aid readability Transcripts should not contain quotation marks exclamation marks colons semicolons single stand alone dashes or ellipses in transcribing Punctuation should be written as it normally appears in standard writing with no additional spaces around the punctuation marks 7 1 3 Numbers All numerals should be written out as complete words instead of number characters They should be written as spoken using the lt foreign gt or lt non PTH gt tag as needed see section 7 4 1 for more details written pronounced number chara
20. overlapping segments two or more distinct speakers using the normal XTrans functionality Overlapping segments are represented in the waveform display as overlapping horizontal bars as shown in the image below Note that using the mouse for segmentation makes it easier to leave unintended small gaps in consecutive segments of continuous speech Using the keyboard shortcuts for segmentation avoids this problem The LAG Listen All Gaps feature in XTrans allows annotators to review all unsegmented material in a file LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 5 of 18 Version 2 August 30 2006 4 26 0485 4 15 1050 4 29 7408 14 6357 FT ss CCC 4250 490 4350 4400 4450 4500 4550 so normal 4 Sentence Units SU Segmentation begins with identification of sentence unit boundaries A sentence unit SU is a natural grouping of words produced by a single speaker SUs have semantic cohesion that is they can have some inherent meaning when taken in isolation and they have syntactic cohesion that is they have some grammatical structure In written language sentences are usually designated by punctuation like periods or question marks When creating SU boundaries for spoken language our goal is to identify a semantically and syntactically cohesive group of words that constitute a reasonable sentence like unit Sentence units are the most basic kind of segment in the QRTR
21. re than one related topic When reports of similar content are adjacent to one another in a broadcast it is often difficult to tell where one story ends and the next begins Annotators should rely on audio cues speaker changes music pauses to inform their judgments When in doubt do not create a new section boundary e Conversations include highly interactive segments of a broadcast including roundtable discussions interviews call in segments debates and the like Some conversation sections are quite long and can contain multiple topics Annotators should create a new section boundary only at natural breaks in the flow of conversation for instance when there is a major shift in topic or when a new panelist joins a roundtable discussion If in doubt the annotator should avoid creating a new conversation boundary It may sometimes be difficult to tell the difference between a report and a conversational segment When in doubt annotators should use report e Non news text includes segments like commercials station identifications public service announcements promotions for upcoming shows and long musical interludes Note that non news sections are not segmented transcribed or further annotated in any way including speaker ID or SU segmentation Once a non news section has been identified and labeled it should be ignored for the rest of the transcription task If multiple non news sections follow one another within a transcript the
22. task Each SU should be contained within its own segment Segments should not contain multiple SUs and single SUs should not be divided across multiple segments We distinguish three types of SUs statements questions and incomplete sentences After identifying the boundaries of an SU and creating a corresponding segment annotators can use XTrans to assign the segment type In general the SU segment types are consistent with standard end of sentence punctuation used during transcription as follows Punctuation SU Type Symbol period end of sentence markup for Statement SUs question mark end of sentence markup for Question SUs double dash end of sentence markup for Incomplete SUs Annotators will note that standard punctuation typically includes commas as well For purposes of the QRTR task we do not identify an SU or sub SU unit that corresponds to a comma Commas may be added into transcripts for human readability but it should be understood that the existence of a comma does not imply the existence of a sentence unit See Section 7 1 2 for additional discussion of punctuation in QRTR transcripts The sections that follow provide language specific rules for identifying SUs of each type 4 1 1 Statement SUs Statements are declarative sentences or fragments and are usually punctuated by a period or exclamation point For instance 4 Note however that incomplete SUs may contain incomplete semantic
23. ugust 30 2006 word Instead there is a restricted set of filled pauses for each language with established spelling conventions For Chinese filled pauses are limited to WJ WE and Ii 7 2 2 Partial Words When a speaker breaks off in the middle of the word annotators transcribe as much of the word as can be made out In Chinese speakers almost always break off at the end of a character so annotators should simply type the character that they hear If the transcriber is sure that the character is part of a word rather than being spoken in isolation he she should put a dash directly after the character Note that this should be distinuished from a word which is repeated several times which should not get any special mark ups For example SEE SENA ERA E KPR RER DD There are all kinds of animals in af af africa 7 2 3 Mispronounced Words A plus symbol is used for obviously mispronounced words not regional or non standard dialect pronunciation Annotators should transcribe using the standard spelling and should not try to representthe pronunciation Just transcribe the word using the standard spelling adding the plus sign to signal that the word is pronounced incorrectly t pronounced as pu ALB fii Keep in mind that this symbol should only be used for obviously mispronounced words Dialect pronunciations or other common variants of words should not be marked as mispronunciations 7 2 4 Idiosyncratic Words
24. uld follow standard conventions The spelling of speaker IDs must be consistent within a broadcast file and wherever feasible across different broadcast files as well It is also important that the spelling of names within a transcript match the spelling of the name in within the speaker ID label For instance if the transcript uses the transliteration Osama bin Laden then the speaker ID should also use Osama not Usama When a speaker is not identified by name within a recording the speaker should be labeled with a unique numerical identifier e g speaker14 Each anonymous speaker is assigned a unique number that should be used for every instance of that speaker throughout the broadcast Anonymous speaker IDs cannot be re used for different speakers in the same file regardless of gender or speaker type The XTrans toolkit requires annotators to provide speaker ID for each SU annotation 7 Note that the LRS Listen Random Segment and LAS Listen All Segments functions in XTrans are helpful for verifying speakerlD assignment LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 11 of 18 Version 2 August 30 2006 6 3 Native and Non native Speakers In addition to labeling speaker type and name annotators also indicate when a speaker is non native that is when they use a language variety other than the target or when they speak the target language with a discernable foreign accent Targets for the c
25. urrent task are o Arabic Modern Standard Arabic MSA o Chinese Mainland Mandarin Chinese o English American English Speakers using other varieties dialects of these languages or speaking these languages with a heavy non standard accent for instance Cantonese accented Mandarin or British English should be marked as non native In the case of Chinese nearly all speakers will be native speakers of some regional variety rather than native speakers of Putonghua A native speaker of any Chinese dialect who is talking in Putonghua should be considered native for purposes of speakerlD labeling Do not mark native Chinese speakers as non native when they are speaking Putonghua simply because you can detect a regional accent Only speakers who are clearly not native speakers of Putonghua or who speak Putonghua with a discernable foreign language accent should be considered non native See Section 7 4 2 for additional discussion of Arabic dialects in broadcast transcripts 7 Transcription Quick rich transcription requires annotators to produce a verbatim transcript of all speech within a file and to add minimal markup to capture salient features of the speech Standard writing conventions including orthography spelling and punctuation are used for ease of comprehension and readability Transcripts must be produced in UTF 8 Unicode encoding Transcripts should be spell checked for common misspellings or typographical err
26. words if they are not in the audio and do not delete words that are spoken even if they are ungrammatical 3 Do not try to normalize dialectal words 4 Do not attempt to transcribe accent features Use standard orthography 5 Do not skip words that are hard to understand Use LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 17 of 18 Version 2 August 30 2006 Appendix 1 Recommended Strategy There are many different ways to interact with XTrans to create a time aligned transcript The following is a synopsis of LDC s recommended strategy for creating broadcast transcripts with XTrans Note that most of these functions are keyboard rather than mouse based commands For quick transcription it is strongly recommended that transcribers choose keyboard over mouse based functions as much as possible This takes a little getting used to but you will find it much faster and easier to use the keyboard only rather than switching between keyboard and mouse and it s easier on your wrists Consult the XTrans user manual for additional information Quick Guide for Quick Transcription 1 open audio file Fil gt Open audio file 2 open new transcript file File gt New 3 associate audio and transcript Edit gt Blindly associate transcript to audio 4 begin playback and mark segment start Alt M 5 stop playback and mark segment end Alt M 6 insert segment Ctrl N Ctrl Insert on nix 7 assign spe
27. y should be grouped together as a single section This is LDC Transcription Guidelines QRTR Broadcast Chinese XTrans Format 10 of 18 Version 2 August 30 2006 different from multiple consecutive news or conversational reports which should be separated into multiple sections 6 Speaker Identification In addition to identifying SUs and section boundaries annotators also label the identity of speakers within a broadcast Speaker IDs are required with each SU segment Each speaker label has three elements speaker type required non native status optional and speaker name if available 6 1 Speaker Type All speakers must be assigned a speaker type There are four speaker types as follows Female used for adult females Male used for adult males Child used for children of either sex Other used for speakers in unison non human computer voices altered voices unknown speaker sex etc 6 2 Names and Identifiers All speakers must be identified byname When name is not known annotators use a unique identifier for each speaker When names are known they should be written out in full For names with multiple spellings or transliterations the most common variant should be used If in common practice the name contains a middle initial or appositive like Jr these should be included and spelled out in full All names must be written in English using the most common transliteration Capitalization sho
Download Pdf Manuals
Related Search