Challenges to provide great user experience using natural language processing model in B2B product on the web

$IBMMFOHFTUPQSPWJEFHSFBUVTFSFYQFSJFODF VTJOHOBUVSBMMBOHVBHFQSPDFTTJOHNPEFM JO##QSPEVDUPOUIFXFC 3ZVJDIJ5BOJNPUP 4UPDLNBSL *OD

4UPDLNBSL *OD *OUSPEVDUJPO • 1SPWJEFTFBSDIBOEBOBMZTJTXFCBQQ PGCVTJOFTTOFXTBMMPWFSUIFXPSME • 4VQQPSUZPVSNBSLFUSFTFBSDIPS DPNQFUJUJWFBOBMZTJTJOUIFCVTJOFTT TDFOF
• 8JUIOBUVSBMMBOHVBHFQSPDFTTJOH BOENBDIJOFMFBSOJOH • 1&45BOBMZTJT • $PNQBOZOBNFFYUSBDUJPO • 2VFSZTVHHFTUJPO

4UPDLNBSL *OD 2VFSZTVHHFTUJPOXJUINBDIJOFMFBSOJOH 3FMFWBOUUPUIFJOQVU JOUIFCVTJOFTTTDFOF 4JNJMBSUP UIF JOQVU BOEEFSJWFEGSPN TFBSDIIJTUPSZ
4FBSDI#BS 4VHHFTUJPOT 4VHHFTUJPOT

4UPDLNBSL *OD 0VUMJOFPGRVFSZTVHHFTUJPOQSPDFTT 1. Generate a language model beforehand (backend)
2. Calculate query suggestions in real-time (backend or frontend) Large corpus of the web news Trained language model Search query Trained language model Word vector Similar words in the vocabulary Similar words in the search history

4UPDLNBSL *OD #BDLFOEPSGSPOUFOE #BDLFOETFSWFS 'SPOUFOE #SPXTFS • /POFUXPSLMBUFODZJOSFBMUJNFTVHHFTUJPO • 1SPUFDUJPOPGTFBSDIIJTUPSZBTBDPOGJEFOUJBM
JOGPSNBUJPO • -PDBM NBDIJOFEFQFOEFOU QFSGPSNBODF • 3FTPVSDF NBOBHFNFOU • 6TFPGPSJHJOBMJNQMFNFOUBUJPOPGMJCSBSJFT • /FUXPSLMBUFODZ

4UPDLNBSL *OD • Build WASM binary 'SPOUFOEJNQMFNFOUBUJPO C++ WASM &NTDSJQUFO

4UPDLNBSL *OD • Calculate query suggestions in browser 'SPOUFOEJNQMFNFOUBUJPO Trained
language model Search query WASM trained language model JavaScript Word vector Similar words Suggestions JavaScript Trained language model

4UPDLNBSL *OD 1FSGPSNBODF #BDLFOETFSWFS 'SPOUFOE #SPXTFS • 8JUIJONT GPSTVHHFTUJPOTQFSJOQVUDIBOHF •
4PNFUJNFTUPPLNPSFUIBOTFDPOETUP EPXOMPBE TUPSFBOEJOJUJBMJ[FMBOHVBHFNPEFMJO UIFGJSTUQBHFMPBEJOHBGUFSBOVQEBUFPG MBOHVBHFNPEFM • .PSFUIBONT GPSTVHHFTUJPOTQFSJOQVU DIBOHF

4UPDLNBSL *OD 1FSTQFDUJWF • 3VOOJOHNBDIJOFMFBSOJOHQSPDFTTFTXJUIMBOHVBHFNPEFMPOUIFXFCJTUIFGBTDJOBUJOH XBZUP • QSPWJEFBTFBNMFTTBOEJOUFSBDUJWFVTFSJOUFSGBDF • QSPUFDUBDPOGJEFOUJBMJOGPSNBUJPO
• "TMBOHVBHFNPEFMTDPVMECFMBSHF PWFS.# BOEJUXPVMEUBLFNVDIUJNFUPEP JOJUJBMMPBEJOH XFTIPVMEUSZUP • JNQSPWFOFUXPSLMBUFODZ CBDLFOEGSPOUFOE • SFEVDFUIFNPEFMTJ[FUPCFPQUJNJ[FEUPCSPXTFSFOWJSPONFOU CBDLFOEGSPOUFOE • TQFFEVQEBUBTUPSF*0 GSPOUFOE • SVOUIFXIPMFJOJUJBMJ[BUJPOQSPDFTTJOUIFCBDLHSPVOE GSPOUFOE

5IBOL:PV • 5XJUUFS!SZVJDIJUU • -JOLFE*OSZVJDIJUU • 4UPDLNBSL*OD

Challenges to provide great user experience using natural language processing model in B2B product on the web

Challenges to provide great user experience using natural language processing model in B2B product on the web

Ryuichi Tanimoto

More Decks by Ryuichi Tanimoto

Other Decks in Technology

Featured

Transcript

$IBMMFOHFTUPQSPWJEFHSFBUVTFSFYQFSJFODF VTJOHOBUVSBMMBOHVBHFQSPDFTTJOHNPEFM JO##QSPEVDUPOUIFXFC 3ZVJDIJ5BOJNPUP 4UPDLNBSL *OD

4UPDLNBSL OD OUSPEVDUJPO • 1SPWJEFTFBSDIBOEBOBMZTJTXFCBQQ PGCVTJOFTTOFXTBMMPWFSUIFXPSME • 4VQQPSUZPVSNBSLFUSFTFBSDIPS DPNQFUJUJWFBOBMZTJTJOUIFCVTJOFTT TDFOF

4UPDLNBSL *OD 2VFSZTVHHFTUJPOXJUINBDIJOFMFBSOJOH 3FMFWBOUUPUIFJOQVU JOUIFCVTJOFTTTDFOF 4JNJMBSUP UIF JOQVU BOEEFSJWFEGSPN TFBSDIIJTUPSZ

4UPDLNBSL *OD 0VUMJOFPGRVFSZTVHHFTUJPOQSPDFTT 1. Generate a language model beforehand (backend)

4UPDLNBSL *OD #BDLFOEPSGSPOUFOE #BDLFOETFSWFS 'SPOUFOE #SPXTFS • /POFUXPSLMBUFODZJOSFBMUJNFTVHHFTUJPO • 1SPUFDUJPOPGTFBSDIIJTUPSZBTBDPOGJEFOUJBM

4UPDLNBSL *OD • Build WASM binary 'SPOUFOEJNQMFNFOUBUJPO C++ WASM &NTDSJQUFO

4UPDLNBSL *OD • Calculate query suggestions in browser 'SPOUFOEJNQMFNFOUBUJPO Trained

4UPDLNBSL *OD 1FSGPSNBODF #BDLFOETFSWFS 'SPOUFOE #SPXTFS • 8JUIJONT GPSTVHHFTUJPOTQFSJOQVUDIBOHF •

4UPDLNBSL *OD 1FSTQFDUJWF • 3VOOJOHNBDIJOFMFBSOJOHQSPDFTTFTXJUIMBOHVBHFNPEFMPOUIFXFCJTUIFGBTDJOBUJOH XBZUP • QSPWJEFBTFBNMFTTBOEJOUFSBDUJWFVTFSJOUFSGBDF • QSPUFDUBDPOGJEFOUJBMJOGPSNBUJPO

5IBOL:PV • 5XJUUFS!SZVJDIJUU • -JOLFEOSZVJDIJUU • 4UPDLNBSLOD