対話状態推定のための外部知識ベースを利用した意味的素性の提案

全文

(1)Vol.2015-SLP-109 No.13 2015/12/2. IPSJ SIG Technical Report. 1. 1. 1. 1. Graham Neubig1. Dialog State Tracking Challenge 4 (DSTC4) Term Memory (LSTM). 1. Shared Task Long-Short. DSTC4 Wikidata. Wikidata. 1.. [1] Wikidata Wikidata. Dialog State Tracking Challenge (DSTC) Shared Task. DSTC3. DSTC DSTC4. 2. LSTM DSTC. Recurrent Neural Networks. (RNN) [2][3]. RNN. DSTC DSTC4. (Long-Short. LSTM. Term Memory) LSTM. [4]. RNN. LSTM 1. Graduate School of Information Science, Nara Institute of Science and Technology. ⓒ 2015 Information Processing Society of Japan. DSTC4 LSTM. 1.

(2) Vol.2015-SLP-109 No.13 2015/12/2. IPSJ SIG Technical Report. doc2vec[5][6]. 1. 300. 1 DSTC4. LSTM. MAIN. SUBCATEGORY. AREA. COUNTRY,. CITY,. DISTRICT,. NEIGHBORHOOD DET. 3.. ACCESS,. BELIEF,. BUILDING,. EVENT, PRICE, NATURE, HISTORY,. Wikidata Wikidata. FEE. MONUMENT,. ATTRACTION, SERVICES, PROD-. Wikipedia. UCTS. [7] Wikidata. labels. aliases. MEAL,. STROLL, VIEW. descriptions. FOOD. —. LOC. TEMPLE, RESTAURANT, SHOP, CULTURAL, GARDEN, ATTRAC-. claims. TION, HOTEL, WATERSIDE, EDUCATION, ROAD, AIRPORT TIME. DATE, INTERVAL, START, END, OPEN, CLOSE. labels. TRSP. STATION, TYPE. WEATHER. —. aliases. SUBCATEGORY MAIN. 1 1. Belgium. descriptions(. country. ). country. 1. area. 1 3.2 Wikidata. claims. labels. 3.1 DSTC4 [8]. 1 2 claims ”. Ramen. ”instance of Japanese cuisine doc2vec. Wikidata. claims. descriptions. 2 3.2.1 3. 1. MAIN. SUBCATEGORY Wikidata. claims. descriptions 1. 44. Wikidata claims 1500. 287. SUBCATEGORY. ⓒ 2015 Information Processing Society of Japan. 2.

(3) Vol.2015-SLP-109 No.13 2015/12/2. IPSJ SIG Technical Report. 2. Sem. Tags Sem. Tags + Rel. Words Accuracy Precision Recall F1 0.0239. 0.300. 0.287. 0.293. + Sem. Tags. 0.0239. 0.303. 0.292. 0.297. + Sem. Tags. 0.0244. 0.301. 0.288. 0.295. + Rel. Words 2 3. TRANSPORTATION. + Sem. Tags. N. F1. 471. 0.425. 513. 0.524. 42. 0.0990. • Accuracy. 3. • Commons category. • Precision. • instance of. • Recall. • part of Wikidata. claims 1. 1. • F-measure Precision. Recall 3. (1) 3. Belgium Europe. claims. (Sem. Tags). (2) (Sem. Tags + Rel. Words). United Nations (3). 1 3.2.2 1. Wikidata 4.2 7 (Sem. Tags). • subclass of. • Commons category. (Sem. Tags + Rel. Words). • capital. 2. 4.2.1. • continent. (Sem. Tags) F. • part of. 0.4%. • member of. • topic’s main category. TRANSPORTATION TYPE. 3. doc2vec 4.2.2 4.2.1. 4.. Wikidata (Sem. tags +. Rel. Words) F 2. 0.1% TRANSPORTATION. 4.2.1 TYPE. 4.2.1 4.2.3 4.1. 4 4. ”topic’s. main category”. ⓒ 2015 Information Processing Society of Japan. 3.

(4) Vol.2015-SLP-109 No.13 2015/12/2. IPSJ SIG Technical Report. 4 Accuracy. Precision. Recall. F1. subclass of. 0.0246. 0.300. 0.285. 0.292. Commons category. 0.0179. 0.281. 0.309. 0.294. capital. 0.0230. 0.303. 0.288. 0.295. continent. 0.0234. 0.301. 0.291. 0.296. part of. 0.0249. 0.312. 0.287. 0.299. member of. 0.0237. 0.314. 0.292. 0.302. topic’s main category. 0.0254. 0.317. 0.292. 0.304. FROM. TO ”Commons category”. TICKET. 5%. ACCOMMODATION. ”part of”, ”mem-. ber of”, ”topic’s main category”. 3%. ACCOMMODATION. PLACE 3. 4% SHOPPING. ”continent”, ”subclass of”, ”Commons category”. ”Commons category”. ACCOMMODATION. 0.2%. ACCOMMODATION. SHOPPING TYPE OF PLACE. NEIGHBOURHOOD. 3.9% NEIGHBOURHOOD 2.9%. ”Commons. category”. NEIGH-. BOURHOOD. LSTM. 0.6%. FOOD. 5.. ”continent” ”capital” Wikidata ”part of”, ”topic’s. main category”, ”subclass of”. 3. FOOD. 1%. ”continent” ”capital” FOOD. CUISINE, TYPE OF PLACE,. PLACE ”Common category” FOOD. MEAL TIME DISH 2%. [1]. 1%. ATTRACTION. ”part of”, ”sub-. [2]. class of”, ”topic’s main category” 0.2%. ”conti-. nent”, ”capital”. 0.1%. [3]. ”Commons category” ATTRACTION. TYPE OF PLACE. TIME. [4]. 9% 7.6% [5]. PLACE 3.7%. [6]. ATTRACTION. NEIGHBOUR. TRANSPORTATION. [7]. ”Commons category”, ”topic’s main category” 4%. [8]. ”member of”. 3.2%. 3 TRANSPORTATION. Yi Ma, Paul A. Crook, Ruhi Sarikaya, Eric Fosler-Lussier: Knowledge graph inference for spoken dialog systems. In Proc. IEEE-ICASSP, 2015. Matthew Henderson, Blaise Thomson, and Steve Young: Robust dialog state tracking using delexicalised recurrent neural networks and unsupervised adaptation. In Proc. IEEE-SLT, pages 360–365. IEEE, 2014. Matthew Henderson, Blaise Thomson, and Steve Young: Word-based dialog state tracking with recurrent neural networks. In Proc. SIGDIAL, page 292, 2014. Sepp Hochreiter and J¨ urgen Schmidhuber: Long shortterm memory. Neural computation, 9(8):1735–1780, 1997. Quoc V Le and Tomas Mikolov: Distributed representations of sentences and documents. In Proc. ICML, 2014. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean: Distributed representations of words and phrases and their compositionality. In Proc. NIPS, pages 3111–3119, 2013. D. Vrandeˇcić and M. Krötzsch: Wikidata: a free collaborative knowledgebase, Communications of the ACM, vol. 57, no. 10, pp. 7885, 2014. Seokwan Kim, Luis Fernando D’Haro, Rafael E. Banchs, Jason Williams, and Matthew Henderson: The Fourth Dialog State Tracking Challenge. In Proc. IWSDS, 2016.. TYPE. 10%. ⓒ 2015 Information Processing Society of Japan. 4.

(5)