基于半监督式学习之广播节目语音逐字稿自动转写系统Automatic.PDFVIP

The 2017 Conference on Computational Linguistics and Speech Processing ROCLING 2017, pp. 197-212 飪 The Association for Computational Linguistics and Chinese Language Processing 鍩烘柤鍗婄洠鐫ｅ紡瀛哥繏涔嬪唬鎾瘈鐩獮闊抽€愬瓧绋胯嚜鍕曡綁瀵郴绲 Automatic Transcription of Broadcast Radio Speech Based on Quality Estimation-Guided Semi-Supervised Training 鐜嬫槦鏈圫ing-Yue Wang,瑷卞惓鑿 Wu-Hua Hsu, 寤栧厓鐢 Yuan-Fu Liao 鍦嬬珛鑷哄寳绉戞妧澶у闆诲瓙宸ョ▼绯 Department of Electronic Engineering, National Taipei University of Technology , asmayday24@, yfliao@.tw 鎽樿 寤ｆ挱绡€鐩 瑁戒綔鏅傞€氬父鍙湁鏀堕寗瑾為煶瑷婅櫉 锛屾矑鏈変繚鐣欑浉灏嶆噳鐨勭瘈鐩収瀹硅┊閲嬭硣鏂 锛坢etadata 锛夛紝灏庤嚧绡€鐩挱鍑哄緦 锛屽緢闆ｆ绱㈢瘈鐩収瀹 鎴栨槸锛 鍔犱互绲勭箶鍐嶅埄鐢ㄣ€傞嚌灏嶆鍟忛锛 甯哥敤鐨勬柟娉曟槸 浠ヨ獮闊宠鲸瑾嶅櫒锛岃嚜鍕曡綁瀵唬鎾瘈鐩収瀹癸紝鐢㈢敓瑾為煶閫愬瓧绋匡紝浣嗘槸鐩墠缂轰箯 宸叉瑷樺ソ鐨勫唬鎾獮闊宠獮鏂欏韩锛屽洜姝 鐒℃硶瑷撶反鍑洪仼鍚堣綁瀵唬鎾瘈鐩殑瑾為煶杈ㄨ瓨鍣ㄦ墍浠ャ€ 鍦 鏈珫鏂囦腑锛屾垜鍊 鎺㈣◣濡備綍鍚屾檪浣跨敤瑾為煶瑷婅櫉 鐗瑰镜鍙冩暩銆佽鲸瑾嶅櫒杈ㄨ獚 绲愭灉鑸囪獮瑷€妯″瀷 鍙冩暩锛 瑷撶反涓€瑾為煶鍝佽唱浼扮畻 (Quality Estimation 锛孮E鍣? 锛屽彇浠ｅ偝绲卞彧渚濊炒瑾為煶杈ㄨ獚鍣ㄧ殑淇″績鍊 浼扮畻 锛圕onfidence Measure 锛夛紝寰炴簮婧愪笉绲 锛屼絾鏈瑷樼殑澶ч噺寤ｆ挱瑾炴枡涓紝鎸戦伕閬╁悎瑷撶反 瑾為煶杈ㄨ獚鍣ㄧ殑瑾炴枡 锛岄€茶鍗婄洠鐫ｅ紡鑱插妯″瀷瑷撶反锛屼互鎻愬崌杞夊寤ｆ挱瑾炴枡閫愬瓧绋跨殑 鏁堣兘銆 瀵﹂涓 浠ヤ竴涓嶄匠閷勯煶鍝佽唱NER-set1 鑸囦竴鍎壇 NER-set2 涔嬪唬鎾瘈鐩脯瑭﹁獮鏂 闆嗭紝娓│绋 瀛愯獮闊宠鲸瑾嶅櫒 鑸囩稉鍗婄洠鐫ｅ紡瑷撶反寰岋紝鏂扮殑瑾為煶 杈ㄨ獚鍣ㄨ綁瀵獮闊抽€愬瓧绋跨殑鏁堣兘 銆傚椹楃祼鏋 椤ず 缍撳崐鐩ｇ潱寮忚〒绶村緦锛屾柊鐨勮獮闊 杈ㄨ獚鍣ㄥ彲浠ユ妸NER-set1 鑸嘚ER-set2鐨 瀛楀厓杈ㄨ獚閷 鐜囷紙CER 锛夊緸鍘熷绋瓙妯″瀷鐨 25%鑸?4.24% 锛屽浣 鑷?3.61%鑸?3.24% 銆傛澶栵紝鑻ラ€蹭竴 姝ユ敼鐢 閫查殠瑾炶█妯″瀷锛屾洿鍙皣 CER鍐嶆敼鍠 鑷?3.25%鑸?2.63% 銆 闂滈嵉瑭烇細 鍗婄洠鐫ｅ紡瀛哥繏銆佸搧璩及绠椼€佷俊蹇冨害瑭曚及銆佽獮闊宠鲸瑾 绯荤当 197 1. 绨′粙 寤ｆ挱绡€鐩殑瑾為煶璩囨枡婧愭簮涓嶇禃 锛屼絾鍥犱汉鍔涖€佽硣婧愮瓑鍥犵礌锛屽唬鎾瘈鐩＝浣滃畬鎴愬緦锛岄€ 甯稿彧鏈変繚鐣欐渶寰岃鎾嚭鐨勮獮闊宠▕铏燂紝娌掓湁灏囬寗瑁介亷绋嬩腑鐨勭敤鍒扮殑鐩搁棞璩囨枡锛屾暣鐞嗘垚寰岃ō 璩囨枡锛 metadata 锛夈€傚皫鑷寸瘈鐩挱鍑哄緦锛屽緢闆ｆ鍐嶆 绱㈢瘈鐩収瀹癸紝鎴栨槸鍔犱互绲勭箶鍐嶅埄鐢ㄣ€傚洜 姝ゆ垜鍊戝笇鏈涜兘澶犺綁瀵唬鎾瘈鐩敘鐢熻獮闊抽€愬瓧绋匡紝 浠ヤ究灏囧唬鎾瘈鐩 绲勭箶鎴愭湁鑱叉浉 璁撻€欎簺锛 澶ч噺鐨勮獮闊宠硣鏂欏彲浠ユ湁鏇村鐨勫姞鍊奸亱鐢ㄣ€傞櫎浜嗗彲浠ヨ畵鑱界溇鑳藉瀹规槗鍦 浠ユ枃瀛楁绱㈢殑鏂瑰紡锛 鍘绘壘鍒版渶闂滈嵉鐨勮瑳杩板収瀹归儴鍒嗭紝灏ゅ叾鏄悕浜哄湪绡€鐩腑鎵€瑾殑鏁呬簨銆佹兂娉曟€濈董銆佹柊鐭ョ瓑绛 锛 涔