Content 此篇論文主要完成了:1.通過數(shù)學(xué)推導(dǎo)汰蓉,找到并證明了限制RNN-based LMs的性能瓶頸之一——Softmax Bottleneck問題2.針對(duì)這個(gè)瓶頸,提...
Content 此篇論文主要完成了:1.通過數(shù)學(xué)推導(dǎo)汰蓉,找到并證明了限制RNN-based LMs的性能瓶頸之一——Softmax Bottleneck問題2.針對(duì)這個(gè)瓶頸,提...
相信很多人都遇到過訓(xùn)練一個(gè)deep model的過程中辱匿,loss突然變成了NaN键痛。在這里對(duì)這個(gè)問題做一個(gè)總結(jié)。 一般來說匾七,出現(xiàn)NaN有以下幾種情況: 如果在迭代的100輪以內(nèi)...
2018 · NAACL · BP · Deep Contextualized Word Representations 問題:以往的詞向量表示的語義信息并不豐富絮短,沒有挖掘到...
2018 · NIPS · Learned in Translation: Contextualized Word Vectors 想法來源:從CV中借鑒的想法,深層的神經(jīng)網(wǎng)...
2018 · ICLR · A NEW METHOD OF REGION EMBEDDING FOR TEXT CLASSIFICATION 想法來源:為了挖掘深層的語法信息...
2018 · ICLR · SKIP RNN: LEARNING TO SKIP STATE UPDATES IN RECURRENT NEURAL NETWORKS 想法來...
2018· ACL · Unsupervised Learning of Sentence Embeddings using Compositional n-Gram Fea...
2018 · ACL · Some of Them Can be Guessed! Exploring the Effect of Linguistic Context in...
2018 · ACL · LSTMs Exploit Linguistic Attributes of Data 想法來源:實(shí)驗(yàn)性質(zhì)文章昨忆。主要是測(cè)試LSTM的記憶能力丁频。 價(jià)值...
2018 · ACL · Exploring Semantic Properties of Sentence Embeddings 想法來源:探索性實(shí)驗(yàn)性質(zhì)文章。 價(jià)值:可以...
2017 · ICLR · EFFICIENT VECTOR REPRESENTATION FOR DOCUMENTS THROUGH CORRUPTION 想法來源:文檔表...
2018 · ACL · What you can cram into a single $&!#* vector-Probing sentence embeddings f...
2017 · EMNLP · Supervised Learning of Universal Sentence Representations from Natural L...
2017 · ICLR · A STRUCTURED SELF-ATTENTIVE SENTENCE EMBEDDING · 想法來源:監(jiān)督任務(wù)+self-attention...
2016 · ICLR · IMPROVING PERFORMANCE OF RECURRENT NEURAL NETWORK WITH RELU NONLINEARITY ...
2018 · ICLR · AN EFFICIENT FRAMEWORK FOR LEARNING SENTENCE REPRESENTATIONS · 想法來源:文章從分布...
2017 · Arxiv · Discourse-Based Objectives for Fast Unsupervised Sentence Representation...
想法來源:很多的詞向量表達(dá)都是將詞作為單一體來看待的邑贴,沒有考慮參數(shù)的共享席里。從感知角度來說,就是失去了原有的內(nèi)部結(jié)構(gòu)拢驾。 價(jià)值:a. 學(xué)習(xí)到詞與詞形態(tài)學(xué)方面的信息奖磁,詞與詞交集越多...
EMNLP · 2017 · Mimicking Word Embeddings using Subword RNNs · 想法來源:subword 價(jià)值:利用了subwor...