Ƙarfafa Masu Fassara Manyan Harsunan Harsuna ta hanyar Tunawa da Fassara

1. Gabatarwa

Wannan takarda tana binciko wata sabuwar hanya don haɓaka fassarar inji (MT) ta hanyar amfani da ƙwarewar koyo a cikin mahallin na Manyan Harsunan Harsuna (LLMs). Babban jigo shine cewa Tunawa da Fassara (TMs)—ma'ajiyar bayanai na fassarorin mutane na baya—za su iya zama ingantattun umarni kaɗan don LLMs, suna jagorantar su don samar da ingantattun fassarori masu dacewa da yanki ba tare da buƙatar canje-canjen gini ko daidaitawa ba.

Aikin ya tsara kansa a kan hanyoyin da suka gabata waɗanda ko dai suna buƙatar gyara tsarin ƙirar Fassarar Injin Jijiya (NMT) ko gina ma'ajiyar ilimin fassara daban. Sabanin haka, hanyar da aka gabatar, Umarnin Tunawa da Fassara don Manyan Harsunan Harsuna (TMP-LM), fasaha ce mai sauƙi, kawai ta umarni wacce ke amfani da ikon LLM na asali na fahimta da bin umarnin da aka gabatar a cikin taga mahallinsa.

2. Hanyar Aiki: Umarnin Tunawa da Fassara don LLMs (TMP-LM)

TMP-LM tsari ne mai sauƙi amma mai ƙarfi wanda ke shigar da ilimin fassara cikin LLM ta hanyar gabatar da misalan TM masu dacewa ga tambayar fassara. Tsarin ya haɗa da: 1) Maido da jimlolin tushe masu kama da su da fassarorinsu daga TM don jimlar shigar da aka bayar. 2) Tsara waɗannan nau'i-nau'i (tushe, manufa) zuwa ingantaccen umarni bisa wani samfuri na musamman. 3) Gabatar da wannan umarni, sannan jimlar tushe ta sabuwa, ga LLM don fassara.

2.1. Ƙirar Samfuri na Umarni

Takarda tana binciko salo daban-daban na umarni don isar da aikin fassara da misalai ga LLM yadda ya kamata. An haskaka samfura biyu na farko:

Samfuri na Umarni (INSTRUCTION): Yana amfani da umarnin harshe na halitta. Misali: "Idan fassarar 'X1' daga Turanci zuwa Faransanci ita ce 'Y1' kuma fassarar 'X2' ita ce 'Y2', to menene fassarar 'X_sabon'? Ana buƙatar sakamakon fassara kawai."
Samfuri Mai Tsari (CODE): Yana amfani da tsari mafi ƙa'ida, nau'i-nau'i na maɓalli-daraja. Misali: "[src-lang]=[X1] [tgt-lang]=[Y1] [src-lang]=[X2] [tgt-lang]=[Y2] [src-lang]=[X_sabon] [tgt-lang]="

Zaɓin samfuri yana tasiri sosai ga aikin LLM, tare da samfuran da aka tsara galibi suna haifar da fitarwa masu daidaito ta hanyar rage shubuha.

2.2. Tsarin TMP-LM

Ana iya tattara tsarin ginshiƙi. Idan aka ba da jimla $x$, aikin ma'ajiyar TM $R(x)$ yana samun nau'i-nau'i $k$ mafi kama da tushe-manufa $(x_i^{tm}, y_i^{tm})$. Aikin mai gina umarni $C(\{(x_i^{tm}, y_i^{tm})\}_{i=1}^k, x)$ yana tsara waɗannan zuwa cikakken umarni $P$. LLM, wanda aka nuna da $M$, sannan ya samar da fassarar: $\hat{y} = M(P)$.

Ingancin ya dogara ne da ikon LLM na yin bincike na kwatankwacin a cikin mahallin—gano tsarin da aka bayar a cikin misalan da aka bayar kuma ya yi amfani da shi ga sabuwar tambaya.

3. Tsarin Gwaji & Sakamako

3.1. Bayanan Gwaji da Ma'auni

An gudanar da gwaje-gwaje akan ayyukan fassara a cikin harsuna da yawa (misali, Turanci-Jamusanci, Turanci-Sinanci) da yankuna (Doka, IT, Likita). Babban LLM da aka yi amfani da shi shine OpenAI's text-davinci-003. Ma'auni sun haɗa da ƙaƙƙarfan, tsarin NMT na musamman na yanki waɗanda aka horar da su akan manyan tarin harsuna biyu.

Fitattun Gwaje-gwaje

Samfuri: GPT-3.5 (text-davinci-003)
Ma'aunin Ƙima: Maki BLEU
Muhimmin Kwatanta: TMP-LM vs. Tsarin NMT na Zamani na Yanki

3.2. Muhimman Sakamako da Bincike

Sakamakon ya kasance mai ban mamaki:

Babban Ribar Maki BLEU: Yin amfani da ingantattun umarnin TM ya inganta aikin fassara na LLM na sifili ta maki BLEU 20 zuwa 30 a cikin ayyuka daban-daban. Wannan yana canza LLM daga matsakaicin mai fassara zuwa ƙwararren mai fassara.
Gasawa tare da SOTA NMT: Aikin LLM da aka umarta ya kasance yayi kama da, kuma wani lokacin ya wuce, na tsarin NMT na zamani waɗanda aka horar da su musamman akan manyan bayanan cikin yanki. Wannan babban bincike ne, domin yana nuna cewa LLMs tare da umarni masu dacewa za su iya dacewa da aikin samfuran na musamman ba tare da horo na musamman ba.
Hankali na Samfuri: Samfurin da aka tsara (CODE) gabaɗaya ya haifar da fassarori masu dogaro da inganci fiye da samfurin harshe na halitta (INSTRUCTION), yana jaddada mahimmancin ingantaccen aikin umarni.

Bayanin Jadawali (A ɓoye): Jadawali na sanduna zai nuna ƙungiyoyi uku ga kowane nau'in harshe/yanki: 1) LLM Sifili (BLEU ƙasa), 2) LLM + TMP-LM (BLEU mai girma sosai), 3) Ma'auni na SOTA NMT (BLEU mai girma, kama da ƙungiya 2). Sandunan ƙungiyoyi na 2 da 3 za su kasance sun yi kama da juna, dukansu sun fi girma fiye da ƙungiya 1.

4. Binciken Fasaha & Muhimman Fahimta

Muhimman Fahimta: Bayanin ƙwararrun takarda shine cewa ƙarfin fassara na LLM ba a tsare ba amma aiki ne na mahallinsa. Samfurin danye mai fassara ne mara kyau, amma lokacin da aka shuka mahallinsa tare da misalan fassara masu dacewa, masu aminci (TMs), yana buɗe aikin da ya yi kama da tsarin NMT na musamman. Wannan yana sake tsara LLMs daga samfuran tsaye zuwa injunan fassara masu ƙarfi, masu shirye-shirye na mahallin. Ya yi daidai da babban canjin tsarin da masu bincike a Cibiyar Bincike kan Samfuran Tushe ta Stanford suka haskaka, waɗanda suka nuna cewa "ilimi" da "ƙwarewar" samfuri suna ƙara bayyana ta hanyar kunna tushen umarni maimakon ma'auni na tsaye kawai.

Tsarin Hankali: Hujja tana da kyau kuma tana jan hankali. 1) LLMs suna da ƙwarewar koyo mai ƙarfi a cikin mahallin da ikon bin umarni (kamar yadda aka nuna a cikin ayyuka kamar "Horar da samfuran harshe don bin umarni tare da amsawar ɗan adam" na Ouyang et al.). 2) Fassara aiki ne da aka bayyana yadda ya kamata wanda za a iya bayyana shi ta hanyar misalai. 3) TMs an tsara su, nau'i-nau'i masu inganci. 4) Don haka, gabatar da TMs a matsayin misalan a cikin mahallin ya kamata, kuma yana, inganta ingancin fassara sosai. Hankali yana da tsari kuma shaidar gwaji tana da ƙarfi.

Ƙarfi & Kurakurai: Ƙarfin ba shakku ne: hanya mai sauƙi, mara kutsawa tana haifar da riba mai yawa. Yana daidaita ingantaccen MT ta hanyar amfani da kadarorin TM da ke akwai da LLMs na kasuwa. Duk da haka, kurakuran suna cikin dogaro. Na farko, yana dogaro sosai akan inganci da dacewar daidaitattun TM—shara a ciki, shara a waje. Na biyu, ya gaji duk iyakokin LLM: farashi, jinkiri, da ƙuntatawa na taga mahallin (kamar matsalar "Bace-a-tsakiya" da Liu et al. suka gano). Na uku, kamar yadda takarda ta nuna, hanyar tana da rauni; samfurin umarni mara kyau zai iya rage aiki. Ya fi alchemy fiye da injiniya a wannan mataki.

Fahimta Mai Aiki: Ga masu aiki, wannan kira ne mai haske don daina kallon LLMs a matsayin masu fassara na waje kuma a fara kallon su a matsayin tsarin da za a iya inganta umarni. Saka hannun jari dole ne ya canza daga horon samfuri zuwa gina ingantattun tsarin ma'ajiyar bayanai don TMs da haɓaka daidaitattun samfuran umarni, ingantattun samfuran umarni don yankuna daban-daban (kamar yadda al'umma ta daidaita daidaitawar BERT). Ga masu bincike, iyakar gaba ita ce sanya wannan tsari ya zama mai ƙarfi da inganci—bincika yadda za a matsar da ilimin TM zuwa cikin umarni mafi inganci ko yadda za a haɗa umarni tare da daidaitawar sauƙi don rage tsawon mahallin da farashi.

5. Tsarin Bincike: Misali Ba na Lamba ba

Yi la'akari da kamfanin fassarar doka tare da babban TM na sassan kwangila. A baya, tsarin NMT zai buƙaci sake horarwa akan sabbin bayanan doka don ingantawa. Tare da TMP-LM:

Shigarwa: Sabuwar jimlar tushe: "La cláusula de indemnización sobrevivirá a la terminación de este Acuerdo."
Ma'ajiyar Bayanai: Tsarin yana bincika TM na doka kuma ya sami sassan da suka yi kama da juna, waɗanda aka fassara a baya:
- TM1: Tushe: "This confidentiality obligation shall survive the expiration of the contract." → Manufa: "La obligación de confidencialidad sobrevivirá a la expiración del contrato."
- TM2: Tushe: "The warranty shall survive delivery and inspection." → Manufa: "La garantía sobrevivirá a la entrega y la inspección."

Gina Umarni (salo na CODE): Tsarin ya gina wannan umarni don LLM:

[src-lang]=[This confidentiality obligation shall survive the expiration of the contract.] [tgt-lang]=[La obligación de confidencialidad sobrevivirá a la expiración del contrato.]
[src-lang]=[The warranty shall survive delivery and inspection.] [tgt-lang]=[La garantía sobrevivirá a la entrega y la inspección.]
[src-lang]=[The indemnity clause shall survive termination of this Agreement.] [tgt-lang]=

Fitarwa: LLM, yana gane tsarin ("X zai rayu Y" → "X sobrevivirá a Y"), yana samar da fassarar mai daidaitaccen salo da daidaitaccen doka: "La cláusula de indemnización sobrevivirá a la terminación de este Acuerdo."

Wannan tsari yana mai da LLM zuwa mataimakin fassara mai sanin mahallin wanda ya bi ƙa'idodin kalmomi da salon kamfanin.

6. Aikace-aikace na Gaba & Hanyoyin Bincike

Tsarin Haɗin kai mai Ƙarfi: Tsarin MT na gaba na iya canzawa cikin sauƙi tsakanin NMT da aka daidaita don rubutu na gaba ɗaya da TMP-LM don yankuna masu wadataccen TMs (doka, likita, fasaha), yana inganta inganci da farashi.
Bayan TMs na Harsuna Biyu: Faɗaɗa ra'ayi zuwa tunawa da fassarar harsuna da yawa, yana ba da damar fassarar pivot kaɗan ko daidaitawar salo a cikin harsuna da yawa.
Koyo Mai Ƙarfi & Tsarin TM: Yin amfani da makin amincewar LLM ko rashin yarda da TMs da ke akwai don alamar kurakurai masu yuwuwa a cikin TMs na ɗan adam ko don ba da shawarar sabbin shigarwa ga masu gyara bayan ɗan adam, ƙirƙirar madauki na fassara mai inganta kansa.
Haɗawa tare da Ƙananan LLMs na Musamman: Yin amfani da TMP-LM ga LLMs masu inganci, buɗe tushe (kamar Llama ko Mistral) waɗanda aka daidaita musamman don ayyukan fassara, rage dogaro ga manyan, manufa gabaɗaya, da APIs masu tsada.
Ma'auni na Daidaitaccen Umarni: Al'umma tana buƙatar ma'auni kamar "Prompt-MT" don tantance dabarun umarni daban-daban don fassara a cikin LLMs daban-daban, kama da rawar WMT don NMT na gargajiya.

7. Nassoshi

Mu, Y., Reheman, A., Cao, Z., et al. (2023). Ƙarfafa Masu Fassara Manyan Harsunan Harsuna ta hanyar Tunawa da Fassara. arXiv preprint arXiv:2305.17367.
Ouyang, L., Wu, J., Jiang, X., et al. (2022). Horar da samfuran harshe don bin umarni tare da amsawar ɗan adam. Ci gaba a cikin Tsarin Bayanai na Jijiya, 35.
Khandelwal, U., Levy, O., Jurafsky, D., et al. (2021). Gabaɗaya ta hanyar tunawa: Samfuran harshe na kusa da juna. Taron Ƙasa da Ƙasa akan Wakilcin Koyo (ICLR).
Bommasani, R., Hudson, D. A., Adeli, E., et al. (2021). A kan damammaki da haɗarin samfuran tushe. Cibiyar Bincike kan Samfuran Tushe ta Stanford.
Liu, N. F., Lin, K., Hewitt, J., et al. (2023). Bace a tsakiya: Yadda samfuran harshe ke amfani da dogon mahallin. arXiv preprint arXiv:2307.03172.
Reheman, A., Cao, Z., Li, B., et al. (2023). Koyo guda ɗaya don fassarar injin jijiya tare da tunawa da fassara. Binciken Ƙungiyar Lissafi na Computational.