杩戝勾鏉ワ紝娣卞害瀛︿範宸茬粡鎴愪负浜?NLP 棰嗗煙鐨勬爣閰嶆妧鏈紝2022 骞?0 鏈?15 鏃モ€滃皬绾功 REDtech 闈掑勾鎶€鏈矙榫欌€濇椿鍔ㄤ腑锛?b>鎴戜滑闈炲父鑽ｅ垢鍦伴個璇峰埌浜嗕笂娴峰鏃﹀ぇ瀛﹁绠楁満瀛﹂櫌閭遍敗楣忔暀鎺堬紝閭辨暀鎺堝垎浜簡銆婅瑷€妯″瀷鍗虫湇鍔′笌榛戠浼樺寲銆嬫姤鍛婏紝璇︾粏璁茶В浜嗚瑷€妯″瀷瓒婃潵瓒婂ぇ鐨勮儗鏅笅瀵逛簬鏂板簲鐢ㄦā寮忕殑鎺㈢储銆?/b>
閭遍敗楣?/b>锛氬浗瀹朵紭闈掕幏寰楄€咃紝浜庡鏃﹀ぇ瀛﹁幏寰楃悊瀛﹀澹拰鍗氬＋瀛︿綅銆備富瑕佷粠浜嬭嚜鐒惰瑷€澶勭悊銆佹繁搴﹀涔犵瓑鏂瑰悜鐨勭爺绌讹紝鍙戣〃 CCF A/B 绫昏鏂?70 浣欑瘒锛岃幏寰?ACL 2017 鏉板嚭璁烘枃濂栵紙CCF A 绫伙級銆丆CL 2019 鏈€浣宠鏂囧銆併€婁腑鍥界瀛︼細鎶€鏈瀛︺€?021 骞村害楂樺奖鍝嶅姏璁烘枃濂栵紝鏈?5 绡囪鏂囧叆閫?PaperDigest 鍙戝竷鐨?IJCAI/ACL/EMNLP 鐨勬渶鏈夊奖鍝嶅姏璁烘枃锛堣寮曠敤鏁拌繘鍏ュ墠褰撳眾浼氳鐨?20 鍚嶏級銆傚嚭鐗堝紑婧愪笓钁椼€婄缁忕綉缁滀笌娣卞害瀛︿範銆嬶紝Github 鍏虫敞鏁?1.5 涓囷紝璞嗙摚璇勫垎 9.4 鍒嗐€備富鎸佸紑鍙戜簡寮€婧愭鏋?FudanNLP 鍜?FastNLP锛屽凡琚浗鍐呭鏁扮櫨瀹跺崟浣嶄娇鐢ㄣ€?015 骞村叆閫夐灞婁腑鍥界鍗忛潚骞翠汉鎵嶆墭涓惧伐绋嬮」鐩紝2018 骞磋幏閽变紵闀夸腑鏂囦俊鎭鐞嗙瀛︽妧鏈闈掑勾鍒涙柊濂栦竴绛夊锛?020 鑾风鍥涘眾涓婃捣楂樻牎闈掑勾鏁欏笀鏁欏绔炶禌浼樼瓑濂栵紝2021 骞磋幏棣栧眾涓婃捣甯傝绠楁満瀛︿細鏁欏鎴愭灉濂栦竴绛夊锛堢涓€瀹屾垚浜猴級绛夈€傚煿鍏诲鐢熷娆¤幏寰椾竴绾у浼氫紭鍗氥€佸井杞鑰呫€佺櫨搴﹀瀛﹂噾绛夈€?/p>
浠ヤ笅鍐呭鏍规嵁閭辨暀鎺堢幇鍦烘姤鍛婃暣鐞?/i>
1. 鑳屾櫙
鍦ㄩ璁粌鏃朵唬锛屾垜浠ぇ閮ㄥ垎鐨勭爺绌跺伐浣滃垎涓や釜澶х被鍒細涓婃父濡備綍鍋氭ā鍨嬬殑棰勮缁冿紝涓嬫父濡備綍鍋氱簿璋?/b>锛屼篃灏辨槸鎶婇璁粌妯″瀷杩佺Щ鍒颁笅娓镐换鍔°€傚湪涓婃父鏈夊緢澶氱殑鍏徃涓诲濡?Open AI銆丟oogle 绛夛紝浠栦滑鎶婂ぇ妯″瀷鍋氬緱闈炲父澶э紝鏄剧ず鍑哄緢澶氬緢浼樼鐨勮兘鍔涘 Few-shot 灏忔牱鏈殑鑳藉姏銆傚湪鍙傛暟鐨勬暟閲忕骇涓婂崌涔嬪悗锛屽皬鏍锋湰鐨勫涔犺兘鍔涘氨鍙樺緱闈炲父寮猴紝杩欏氨鏄ぇ瀹惰繕鏄湪涓嶅仠鍦拌杩欎釜妯″瀷鍙樺緱瓒婃潵瓒婂ぇ鐨勫師鍥犮€?/p>
浣嗘槸闅忕潃妯″瀷瓒婃潵瓒婂ぇ锛岃繖绉嶉璁粌鍔犵簿璋冪殑妯″紡鍙樺緱涓嶅彲琛屼簡锛屼竴鏂归潰涓婃父鍒堕€犲ぇ妯″瀷鐨勫叕鍙镐笉鎰挎剰鎶婂畠寮€婧愶紝鍙︿竴鏂归潰涓嬫父鐨勫簲鐢ㄥ巶鍟嗕篃涓嶅お鍙兘鎶婂畠涓嬭浇涓嬫潵锛屽嵆浣夸笅杞戒篃寰堥毦鏈夎祫婧愯繍琛屻€?/p>
鎵€浠ユ垜浠杩芥眰涓€绉嶆柊鐨勫簲鐢ㄦā寮?/b>銆傛瘮濡備互 GPT 涓轰唬琛ㄦ彁鍑烘潵鐨?in-context learning锛堝湪涓婁笅鏂囦腑瀛︿範锛夛紝缁欓璁粌妯″瀷杈撳叆涓€浜涙彁绀烘垨鑰呬緥瀛愶紝璁╁師妯″瀷鏍规嵁杩欎簺渚嬪瓙鍦ㄤ笅娓镐换鍔′笂杩涜閫傞厤锛屽畠鐨勬晥鏋滀篃闈炲父濂斤紝浠?GPT-3 涓轰緥鐨勪竴浜涙ā鍨嬭〃鐜板緱闈炲父鎯婅壋銆俰n-context learning 鎴愪负鎴戜滑鍦ㄨ繖涓鍩熶笂鐮旂┒鐨勯噸鐐广€?/p>2. Language-Model-as-a-Service 璇█妯″瀷鍗虫湇鍔?/h2>
濡傛灉妯″瀷鏄儴缃插湪鏈嶅姟绔殑锛岀浉褰撲簬鎶婅瑷€妯″瀷鍋氭垚涓€涓湇鍔★紝鎴戜滑灏辨彁鍑轰簡鈥滆瑷€妯″瀷鍗虫湇鍔♀€?/b>鐨勬蹇点€傝瑷€妯″瀷鍗虫湇鍔′簨瀹炰笂宸茬粡鏄竴涓緢鎴愮啛鐨勫簲鐢ㄤ簡锛屾湁寰堝鐨勫簲鐢ㄤ篃閮芥槸鍩轰簬璇█妯″瀷鍗虫湇鍔＄殑鑳藉姏銆傚儚 GPT-3 寮€鍙戠殑涓€浜涗笅娓哥殑鏈嶅姟鈥斺€旀垜浠彲浠ョ敤 GPT-3 鐢熸垚涓€涓綉椤垫寜閽紝鐢ㄥ畠鎶婅嚜鐒惰瑷€杞寲鎴愭暟瀛﹀叕寮忕瓑绛夈€?/p>
鍦ㄨ瑷€妯″瀷鍗虫湇鍔′腑鎴戜滑浼氬瓨鍦ㄤ袱涓寫鎴橈細
鏈嶅姟鐨勫師妯″瀷鏄粈涔堬紵
濡備綍鎶婂畠閫傞厤鍒颁笅娓镐换鍔″綋涓紵
3. United Foundation Model
缁熶竴鐨勯璁粌妯″瀷鐨勭洰鏍囨槸鐢ㄤ竴涓ā鍨嬫潵閫傞厤鎵€鏈夌殑鑷劧璇█澶勭悊浠诲姟锛屾瘮濡傛垜浠璁粌涓€涓ā鍨嬶紝璁╁畠鏃㈣兘澶熸敮鎸佺悊瑙ｄ换鍔★紝涔熷彲浠ユ敮鎸佺敓鎴愪换鍔°€?/p>
CPT锛氫竴绉嶉潪瀵圭О鐨勯璁粌 Transformer
鍦ㄤ紶缁熺殑棰勮缁冩ā鍨嬩笂鏈夊嚑绫讳唬琛紝濡備互 BERT 涓轰緥鐨勭悊瑙ｆā鍨嬶紝GPT 涓轰唬琛ㄧ殑鐢熸垚妯″瀷锛岃繕鏈?BART銆傞偅涔堣兘涓嶈兘鎶婂畠浠眹鎬诲埌涓€璧峰憿锛熸垜浠彁鍑轰簡涓€涓柊鐨勬ā鍨?CPT锛屽畠鐨勬牳蹇冩€濇兂灏辨槸灏嗙悊瑙ｄ换鍔″拰鐢熸垚浠诲姟鍚堝苟鍒颁竴璧?/b>锛屾瘮濡傛垜浠妸 BERT 鍜?BART 鍚堝苟鍒颁竴璧风殑鏃跺€欙紝鍙戠幇閮介渶瑕佷竴涓叡鍚岀殑缂栫爜鍣紝鍏变韩缂栫爜鍣ㄥ悗鎴戜滑寰楀埌濡備笅鍥捐繖绉嶅舰鐘躲€?/p>
瀹冨悓鏍锋槸 Transformer 鐨?Encoder-Decoder 鏋舵瀯锛屼絾鍏跺乏杈瑰彲浠ョ敤鏉ュ仛鐞嗚В锛屽彸杈瑰彲浠ュ仛鐢熸垚锛屽湪寰堝涓枃棰勮缁冧换鍔′笂閮借兘澶熻揪鍒扮洰鍓嶆渶濂界殑鏁堟灉锛屽悓鏃堕潪瀵圭О鐨?Transformer 鐨?Encoder-Decoder 鏋舵瀯锛屼篃浣垮叾鐢熸垚鏁堢巼鎻愬崌浜?鍊嶄互涓娿€?/p>
Seq2Seq Masked Language Modeling
鐩墠锛岃嚜鐒惰瑷€澶勭悊褰撲腑锛岃兘澶熸敮鎸侀潪甯稿浠诲姟绫诲瀷鐨勮瑷€妯″瀷鏂瑰紡灏辨槸搴忓垪鍒板簭鍒楁ā鍨嬶紝涓€涓吀鍨嬬殑浠ｈ〃灏辨槸 T5锛屽畠鍙互鎶婂緢澶氱殑鑷劧璇█澶勭悊浠诲姟閮借浆鍖栨垚涓哄簭鍒楀埌搴忓垪鐨勫舰寮忋€傚鏋滃彲浠ヨ繖鏍疯浆鍖栵紝鎴戜滑鐨勫悗鍙板幓閮ㄧ讲涓€涓繖鏍峰簭鍒楀埌搴忓垪鐨勫熀纭€妯″瀷锛屽氨鍙互鐢ㄦ潵鏀寔涓嬫父浠诲姟浜嗐€?/p>
浣嗘槸鐢?T5 澶勭悊鑷劧璇█澶勭悊浠诲姟鏃朵緷鐒舵槸闈炲父鏈夋寫鎴樻€х殑锛屽湪鏇村鐨勫簲鐢ㄥ綋涓紝涓€浜涗换鍔￠€氬父鏉ヨ鏄瘮杈冮毦浠ヨ浆鍖栫殑銆傛瘮濡?ABSA锛堝湪鑷劧璇█澶勭悊鏂归潰绾х殑鎯呮劅鍒嗘瀽锛夈€傝繖閲岀粰鍑轰竴鍙ヨ瘽 鈥淒rink are always well made鈥濓紝鍏朵腑鏈変竴涓瘎浠峰璞★紝杩樻湁涓€涓瘎浠疯瘝浠ュ強浠栫殑鎯呮劅鍊惧悜锛岃繖浜涢兘闇€瑕佷粠杩欎釜鍙ュ瓙涓娊鍙栧嚭鏉ャ€?/p>
浜嬪疄涓婏紝ABSA 浠诲姟鍙堝垎涓哄緢澶氱殑瀛愪换鍔★紝涓嶅悓鐨勫瓙浠诲姟鐢ㄤ簬澶勭悊涓嶅悓鐨勬儏鍐点€傛瘮濡傝鍍?a1 杩欎釜浠诲姟灏辨槸鍙娊鍙栨柟闈㈣瘝锛岃繕鏈?o1 杩欎釜浠诲姟鍙娊鍙栬瘎浠疯瘝锛屼笉鍚屼换鍔＄殑褰㈠紡閮戒笉涓€鏍凤紝鎵€浠ュ埌鐩墠涓烘娌℃湁涓€涓ā鍨嬭兘澶熷悓鏃舵敮鎸佸湪 ABSA 浠诲姟閲岄潰鎵€鏈夌殑瀛愪换鍔°€?/p>
閭ｄ箞鑳戒笉鑳界敤鐢熸垚搴忓垪鍒板簭鍒楁ā鍨嬬殑鏂瑰紡鏉ュ悓鏃跺鐞?涓瓙浠诲姟鍛紝浜嬪疄涓婅繖涓ā鍨嬩篃闈炲父绠€鍗曪紝鎴戜滑鍙互鎶?ABSA 浠诲姟鍋氫竴涓簭鍒楃敓鎴愪换鍔★紝鎶婂畠鍙樻垚涓€涓娊鍙栧璞＄殑搴忓垪涓嬫爣鐨勭敓鎴愶紝姣斿璇存垜浠鎶藉彇 aspect term 鈥渨ine list鈥濓紝鎴戜滑鍙渶瑕佽緭鍑哄畠鐨勮捣濮嬩綅缃?1锛岃繕鏈夊畠鐨勭粨鏉熶綅缃?2锛屽啀鎶?鈥渟ervice鈥濓紝涔熸槸寮€濮嬩綅缃拰缁撴潫浣嶇疆锛屽嵆 鈥?2, 12鈥濓紝浠ュ簭鍒楃殑鏂瑰紡鎶婂畠鐨勪綅缃敓鎴愬嚭鏉ュ嵆鍙€?/p>
瀵逛簬涓夊厓缁勭殑浠诲姟锛屽氨鐢熸垚鈥渨ine list鈥?1, 2锛屽啀鐢熸垚瀵瑰簲鐨?Opinion 鈥渋nteresting鈥濓紝鍐嶇敓鎴愬畠鐨勬儏鎰熷€惧悜锛岃繖鏍锋垜浠氨鎶?ABSA 鐢ㄧ粺涓€鐨勫簭鍒楀埌搴忓垪鐨勫舰寮忛噸鏂板舰寮忓寲锛屾鏃舵垜浠氨鍙互鐢ㄤ竴涓ā鍨嬫潵鏀寔鎵€鏈夌殑7涓瓙浠诲姟锛屽畠缁熶竴妗嗘灦灏卞彉寰楅潪甯哥畝鍗曪紝鐢ㄤ竴涓?BART 鐨?Encoder-Decoder 灏辫兘澶熷幓澶勭悊浜嗐€傝繖涓伐浣滀笉浣嗗舰寮忕畝鍗曪紝鐢ㄤ竴涓妧鏈ā鍨嬪氨鍋氫簡鎵€鏈夌殑瀛愪换鍔★紝鍚屾椂涔熷緱鐩婁簬杩欎簺棰勮缁冩ā鍨嬶紝鏁堟灉涔熸瘮鍏朵粬鍒嗗紑瀹屾垚鐨勬柟寮忔洿濂姐€?/p>
鍚屾牱鎴戜滑鎶婅繖涓兂娉曚篃鐢ㄥ埌 NER锛堝懡鍚嶅疄浣撹瘑鍒級涓婏紝NER涔熸槸鍦ㄨ嚜鐒惰瑷€澶勭悊涓潪甯搁噸瑕佺殑涓€绫讳换鍔°€傚湪 NER 閲屾湁闈炲父鐨勫鐨勫瓙浠诲姟锛?/p>
鏈夎繛缁殑 NER锛歂ER 涓殑璇嶆槸杩炵画鍑虹幇鐨勶紱
杩樻湁鏄祵鍏ョ殑 NER锛氬湪涓€涓疄浣撻噷闈㈠祵濂楀彟澶栦竴涓疄浣擄紱
浠ュ強涓嶈繛缁殑 NER锛氫竴涓疄浣撳彲鑳芥槸涓嶈繛缁殑鍦ㄦ鏂囧嚭鐜般€?/p>
浼犵粺瑙ｅ喅鏂瑰紡鏄噰鐢ㄤ笉鍚岀殑绠楁硶鏉ュ畬鎴愶紝姣斿杩炵画鐨?NER 灏变細鐢ㄥ簭鍒楁爣娉紝涓嶈繛缁殑 NER 鍩烘湰涓婂埄鐢ㄨ浆绉绘柟娉曘€?/p>
搴忓垪鏍囨敞寰堥毦澶勭悊涓嶈繛缁殑 NER锛?鎵€浠ヨ繖浜涙柟娉曚箣闂翠笉閫氱敤锛屾垜浠篃鍙互鐢ㄥ簭鍒楀埌搴忓垪鐨勬柟娉曞皢 3 绉?NER 鐨勫瓙浠诲姟鍋氫竴涓粺涓€锛屽悓鏍风被浼间簬 ABSA 涓殑鍋氭硶銆?/p>
鎴戜滑鎶?NER 鐢熸垚鍑烘潵锛屾瘮濡傝鎶藉彇 鈥渕uscle pain鈥濓紝鎴戜滑灏辩敓鎴愬畠瀵瑰簲鐨勪綅缃紝鐒跺悗鍐嶇敓鎴愬畠瀵瑰簲鐨勫疄浣撶殑绫诲瀷鍗冲彲銆傚悓鏍蜂篃鍙互鐢ㄥ熀纭€鐨?BART Encoder-Decoder 锛岃繖鏍峰畠灏卞彲浠ラ潪甯告柟渚垮湴鍘诲仛鍚勭涓嶅悓绫诲瀷鐨?NER銆傝繖绉嶆柟寮忔晥鏋滀篃闈炲父濂斤紝鐩墠鍦ㄤ富娴佺殑 NER 鏁版嵁闆嗕笂閮借兘杈惧埌闈炲父濂界殑鏁堟灉銆?/p>
4. Efficient Tuning Algorithm
鏈変簡鍩虹鐨勭粺涓€棰勮缁冩ā鍨嬩箣鍚庯紝鎴戜滑鎬庝箞鏇村姞鏈夋晥鍦版妸瀹冭縼绉诲埌涓嬫父鐨勫悇绉嶄笉鍚屼换鍔′笂鍛紵杩欓噷灏卞垎浜嗗緢澶氱鏂瑰紡锛?/p>
饾挻-Tuning锛堟爣绛捐皟閫傦級
瀵逛簬涓€涓璁粌妯″瀷锛岃緭鍏ヤ竴涓彞瀛愭椂锛屾垜浠厛鍘绘彁鍙栧畠鐨?Feature 鏋勬垚 Feature Space锛堢壒寰佺┖闂达級锛屽啀鎶?Feature Space 鍋氬弬鏁拌皟鑺傦紝鍚?Label Space锛堟爣绛剧┖闂达級鍘诲仛鏄犲皠锛岃繖灏辨槸浼犵粺鐨?Fine tuning銆傜敱浜庣壒寰佺┖闂村拰鍙傛暟绌洪棿闈炲父澶э紝杩欎釜宸ヤ綔閫氬父闇€瑕佸ぇ閲忕殑鏁版嵁鍘诲仛璋冭妭銆?/p>
鎴戜滑鑳戒笉鑳芥兂鍙﹀涓€涓柟娉曞憿锛熸槸鍚﹀彲浠ユ妸 Feature Space 鍥哄畾锛岃€屽幓璋?Label Space锛屾妸鏍囩绌洪棿鍚戠壒寰佺┖闂撮潬鎷紵鎴戜滑閫氬父鐢?鈥測鈥?鏉ヨ〃绀烘爣绛撅紝鎵€浠ユ妸杩欎釜鏂规硶鍙仛 鈥?b>饾挻-Tuning鈥濄€?/p>
杩欎釜鏂规硶鏉ヨ嚜浜庢垜浠洿鏃╀箣鍓嶇殑涓€涓伐浣滐紝杩欎釜宸ヤ綔鍙互灏嗘枃鏈换鍔¤浆鍖栨垚鏂囨湰鍖归厤浠诲姟銆備紶缁熺殑鏂囨湰鍒嗛厤鏄粰浣犱竴涓彞瀛愯緭鍏ュ畠鐨勬爣绛撅紝姝ゆ椂鎴戜滑鍏跺疄骞舵病鏈夊お鍘诲埄鐢ㄦ爣绛剧殑淇℃伅锛屾瘮濡傝杩欎釜鏍囩鎴戜滑鍙互鐢ㄤ竴鍙ヨ瘽鏉ユ弿杩扮殑璇濓紝鎴戜滑灏卞彲浠ユ妸鍒嗙被浠诲姟鍙樻垚涓€涓枃鏈尮閰嶄换鍔★紝鐪嬭繖涓彞瀛愬拰杩欎釜鏍囩鏈夋病鏈夎繘琛岀浉浜掔殑鍖归厤銆傞€氳繃杩欑娉涘紡鐨勮浆鍙橈紝鎴戜滑灏卞彲浠ラ潪甯歌交鏉剧殑鍘绘彁鍗囨枃鏈垎绫荤殑鎬ц兘銆?/p>
鈥?饾挻-Tuning鈥?涔熸槸绫讳技杩欑鑰冭檻銆傛垜浠皢鏍囩鎴栬€呮槸鏍囩鐨勮〃杩颁綔涓鸿緭鍏ワ紝灏辨瀯閫犲鍥炬灦鏋勶紝宸﹁竟鏄璁粌妯″瀷锛屽叾鍙傛暟鏄浐瀹氫笉鍔ㄧ殑锛屽彧鐢ㄦ潵鎻愬彇 Feature锛屽彸杈硅緭鍏ヤ竴浜涙爣绛撅紝涔熷氨鏄?鈥?b>饾挻 鈥濓紝鍚屾椂杩樻湁涓€涓?Task token锛孴ask token 鐢ㄦ潵鏈€鍚庡幓鎸囧嚭鏈€缁堢殑鏍囩鏄摢涓€涓紝瀹冧篃缁忚繃涓€涓?Transformer 鐨勬灦鏋勶紝绫讳技浜庝竴涓?Encoder-Decoder 鐨勬灦鏋勶紝鍙笉杩囧乏绔槸涓嶅仛璋冩暣鐨勶紝鎴戜滑鍙渶瑕佽皟鍙崇鐨勫弬鏁般€傚彸绔殑瑙勬ā閫氬父姣旇緝灏忥紝鎵€浠ュ畠鐨勬晥鐜囨槸闈炲父楂樼殑銆?/p>
鍦ㄦ灦鏋勪笂锛屸€?饾挻-Tuning鈥濆尯鍒簬鈥淔ine-Tuning鈥? 鈥淎dapter-Tuning鈥? 鈥淧rompt-Tuning鈥?涓嶉渶瑕佽绠?PTM 鏈韩鐨勬搴︼紝鎵€浠ュ叾浼樺寲鏁堢巼闈炲父楂樸€?/p>
鍦ㄤ竴浜涢€氱敤鐨勮瑷€鐞嗚В鏁版嵁闆嗕笂锛屸€?饾挻-Tuning鈥?閮借兘姣斿儚 鈥淧-Tuning鈥? 鈥淔ix-Tuning鈥?鏁堟灉瑕佸ソ銆傚綋鐒跺拰 鈥淔ine-tuning鈥?杩樻湁涓€瀹氱殑宸窛锛屾湁寰堝ぇ鐨勬敼杩涚┖闂淬€?/p>
鈥?饾挻-Tuning鈥?鏈€澶х殑浼樼偣灏辨槸璁粌鏁堢巼鐗瑰埆楂樸€傚畠涓嶉渶瑕佽绠楅璁粌妯″瀷鐨勬搴︼紝鎵€浠ラ鍏堝湪鍐呭瓨涓婁細鏈夊緢澶х殑鑺傜渷锛岃妭鐪佺殑杩欎簺鍐呭瓨鎴戜滑瀹屽叏鍙互澧炲ぇ Bech 涔嬬被鐨勪笢瑗匡紝杩涗竴姝ユ彁鍗?鈥?饾挻-Tuning鈥?鐨勬晥鐜囥€?/p>
Black-Box Tuning
闄や簡 鈥?饾挻-Tuning鈥?涔嬪锛岃兘涓嶈兘渚濈劧璋冧竴浜涘弬鏁帮紝浣嗘槸鍚屾牱涓嶉渶瑕佽绠楁搴︼紝鏄惁鑳借揪鍒拌繖鏍风殑鏁堟灉锛?/p>
杩欏氨鏄?Black-Box Tuning锛堥粦绠变紭鍖栵級锛岄粦绠变紭鍖栫殑鏁翠綋鎬濇兂鏄繖鏍风殑锛屾垜浠妸涓€涓璁粌妯″瀷閮ㄧ讲鍦ㄦ湇鍔″櫒绔紝鎶婂畠褰撴垚涓€涓粦鐩掑瓙锛屽畠鍙彁渚涘墠椤圭殑璁＄畻锛屾垜浠繕鍙互閫氳繃澧炲姞涓€浜?Adapt銆丳rompt 鍘昏皟鑺傦紝鎶婂畠閫傞厤鍒颁笉鍚岀殑浠诲姟涓娿€?/p>
閫氬父鍍?Prompt tuning 鐨勬柟娉曪紝闇€瑕侀€氳繃澶фā鍨嬭绠楁搴︼紝鍐嶉€氳繃姊害璋冭妭 Prompt 鍙傛暟锛屽浜庤繖浜涘ぇ妯″瀷閮ㄧ讲鍦ㄦ湇鍔″櫒涓婂 GBT锛?鏄笉鍙鐨勩€傛垜浠笇鏈涙妸棰勮缁冩ā鍨嬬湅鎴愪竴涓粦绠憋紝鍘诲鎵句竴涓?Prompt锛屼娇寰楀畠鍦ㄤ笅娓镐换鍔′笂鐨勬晥鏋滄渶濂姐€備竴鏃︽垜浠笉鑳借幏寰楄繖涓搴︼紝鍏跺疄灏辨妸瀹冭浆鍙樻垚涓€涓粦绠变紭鍖栭棶棰橈紝鎴栬€呮槸鏃犳搴︿紭鍖栭棶棰樸€傚湪鏃╂湡宸ヤ綔鎴栦紭鍖栧伐浣滀腑鎴戜滑鎵惧埌浜嗕竴浜涙湁鏁堢殑鏃犳搴︿紭鍖栨柟娉曪紝浣嗘槸瀹冧粎鍦ㄤ綆缁寸┖闂存瘮杈冩湁鏁堬紝鍦ㄩ珮缁寸┖闂寸敱浜庢悳绱㈢┖闂撮潪甯稿ぇ锛岀洰鍓嶆潵璁茶繕鏄潪甯镐綆鏁堛€?/p>
鐗瑰埆瀵逛簬澶фā鍨嬫潵璁诧紝鍗充究鏄?Prompt锛屽畠鐨勫弬鏁颁篃闈炲父澶с€傛瘮濡傝 50 涓?Prompt token锛屾瘡涓湁 1000 缁寸殑璇濆氨鏄?5 涓囦釜鍙傛暟锛? 涓囦釜鍙傛暟绌洪棿鏄潪甯稿ぇ鐨勩€傛€庢牱鎶婂畠鍦ㄤ綆缁寸┖闂存湁鏁堢殑鏃犳搴︿紭鍖栫敤鍒伴珮缁寸┖闂村憿锛岃繖鏄竴涓寫鎴樸€?/p>
鎵€骞哥殑鏄湪楂樼淮绌洪棿涓笉鏄墍鏈夌殑鍙傛暟閮界瓑鍚岄噸瑕侊紝姣斿绁炵粡缃戠粶涓篃鏈夊緢澶氬弬鏁伴兘鏄啑浣欑殑锛屾湁浜涘弬鏁板苟涓嶆槸閭ｄ箞閲嶈锛屽洜姝ゅ湪杩欎箞澶氱殑鍙傛暟绌洪棿涓紝鏄笉鏄彲浠ュ彂鐜拌繖浜?Prompt 鎴栬€呮槸澶фā鍨嬬殑鏈緛缁村害锛屽畠鐨勬湰鐪熺淮搴﹀彲鑳藉湪闈炲父浣庣淮搴︾殑绌洪棿銆傛垜浠湪浣庣淮鐨勬湰鐪熺淮搴︾┖闂村幓浼樺寲杩欎釜 Prompt锛岀敤鏃犳搴︾殑鏂规硶杈惧埌寰堝ソ鐨勬晥鏋溿€?/p>
鍩烘湰妗嗘灦濡備笅鍥撅紝棣栧厛鎴戜滑鎶?Prompt 鏄犲皠鍒颁綆缁寸┖闂达紝鍦ㄤ綆缁寸┖闂翠腑鐢ㄦ棤姊害浼樺寲鐨勬柟娉曚紭鍖栵紝瀹屾垚涔嬪悗鍐嶆妸瀹冩槧灏勫洖鍘伙紝杩欐牱鍙互閫氳繃鏃犳搴︿紭鍖栫殑鏂规硶鍘讳紭鍖栧ぇ妯″瀷锛屽苟涓旀妸瀹冮€傞厤鍒颁笅娓镐换鍔′笂銆?/p>
鍦ㄦ瘮濡?Few-shot 杩欎簺浠诲姟涓婏紝Black-Box Tuning 鍩烘湰鍙互杩藉钩鍩轰簬姊害鐨勬柟娉曪紝浣嗘槸鏈変竴涓己鐐癸紝鎴戜滑鐨?Prompt 鏈€濂?Pre-train 涓€涓嬨€傞€氳繃 Black-box 鎴戜滑楠岃瘉浜嗗彲浠ョ敤鏃犳搴︾殑鏂规硶杩涜澶ц妯￠璁粌妯″瀷鐨勮皟鍙傦紝浣嗘槸渚濈劧瀛樺湪缂洪櫡锛屽畠鐨?Prompt 璋冭捣鏉ヤ緷鐒堕潪甯稿洶闅撅紝骞朵笖闇€瑕侀璁粌銆?/p>
鎴戜滑鑳戒笉鑳芥妸鎶€鏈繘涓€姝ユ敼鍠勫憿锛熸垜浠氨鎻愬嚭浜嗙浜屼釜鐗堟湰 BBTv2锛屽湪杩欎釜鐗堟湰褰撲腑鎴戜滑鍋氫簡涓€浜涙敼杩涳紝鎴戜滑涓嶉渶瑕佸仛 Prompt 鐨勯璁粌锛屽悓鏃舵敼杩涢殢鏈烘姇褰辩殑鏂规硶锛屽苟涓旈噰鐢?Deep prompt锛屾瘡涓€灞傞兘鍔犱竴浜?Prompt銆備簨瀹炶瘉鏄庤繖浜涚瓥鐣ユ槸鏈夋晥鐨勩€?/p>
鎴戜滑鍙互鐪嬪埌锛岀粡杩囪繖鏍蜂竴浜涙敼杩涗箣鍚庯紝BBTv2 鍦ㄦ瘮濡備竴涓囦釜鍙皟鍙傛暟鐨勬儏鍐典笅锛屽畠杈惧埌浜嗙洰鍓嶆渶濂界殑鏁堟灉锛屾瘮鍩轰簬鍚殑鏂规硶鏁堟灉杩樿濂斤紝骞朵笖涓嶉渶瑕?Pre-train銆備篃灏辨槸璇村浜庤繖浜涘ぇ妯″瀷鏉ヨ锛屾垜浠敤鏃犳搴︽柟娉曞氨鑳藉鎵撹触鍩轰簬姊害鐨勬柟娉曪紝鎵€浠ヨ繖涔熸槸杩欎釜宸ヤ綔鐨勬剰涔夋墍鍦紝涔熺粰灏嗘潵涓€浜涘ぇ妯″瀷鐨勫簲鐢ㄦ彁渚涗簡鍙﹀涓€涓満鏅紝鎶婂ぇ妯″瀷閮ㄧ讲鍦ㄦ湇鍔″櫒绔紝鍙敤瀹冪殑 Forword 灏辫浜嗭紝鎴戜滑璋冨弬涓嶉渶瑕佹搴︼紝鍙渶瑕佸墠椤圭殑璁＄畻銆?/p>
5. Summary
鈥滆瑷€妯″瀷鍗虫湇鍔♀€濇槸鏈枃鐨勪竴涓富瑕佹蹇碉紝璇█妯″瀷鍗虫湇鍔＄殑搴旂敤鎵嬫锛屽ぇ姒傚垎鎴愪簲绫伙細
Text prompt锛?/b>
鍙互浜哄伐璁捐涓€浜涘熀浜庢枃鏈殑 Prompt锛屼絾鏄彉鎴愪簡鐗瑰緛宸ョ▼闂锛岄渶瑕佸伐绋嬪笀涓嶆柇鍘昏瘯锛岀浉褰撹€楄垂绮惧姏銆?/p>
In-context learing锛?/b>
鐩墠鏉ヨ In-context learing 鍦?GPT 瀹為獙涓婃槸闈炲父鏈夋晥鐨勶紝浣嗗湪鍏朵粬妯″瀷涓婅繕闇€瑕佷竴瀹氱殑楠岃瘉锛屼絾鏄畠鏄潪甯告湁鍓嶆櫙鐨勬柟鍚戯紝鍏朵腑鐨勯棶棰樹篃闈炲父鍊煎緱澶у鍘荤爺绌躲€?/p>
Data generation锛?/b>
鎴戜滑鐢ㄥぇ妯″瀷鍘荤敓鎴愪竴浜涙暟鎹紝鍐嶇敤杩欎簺鏁版嵁璁粌涓€涓洿灏忕殑妯″瀷锛岃繖涔熸槸涓€绉嶆柟娉曘€?/p>
Black-box optimization锛?/b>
鍗充笂鏂囨墍杩扮殑 Black-box tuning銆?/p>
Feature-based-learning锛?/b>
鎴戜滑鎶婇璁粌妯″瀷鐨勮緭鍑轰綔涓轰竴绉?Feature锛岃緭鍏ョ粰涓€浜涚壒瀹氱殑妯″瀷銆傗€?饾挻-Tuning鈥?灏辨槸杩欑浣跨敤銆?/p>
6. 鈥淨&A鈥濈幆鑺?/b>
Q锛氳秴澶ч璁粌妯″瀷璇█妯″瀷 Large 鐨勬ā鍨嬪湪宸ヤ笟搴旂敤涓婄殑鍙鎬ф槸鎬庢牱鐨勶紝鍍忓垰鍒氭彁鍒扮殑鏈€杩戞瘮杈冪伀鐨?Diffusion 妯″瀷锛屾垨鑰呰鍏朵粬涓€浜涘鏉傜殑澶氭ā鎬併€侀璁粌妯″瀷锛?/p>
閭遍敗楣忥細鎹垜鎵€鐭ワ紝杩欎簺澶фā鍨嬪湪宸ヤ笟鐣岀殑搴旂敤闈炲父澶氾紝姣斿璇村湪涓€浜涚粓绔换鍔′笂锛岀壒鐐规槸涓€鏃︽湁浜嗗熀纭€搴旂敤锛屼笅娓搁兘涓嶆槸闂銆傜幇鍦ㄧ殑涓昏闂鏄垚鏈紝濡備綍楂樻晥鐨勯€傞厤浠ュ強閫氳繃涓€浜涙ā鍨嬪帇缂╂垨鑰呭叾浠栨柟娉曟潵鎻愰珮鏁堢巼銆?/p>
Q锛氱敓鎴愬紡瀹炰綋鎴栬€呮儏鎰熸娊鍙栧簲鐢ㄥ埌宸ヤ笟鍦烘櫙涓昏浼氶潰涓翠粈涔堟寫鎴橈紵
閭遍敗楣忥細杩欎釜鎸戞垬杩樻槸鍦ㄤ簬锛屾垜浠繖閲岃鐨勮繖浜涙柟娉曪紝鍏跺疄杩樻槸闇€瑕佸ぇ閲忕殑璁粌鏁版嵁鐨勶紝鍦ㄧ湡姝ｇ殑宸ヤ笟鍦烘櫙褰撲腑锛屽緢澶氭椂鍊欐爣鍑嗘暟鎹笉鏄偅涔堝锛岃繖鏍风殑璇濈敓鎴愭ā鍨嬩笉鍍忓叾浠栫殑鏂规硶锛屾垨鑰呰浼犵粺鐨勮缁冩柟娉曟晥鏋滃ソ銆備絾鏄垜杩樻槸鍧氫俊闅忕潃棰勮缁冩ā鍨嬬殑鍙戝睍锛屼細鏈夋槑鏄剧殑鎻愬崌銆?/p>