add paper, organize notes

shaoxiongji · Mar 6, 2019 · b267244a27ccf54ced5c7182ba3944434c613487 · b267244
1 parent da5867f
commit b267244a27ccf54ced5c7182ba3944434c613487
diff --git a/README.md b/README.md
diff --git a/notes/QA/EARL.md b/notes/QA/EARL.md
@@ -0,0 +1,4 @@
+  
+Motivation: identification and linking of entities; identification and linking of relations; identification of query intent; generating formal query
+  
+Preprocessing: keyword tokenizer, entity relation predictor, candidate generation  
+  
+Disambiguation: 1) GTSP solver 2) connection density, adaptive learning  
+  
+Dataset: LC-QuAD
diff --git a/notes/QA/code-mixed.md b/notes/QA/code-mixed.md
@@ -0,0 +1,3 @@
+  
+Task: code-mix simple questions KBQA  
+  
+Method: Triplet-Siamese-Hybrid CNN, TSHCNN; triplet inputs: 1) questions, 2) positive/negtive tuple, 3) questions and positive/negative tuple  
+  
+Datasets: SimpleQuestions (Bordes et al., 2015) dataset，75.9k/10.8k/21.7k training/validation/test  
diff --git a/notes/QA/multi-hop.md b/notes/QA/multi-hop.md
@@ -0,0 +1,5 @@
+  
+Method: ConceptNet multi-hop path from common sense  
+  
+Experiments: generative QA  
+  
+Datasets: NarrativeQA  
+  
+Baseline model: embedding layer, reasoning layer, model layer(self-attention, BiLSTM), answer layer (pointer-generator decoder)  
+  
+Commonsense: multi-hop path, PMI socring, choose path like beam search  
diff --git a/notes/QA/pattern-revising.md b/notes/QA/pattern-revising.md
@@ -0,0 +1,3 @@
+  
+Motivation: learn to rank subject-predicate pairs  
+  
+Method: pattern extraction, pattern revising, joint fact selection  
+  
+Datasets: SimpleQuestions, freebase(FB2M, FB5M)
diff --git a/notes/QA/strong-baselines.md b/notes/QA/strong-baselines.md
@@ -0,0 +1,4 @@
+  
+Motivation: LSTM fine-tuning -> comparable performance  
+  
+Method: entity detection: CRF with features engineering; entity linking: n-gram inverse indexing, Levenshtein Distance; relation prediction: RNNs, CNNs, logistic regression (TF-IDF, bi-gram, word embedding, relation words); evidence integration: m entities and n relations -> 1 entity-relation  
+  
+Datasets: SimpleQuestions  
+  
+Experiments: entity linking, relation prediction, end2end QA 
diff --git a/notes/RE/inner-sentence.md b/notes/RE/inner-sentence.md
@@ -0,0 +1,3 @@
+  
+Motivation: distance supervised relation extraction; false annotation, inner-sentence noise, random feature extraction is not robust.  
+  
+Method: 1) STP: Sub-Tree Parser; BGRU: Bidirectional GRU, entity-wise neural extractor; 3) transfer learning: entity classification -> relation extraction  
+  
+Experiments: held-out evaluation, manual evaluation 
diff --git a/notes/completion/mkbe.md b/notes/completion/mkbe.md
@@ -0,0 +1,4 @@
+  
+Motivation: embedding triplet and multimodal data into vector space  
+  
+Methods: encoders: multimodal data into vector; decoders: generate multi-modal values  
+  
+Datasets: MovieLens-100k, YAGO-10   
+  
+Experiments: link prediction, generating text and images
diff --git a/notes/conversation/commonsense.md b/notes/conversation/commonsense.md
@@ -0,0 +1,4 @@
+  
+Problems: OOV  -> commonsense KG -> triplet without semantic meaning of subgraph  
+  
+Proposed Mehtod: commonsense knowledge aware conversational model, CCM; subgraph, static graph attention; dynamic graph attention; encoder-decoder seq2seq  
+  
+Datasets: ConceptNet, reddit post-response  
+  
+Metirc: perplexity, entity score, crowdsourcing(appropriateness, informativeness) 
diff --git a/notes/dynamic/HyTe.md b/notes/dynamic/HyTe.md
@@ -0,0 +1,3 @@
+  
+Method: graph embedding method considering temporal scopes, represent time as a hyperplane  
+  
+Experiments: link prediction, temporal scoping  
+  
+Datasets: YAGO11k, Wikidata12k  (with time annotations)
diff --git a/notes/embedding/BootEA.md b/notes/embedding/BootEA.md
@@ -0,0 +1,4 @@
+  
+Motivation: lack of enough prior alignment  
+  
+Method: bootstrapping approach to embedding-based entity alignment; alignment editing   
+  
+Datasets: DBP15K, DWY100K  
+  
+
diff --git a/notes/embedding/KBGAN.md b/notes/embedding/KBGAN.md
@@ -0,0 +1,3 @@
+  
+Motivation: knowledge graphs typically only contain positive facts.  
+  
+Method: GAN for negtive sample generation.  
+  
+Experiments: link prediction using FB15k-237, WN18 and WN18RR  
diff --git a/notes/embedding/co-training.md b/notes/embedding/co-training.md
@@ -0,0 +1,4 @@
+  
+Motivation: multilingual KG; low coverage of entity alignment; literal description of entities   
+  
+Method: 1) KGEM: knowledge model(TransE), alignment model(MTransE); 2) DEM: attentive gated recurrent unit encoder(AGRU), cross-lingual embedding; 3) KDCoE: iterative co-training.  
+  
+Datasets: WK3160k extracted from DBPedia  
+  
+Experiments: cross-lingual entity aligment & zero-shot aligment(Hit@1, Hit@10, MRR), cross-lingual knowledge completion (proportion of ranks no larger than 10 Hit@10, mean reciprocal rank MRR)
diff --git a/notes/embedding/edge-labels.md b/notes/embedding/edge-labels.md
@@ -0,0 +1,4 @@
+  
+Motivation: network structure, semantic information of edge  
+  
+Proposed method: structural loss: context node; relational loss: edges  
+  
+Datasets: ArnetMiner, AmazonReviews  
+  
+Experiments: multi-label node classification  
diff --git a/notes/embedding/kg-geometry.md b/notes/embedding/kg-geometry.md
@@ -0,0 +1,3 @@
+  
+Motivation: to fill the gap between effectiveness of KG embeddings and their geometric understanding.  
+  
+Metrics: 1) ATM, alignment to mean; 2) Conicity; 3) VS, vector spread; 4) AVL, average vector length  
+  
+Datasets: Freebase(FB15k), WordNet(WN18)
diff --git a/notes/embedding/rule-learning.md b/notes/embedding/rule-learning.md
@@ -0,0 +1,4 @@
+  
+Motivation: learning first-order rules, scalable techniques  
+  
+Definitions: closed-pathrule, support degree of r, standard confidence, head coverage  
+  
+Proposed method: sampling method; argument embedding; co-occurrence socre function; rule evaluation  
+  
+Datasets: FB15K-237, FB75K, YAGO2s, Wikidata, DBpedia 3.8 
diff --git a/notes/few-shot/one-shot-relational.md b/notes/few-shot/one-shot-relational.md
@@ -0,0 +1,3 @@
+  
+Motivation: long-tail in KG, one-shot setting, metric learning  
+  
+Proposed method: 1) neighbor encoder: subgraph, one-hop neighbor set, encoding; 2) matching processor: LSTM encoding, similarity  
+  
+Datasets: NELL-one, Wiki-One derived from NELL, Wikidata