Riba:
- Ikon fassara nau'ikan panel na azanci, gwajin hedonic / siffatawa da amincin panelist
- Ikon tsara taƙaitaccen bayanan azanci, cire jigo da fassarar ƙididdiga tare da AI
- Ability don tabbatar da fassarar ƙididdiga na AI tare da cikakkun bayanan panel da ƙirar gwaji
Kuna cikin dakin gwaje-gwaje na R&D don sabon yoghurt mai ƙarancin sukari. Ƙungiyar tallace-tallace ta tambaya "Shin masu amfani suna son shi?" Ya tambaya, kuma samarwa yana tambaya, "Shin za a iya bambanta shi da samfurin tunani?" yana mamaki. Yana da fom ɗin hedonic mai maki 9 da masu amfani da mabukaci 60 suka cika, ƙididdige QDA daga komitin siffanta mutum 8 da aka horar da shi, da sakamakon gwajin nuna wariya na triangle. Daruruwan layuka na danyen maki a cikin Excel, da akwatin sharhi mai buɗewa ga kowane ɗan majalisa. Da alama abin sha'awa ne a tambayi mataimakin AI "ya taƙaita wannan bayanan, masu amfani suna son shi?" A cikin wannan rukunin, za mu yi nazarin yadda ake karanta bayanan azanci daidai, amfani da AI don taƙaitawa da cire jigo, kuma mafi mahimmanci, inganta fassarar ƙididdiga ta AI tare da ɗanyen bayanan panel da ƙirar gwaji.
Nau'in gwajin jijiya: yi tambaya daidai tare da gwajin da ya dace
Kuskuren da ya fi kowa a cikin bincike na hankali shine rikita nau'in tambaya da nau'in gwaji. Akwai manyan iyalai guda uku kuma kowannensu yana amsa wata tambaya daban.
- Gwajin Hedonic (tasiri): "Nawa aka fi so?" ya amsa tambayar. Kayan aiki na yau da kullun shine sikelin hedonic mai maki 9 (1 = ba a so sosai, 5 = ba a so ko ba a so, 9 = ana so sosai). Mahalarta taron masu amfani ne marasa ilimi; Manufar ita ce auna yarda / fifiko. Gabaɗaya, ana buƙatar kwamiti na mutane 50-100.
- Gwaje-gwaje masu bayyanawa (QDA): "Wane fasali ne samfurin ke da shi kuma a cikin wane ƙarfi?" ya amsa tambayar. Kwamitin horarwa na mutane 8-12 yana ƙididdige halayen da suka bayyana (misali "daidaitawar kirim," "mai tsami," "dandan 'ya'yan itace") akan ma'aunin ƙarfi na 0-15. Ba ya auna abubuwan so, yana ƙirƙirar bayanan martaba.
- Gwajin wariya: "Shin samfuran biyu sun bambanta da fahimta?" ya amsa tambayar. A cikin gwajin triangle, ana ba da samfurori guda uku ga panel; biyu iri daya ne, daya daban, shi kuma mai gabatar da kara ya zabi “na daban”. Mafi dacewa don gwada ko canji a cikin tsari yana bayyane.
Tukwici: Kafin zaɓar gwajin, cika jimlar ku: "Ina so in sani ko ____." Idan "ana son tazarar", yi amfani da gwajin hedonic, idan "bambanta", yi amfani da gwajin nuna bambanci, kuma idan "mai ƙarfi a cikin wane fasalin", yi amfani da gwajin siffatawa. Idan ba ku ƙididdige wannan dalili ba lokacin ba da bayanan ga AI, za a tsara taƙaitawar ta hanyar da ba daidai ba.
Amincewar yan majalisa da ƙirar gwaji
Kafin ka iya amince da danyen maki, dole ne ka amince da panel kanta. Wasu ƙa'idodi na asali:
- Gwajin makafi: Mai gabatar da shirin bai kamata ya san wane samfurin shine "sabon samfurin" ba kuma wanda shine "reference". Ana gabatar da misalai tare da lambobin bazuwar lambobi 3 (misali 417, 682).
- Diyya na odar gabatarwa: Samfurin farko da aka ɗanɗana yana da fa'ida gabaɗaya (samfurin na farko). Ya kamata a daidaita odar gabatarwa a tsakanin masu fafutuka.
- Maimaitawa: Idan mai ba da shawara iri ɗaya ya sake ƙima samfurin guda tare da lambobi daban-daban, zaku iya auna daidaito (maimaituwa). A cikin kwamitin da aka kwatanta, an sake horar da mawallafin da bai dace ba ko kuma an cire shi.
- Bincika Bincika: Idan akwai samfurori iri ɗaya da aka gabatar a makance kuma mai gabatar da shirin ya ƙididdige su daban-daban, ana zargin bayanan ɗan kwamitin.
Misalin taƙaitaccen bayanan hedonic
Da ke ƙasa akwai taƙaitaccen tebur na gwajin hedonic don masu fafutuka 60. Kwatanta sabuwar dabara (lambar 417) tare da tunani (lambar 682).
fasali
Sabuwar dabara (417) matsakaici.
Magana (682) matsakaici.
Std. karkata (417)
n
Gabaɗaya yabo
7.1
7.4
1.3
60
Dandanna/kamshi
6.8
7.3
1.5
60
daidaito
7.3
7.2
1.1
60
Bayyanar
7.6
7.5
0.9
60
Nufin siya (%)
58%
65%
-
60
Yana da sauƙi a kalli wannan ginshiƙi kuma ku ce "maganin ya ɗan fi kyau, amma bambancin ƙarami ne." Amma bambancin ma'anar 0.3 yana da mahimmancin ƙididdiga ko amo? Wannan shine inda AI ke samar da mafi yawan hallucinations.
Takaitawa, cire jigo da tsara fassarar ƙididdiga tare da AI
Kuna iya amfani da AI azaman mai haɓakawa don ayyuka uku: zayyana ƙididdiga na lambobi, cire jigogi daga maganganun buɗe ido, da tsara fassarori na ƙididdiga. Amma a cikin duka ukun, abin da aka fitar daftari ne, ba yanke shawara ba.
Ƙarfi mai ƙarfi don fitar da jigogi daga sharhin da ba a buɗe ba:
Matsayi: Kai mai nazari ne na azanci. A ƙasa akwai maganganun buɗe ido daga masu amfani da 60 don yoghurt ɗin 'ya'yan itace masu ƙarancin sukari (lambar 417). Ayyukanku: 1) Ƙungiya ta yin sharhi a cikin jigogi 5-7 (misali zaƙi, miya, laushi, ɗanɗanon 'ya'yan itace).2) Kiyaye yawancin maganganun da ke da kyau / korau / tsaka tsaki ga kowane jigo kuma rubuta lambar.3) Ƙara ƙididdiga kai tsaye, taƙaitawa. KADA KA kafa jigon da ba a ambata a cikin sharhi ba.4) KADA KA SANYA ƙwararrun ƙididdiga; Wannan taƙaitaccen bayani ne. Fitowa azaman tebur: | Jigo | Mai kyau | Korau | tsaka tsaki | Sanarwa Misali | Comments:[60 comments pasted here]
Yanzu bari mu ga bambanci tsakanin rarrauna da sauri mai ƙarfi a cikin fassarar ƙididdiga:
RAUNI GASKIYA: "Bincika wannan bayanan na hankali, gaya mani idan sabon samfurin ya fi abin tunani." (AI MAY gyara "eh ya fi kyau" ko "akwai bambanci mai mahimmanci" ba tare da sanin girman samfurin ba, nau'in gwaji, p-darajar.) KARFIN KYAUTA: "A ƙasa akwai ƙimar ƙimar ƙimar hedonic na 9-point tare da ginshiƙan 60 (ginshiƙan 417 da 682). Don haɗa t-gwajin da ake buƙata, amma WRITE zai tantance sakamakon. SPSS/R. - Wanne gwaji ya dace kuma me yasa (hade ko mai zaman kanta) - Ta yaya zan kafa fassarar daban idan p <0.05 ya fito?
Ƙarfin ƙarfi yana sa mai ba da shawara ta hanyar AI; Kuna lissafta lambar.
Muhimmancin ƙididdiga da ingantaccen samfurin
Bari mu ce taƙaitaccen AI ya rubuta "sabon samfurin an ƙididdige shi sosai ƙasa da tunani." Kuna buƙatar tabbatar da wannan tare da ɗanyen bayanai. Bari mu ga dabaru na t-test da aka haɗa tare da lissafi mai sauƙi:
shigo da numpy azaman npfrom ƙididdigar shigo da scipy # gabaɗayan liking maki na 60 panelists (9-point hedonic) sabon = np.array ([7,8,6,7,7,6,8,7, ...]) # code 417, n = 60reference = np.array ([7,8,7,8,7,7,8) Mawallafin guda ɗaya ya zira kwallaye biyun samfurori -> t-testit_stat, p_value = stats.ttest_rel (sabon, reference) difference = new.mean () - reference.mean () print (f"Ma'anar bambanci: {difference:.2f} score") print (f"t = {t_stat:.2f}, p = {p_value): # shi. "Babu wani bambanci mai mahimmanci na ƙididdiga."
Mahimmin mahimmanci a nan: ma'anar bambanci na maki 0.30 na iya zama maras muhimmanci tare da p = 0.18. Maganar AI cewa "nassoshi ya fi kyau" fassarar ce da ba ta da goyan bayan ƙididdiga. Lokacin yin yanke shawara, ya kamata ku dubi p-darajar, girman samfurin, da kuma zato na gwajin, ba ma'anar kadai ba.
Tsanaki: LLMs na iya samar da tabbataccen kalamai kamar "p<0.05, bambancin yana da mahimmanci" ko da ba a ba su ɗanyen bayanan lambobi ba. Wannan hallucination ne. Kada ka rubuta kowane p-darajar, sharhi mai mahimmanci, ko "masu amfani da su suna son" hukunce-hukuncen cikin rahoton dangane da rubutun AI; sake ƙididdigewa a cikin software na ƙididdiga (R, SPSS, JASP, Python).
karamin kaso
A wani kamfani na abin sha, ƙungiyar masu azanci suna kimanta sabuwar dabarar da ke rage sukari a cikin ruwan lemu da kashi 15%. Ƙungiyar tana gudanar da gwajin hedonic mai maki 9 tare da masu amfani da 90 kuma suna ciyar da ɗanyen hoto ga AI don taƙaita sakamakon. Rahoton AI ya ce "sabuwar tsarin an fi son shi sosai (p <0.05)" kuma ƙungiyar ta yanke shawarar kawar da dabarar. Wani babban injiniya yana sake gudanar da danyen bayanai a cikin R: bambanci na gaskiya 7.0 vs 6.8, p = 0.31 - don haka babu wani babban bambanci. Bugu da ƙari, an lura cewa tsari na gabatarwa ba daidai ba ne, sabon tsarin koyaushe yana dandana na biyu kuma yana shafar gajiyar dandano (adaptation). AI duka sun yi p-darajar kuma sun kasa gano kuskuren ƙirar gwaji. Ƙungiyar ta sake maimaita gwajin tare da daidaitaccen ƙira, kuma ana karɓar tsarin ƙananan sukari. Juya zuwa ga ɗanyen bayanai ya ceci samfur mai kyau daga jefar da shi.
Kuskuren gama gari
- Gwajin rikice-rikice na hedonic (son) tare da gwajin siffa (bayanin martaba); Fadin "maki mai girma" ba yana nufin ba a son shi.
- Sanya p-darajar AI-concocted ko "bambanci mai mahimmanci" a cikin rahoton ba tare da yin ƙididdiga mai ƙima ba.
- Yin watsi da girman samfurin; Gabaɗaya "masu amfani da ita sun so" daga gwajin hedonic na mutum 12.
- Rashin daidaita tsari na gabatarwa da kuma yin watsi da samfurin farko / dandano gajiya.
- Bayar da masu ba da shawara su san wane samfurin shine "sabon samfurin" ba tare da gwajin makaho ba.
- Rashin tabbatar da cewa AI ta taƙaita maganganun buɗaɗɗen ƙayyadaddun ƙididdiga / jigogi.
- Auna duk maki daidai ba tare da duba amincin kwamitin ba (daidaicin kwafin makafi).
A takaice
- A cikin gwajin azanci, da farko ƙayyade tambaya: yi amfani da gwajin hedonic don dandano, gwajin triangle don bambanci, gwajin siffa (QDA) don bayanin martaba.
- Aminta da kwamitin kafin aminta da danyen maki: duba aminci tare da gwajin makafi, daidaita odar gabatarwa, sake kunnawa da kwafin makafi.
- Yi amfani da AI don taƙaitaccen ƙididdigewa, cire jigo daga maganganun buɗe ido, da tuntuɓar hanyar ƙididdiga; amma abin da ake fitarwa shi ne daftarin aiki.
- Bambancin ma'ana ba ɗaya bane da mahimmancin ƙididdiga; Ƙananan bambance-bambance na iya zama marasa mahimmanci tare da p-darajar.
- AI na iya samar da p-darajar, fassarar mahimmanci, da hallucination na hukuncin "yatsa"; Sake ƙididdige kowane sakamakon ƙididdiga tare da ɗanyen bayanai a cikin software ɗin ku.
- Shawarar samfur / tsari na ƙarshe; Ana ɗaukar ingantattun ƙididdiga, ƙirar gwaji mai ƙarfi, da amincin ƙungiyar tare, ba bisa rubutun AI ba.
Aikin aikace-aikace
Zana gwajin hedonic mai maki 9 da gwajin triangle don masu ba da shawara na 40-60 akan samfurin da aka yi nazari da kansa ko na hasashen (misali, miya mai rage-gishiri, kek mara amfani). Rubuta ƙirar gwajin ku: nawa mahalarta taron, makafi coding, tsarin daidaita tsarin gabatarwa, da ko za ku yi amfani da kwafin makafi. Ƙirƙiri ingantaccen tef ɗin bayanai (aƙalla layuka 20) kuma ba AI ta hanyoyi daban-daban guda biyu: (1) cire jigo daga maganganun buɗe ido, (2) shawarwarin hanyar ƙididdigewa (a zahiri neman kar a daidaita p-darajar). Sa'an nan kuma gudanar da t-gwajin da kanku a cikin Python/R/JASP kuma ku kwatanta shi da fassarar AI. A cikin bayanin kula mai shafi ɗaya: Bayyana wanne daga cikin maganganun AI ɗin da kuka tabbatar, waɗanda kuka karyata da ɗanyen bayanai, da abin da kuka dogara da shawarar samfurin ku na ƙarshe. A cikin bayanin ku, bayyana aƙalla haɗari ɗaya na son zuciya a ƙirar gwajin ku da yadda kuka rage shi.