Suchergebnisse

International audience ; This paper deals with a survey corpus. We present information retrieval about the speaker. We used finite state transducer cascades and we present here detailed results with an evaluation. This work is part of a French project to enhance the corpus ESLO (sociolinguistic survey taken in the city of Orléans). This survey has been realized in 1968 and the project is to save records in computer format, to transcribe them and to increase the transcription with annotations in XML format. This work was supported by a French ANR contract (ANR-06-CORP-023) and by European fund from Région Centre (FEDER). The corpus represent a collection of 200 interviews with the questions about the life in the city of Orléans: How long have you lived in Orléans for?, What led you to live in Orléans?, Do you like living in Orléans?, etc. and questions about the occupation or the family of the speaker, completed by recordings within a professional or private context. The recording situations are different: interviews, discussions between friends, recordings in microphone hidden, interviews with the political, academic and religious personalities, conversations between a social worker and parents in Psycho Medical Center of Orleans. In total, we have 300 hours of speech estimated to 4,500,000 words. More precisely, we worked on almost 120 transcribed hours representing 112 Transcriber XML files and 32 577 Kb. We worked on 105 files (31 004 Kb) and we evaluated the results on 7 files (1 573 Kb-5.1%). The transcription files have no punctuation marks, but the first letter of proper names is capitalized and acronyms are fully capitalized. We used the CasSys system (Friburger, Maurel, 2004) that computes texts with transducer cascades (Abney, 1996). The cascades we used are hand built: each transducer describes a local grammar for the recognition of some entities. Some times this recognition needs the succession of two or more transducers, in a specific order. More precisely, we used two cascades; the first one, for ...

International audience ; This paper deals with a survey corpus. We present information retrieval about the speaker. We used finite state transducer cascades and we present here detailed results with an evaluation. This work is part of a French project to enhance the corpus ESLO (sociolinguistic survey taken in the city of Orléans). This survey has been realized in 1968 and the project is to save records in computer format, to transcribe them and to increase the transcription with annotations in XML format. This work was supported by a French ANR contract (ANR-06-CORP-023) and by European fund from Région Centre (FEDER). The corpus represent a collection of 200 interviews with the questions about the life in the city of Orléans: How long have you lived in Orléans for?, What led you to live in Orléans?, Do you like living in Orléans?, etc. and questions about the occupation or the family of the speaker, completed by recordings within a professional or private context. The recording situations are different: interviews, discussions between friends, recordings in microphone hidden, interviews with the political, academic and religious personalities, conversations between a social worker and parents in Psycho Medical Center of Orleans. In total, we have 300 hours of speech estimated to 4,500,000 words. More precisely, we worked on almost 120 transcribed hours representing 112 Transcriber XML files and 32 577 Kb. We worked on 105 files (31 004 Kb) and we evaluated the results on 7 files (1 573 Kb-5.1%). The transcription files have no punctuation marks, but the first letter of proper names is capitalized and acronyms are fully capitalized. We used the CasSys system (Friburger, Maurel, 2004) that computes texts with transducer cascades (Abney, 1996). The cascades we used are hand built: each transducer describes a local grammar for the recognition of some entities. Some times this recognition needs the succession of two or more transducers, in a specific order. More precisely, we used two cascades; the first one, for ...

International audience ; This paper deals with a survey corpus. We present information retrieval about the speaker. We used finite state transducer cascades and we present here detailed results with an evaluation. This work is part of a French project to enhance the corpus ESLO (sociolinguistic survey taken in the city of Orléans). This survey has been realized in 1968 and the project is to save records in computer format, to transcribe them and to increase the transcription with annotations in XML format. This work was supported by a French ANR contract (ANR-06-CORP-023) and by European fund from Région Centre (FEDER). The corpus represent a collection of 200 interviews with the questions about the life in the city of Orléans: How long have you lived in Orléans for?, What led you to live in Orléans?, Do you like living in Orléans?, etc. and questions about the occupation or the family of the speaker, completed by recordings within a professional or private context. The recording situations are different: interviews, discussions between friends, recordings in microphone hidden, interviews with the political, academic and religious personalities, conversations between a social worker and parents in Psycho Medical Center of Orleans. In total, we have 300 hours of speech estimated to 4,500,000 words. More precisely, we worked on almost 120 transcribed hours representing 112 Transcriber XML files and 32 577 Kb. We worked on 105 files (31 004 Kb) and we evaluated the results on 7 files (1 573 Kb-5.1%). The transcription files have no punctuation marks, but the first letter of proper names is capitalized and acronyms are fully capitalized. We used the CasSys system (Friburger, Maurel, 2004) that computes texts with transducer cascades (Abney, 1996). The cascades we used are hand built: each transducer describes a local grammar for the recognition of some entities. Some times this recognition needs the succession of two or more transducers, in a specific order. More precisely, we used two cascades; the first one, for ...

International audience ; This paper deals with a survey corpus. We present information retrieval about the speaker. We used finite state transducer cascades and we present here detailed results with an evaluation. This work is part of a French project to enhance the corpus ESLO (sociolinguistic survey taken in the city of Orléans). This survey has been realized in 1968 and the project is to save records in computer format, to transcribe them and to increase the transcription with annotations in XML format. This work was supported by a French ANR contract (ANR-06-CORP-023) and by European fund from Région Centre (FEDER). The corpus represent a collection of 200 interviews with the questions about the life in the city of Orléans: How long have you lived in Orléans for?, What led you to live in Orléans?, Do you like living in Orléans?, etc. and questions about the occupation or the family of the speaker, completed by recordings within a professional or private context. The recording situations are different: interviews, discussions between friends, recordings in microphone hidden, interviews with the political, academic and religious personalities, conversations between a social worker and parents in Psycho Medical Center of Orleans. In total, we have 300 hours of speech estimated to 4,500,000 words. More precisely, we worked on almost 120 transcribed hours representing 112 Transcriber XML files and 32 577 Kb. We worked on 105 files (31 004 Kb) and we evaluated the results on 7 files (1 573 Kb-5.1%). The transcription files have no punctuation marks, but the first letter of proper names is capitalized and acronyms are fully capitalized. We used the CasSys system (Friburger, Maurel, 2004) that computes texts with transducer cascades (Abney, 1996). The cascades we used are hand built: each transducer describes a local grammar for the recognition of some entities. Some times this recognition needs the succession of two or more transducers, in a specific order. More precisely, we used two cascades; the first one, for ...