미리보기 생성 중...
From traditional to modern approaches in IR o Traditional approach RSV(d,q) o Modern approach (guided by machine learning) f(d,q) n Conducting query and document understanding n Representing query and document as multiple feature vectors n Calculating multiple matching scores between query and document n Training ranker with matching scores as features using learning to rank Recherche d'information textuelle 266
Deep learning for semantic matching o Pioneer work for text semantics n Word embeddings o Skip-gram & CBOW o Following methods: Gloves, FastText, … o Principle: learn word representations using contextual informations o Unsupervised methods (not task-driven) Recherche d'information textuelle 268
© http://nn4ir.com/ Dual Embedding Space Model (DESM) [Nalisnick et al., 2016] Recherche d'information textuelle 271
© http://nn4ir.com/ Dual Embedding Space Model (DESM) [Nalisnick et al., 2016] Recherche d'information textuelle 272
© http://nn4ir.com/ Neural Translation Language Model [Zuccon et al., 2015] © Zuccon et al. 2015 Recherche d'information textuelle 273
© http://nn4ir.com/ Neural Translation Language Model [Zuccon et al., 2015] Recherche d'information textuelle 274
© http://nn4ir.com/ Pre-trained word embedding for query term weighting [Zheng and Callan, 2015] Recherche d'information textuelle 275
© http://nn4ir.com/ Pre-trained word embedding for query expansion Recherche d'information textuelle 276
© http://nn4ir.com/ Pre-trained word embedding for query expansion Recherche d'information textuelle 277
From document/query representations to end-to-end neural IR models o Document/query representations are generally learnt independently from the retrieval task n Well designed for semantic matching n Not specifically designed for relevance matching o One solution is to guide representation according to the retrieval task n Goal of retrieval task: ranking n Need to build neural network able to rank documents (learning-to-rank) Recherche d'information textuelle 278
© Nuyen End-to-end neural IR approaches n Representation-based n Interaction-based neural neural models models © Nguyen 2018 Recherche d'information textuelle 279
Learning-to-rank with neural models o Triplet loss ©https://filebox.ece.vt.edu/~jbhuang/te aching/ece6554/sp17/lectures/Lecture_0 8_Siamese_Triplet_Networks.pdf ! = max(cos *,, + −cos *,, − +/01234,0) Recherche d'information textuelle 280
Representation-based neural models Deep Semantic Matching Model (Huang et al, 2014) Recherche d'information textuelle 281
Representation-based neural models Deep Semantic Matching Model (Huang et al, 2014) o Word representations: letter trigram o Document scoring o Loss Recherche d'information textuelle 282
Representation-based neural models Deep Semantic Matching Model (Huang et al, 2014) n Dataset: commercial search engine (over 16.000 queries) Recherche d'information textuelle 283
© http://nn4ir.com/ Representation-based neural models An extension with convolution (sentence level) Recherche d'information textuelle 284
앞 20페이지만 표시됩니다. 전체를 보려면 자료를 다운로드하세요.
댓글이 없습니다. 첫 댓글을 남겨보세요!
댓글 (0)
댓글이 없습니다. 첫 댓글을 남겨보세요!