text stringlengths 0 164k |
|---|
nonumber sum t w in cal Dt |
log ptheta lefttmid w ell t right Eq 20 Crucially our crosslingual objective forces both taggers to share part of the parameter vector theta which allows it to represent morphological regularities between the two languages in a common embedding space and thus enables transfer of knowledge This is no different fro... |
mid w ell First we modify the initial characterlevel LSTM embedding such that it also encodes the identity of the language Second we modify the softmax layer creating a languagespecific softmax Our first architecture has one softmax as in eqtagger over all morphological tags in cal T shared among all the languages To... |
ciMi texttt idell rangle right right |
left textLSTM leftlangle texttt idell ciMi ldots ci1 texttt idell rangle right right nonumber Eq 22 This model creates a languagespecific embedding vector vi but the individual embeddings for a given character are shared among the languages jointly trained on The remainder of the architecture is held constant ... |
<s>[INST] Neural Cross-Lingual Relation Extraction Based on Bilingual Word Embedding Mapping [/INST] Relation extraction RE is an important information extraction task that seeks to detect and classify semantic relationships between entities like persons organizations geopolitical entities locations and events It provi... |
<s>[INST] Visual Natural Language Query Auto-Completion for Estimating Instance Probabilities [/INST] This work focuses on the problem of finding objects in an image based on natural language descriptions Existing solutions take into account both the image and the query BIBREF0 BIBREF1 BIBREF2 In our problem formulatio... |
<s>[INST] Translating Navigation Instructions in Natural Language to a High-Level Plan for Behavioral Robot Navigation [/INST] Enabling robots to follow navigation instructions in natural language can facilitate humanrobot interaction across a variety of applications For instance within the service robotics domain robo... |
<s>[INST] Analysis of Risk Factor Domains in Psychosis Patient Health Records [/INST] Psychotic disorders typically emerge in late adolescence or early adulthood BIBREF0 BIBREF1 and affect approximately 254 of the population BIBREF2 BIBREF3 making them one of the leading causes of disability worldwide BIBREF4 A sub... |
<s>[INST] Morphological Word Segmentation on Agglutinative Languages for Neural Machine Translation [/INST] Neural machine translation NMT has achieved impressive performance on machine translation task in recent years for many language pairs BIBREF0 BIBREF1 BIBREF2 However in consideration of time cost and space capac... |
<s>[INST] Deja-vu: Double Feature Presentation and Iterated Loss in Deep Transformer Networks [/INST] In this paper we propose the processing of features not only in the input layer of a deep network but in the intermediate layers as well We are motivated by a desire to enable a neural network acoustic model to adaptiv... |
<s>[INST] Acquisition of Inflectional Morphology in Artificial Neural Networks With Prior Knowledge [/INST] A widely agreedon fact in language acquisition research is that learning of a second language L2 is influenced by a learners native language L1 BIBREF0 BIBREF1 A languages morphosyntax seems to be no exception to... |
<s>[INST] How Does Language Influence Documentation Workflow? Unsupervised Word Discovery Using Translations in Multiple Languages [/INST] The Cambridge Handbook of Endangered Languages BIBREF3 estimates that at least half of the 7000 languages currently spoken worldwide will no longer exist by the end of this century ... |
<s>[INST] Dense Information Flow for Neural Machine Translation [/INST] Neural machine translation NMT is a challenging task that attracts lots of attention in recent years Starting from the encoderdecoder framework BIBREF0 NMT starts to show promising results in many language pairs The evolving structures of NMT mode... |
<s>[INST] Frozen Binomials on the Web: Word Ordering and Language Conventions in Online Text [/INST] Lists are extremely common in text and speech and the ordering of items in a list can often reveal information For instance orderings can denote relative importance such as on a todo list or signal status as is the case... |
<s>[INST] Casting Light on Invisible Cities: Computationally Engaging with Literary Criticism [/INST] Literary critics form interpretations of meaning in works of literature Building computational models that can help form and test these interpretations is a fundamental goal of digital humanities research BIBREF0 With... |
<s>[INST] Scalable and Accurate Dialogue State Tracking via Hierarchical Sequence Generation [/INST] A Dialogue State Tracker DST is a core component of a modular taskoriented dialogue system BIBREF7 For each dialogue turn a DST module takes a user utterance and the dialogue history as input and outputs a belief estim... |
mathbf hat textrm BiLSTMmathbf eXat mathbf hat1 |
mathbf hut textrm BiLSTMmathbf eXut mathbf hut1 |
mathbf hbt textrm BiLSTMmathbf eXbt mathbf hbt1 |
mathbf ha0 mathbf hu0 mathbf hb0 c0 |
endaligned Eq 7 where c0 is the zeroinitialized hidden state for the BiLSTM The hidden size of the BiLSTM is dm2 We concatenate the forward and the backward hidden representations of each token from the BiLSTM to obtain the token representation mathbf hktin Rdm kin lbrace aubrbrace at each time step t The hidden... |
mathbf s mathbf hs textrm CMRDmathbf ex mathbf c HB Ha Hu |
the final output matrices SHs in Rlstimes dm are concatenations of all generated mathbf s and mathbf hs respectively along the sequence length dimension where dm is the model size and ls is the generated sequence length The general structure of the CMR decoder is shown in Figure 4 Note that the CMR decoder can suppo... |
textbf h0 textrm LSTMtextbf exttextbf qt1 |
where textbf qt is the hidden state of the LSTM textbf q0 is initialized with an average of the hidden states of the belief encoder the system encoder and the user encoder which produces HB Ha Hu respectively mathbf h0 is then summed elementwise with the condition representation mathbf cin Rdm to produce mathbf h1... |
mathbf h1 mathbf h0mathbf c |
mathbf h2 mathbf h1textAttntextbeliefmathbf h1He |
mathbf h3 mathbf h2textAttntextsysmathbf h2Ha |
mathbf h4 mathbf h3textAttntextusrmathbf h3Hu |
mathbf r mathbf h1oplus mathbf h2oplus mathbf h3oplus mathbf h4 in R4dm |
where textAttnk kin lbrace textbelief textsystextusrrbrace are the attention modules applied respectively to HB Ha Hu and oplus means the concatenation operator The gradients are blocked for mathbf h1mathbf h2mathbf h3 during the backpropagation stage since we only need them to work as the supplementary mem... |
mathbf a W1Tmathbf hmathbf b1 in Rdm |
mathbf c textsoftmaxHTa in Rl |
mathbf h Hmathbf c in Rdm |
mathbf ha W2Tmathbf hmathbf b2 in Rdm |
where the weights W1in Rdm times dm W2in Rdm times dm and the bias b1in Rdm b2in Rdm are the learnable parameters The order of the attention modules ie first attend to the system and the user and then the belief is decided empirically We can interpret this hierarchical structure as the internal order for the memor... |
mathbf r1 sigma W1Tmathbf r0mathbf b1 |
mathbf r2 sigma W2Tmathbf r1mathbf b2 |
mathbf r3 sigma W3Tmathbf r2mathbf b3 |
mathbf hs sigma W4Tmathbf r3mathbf b4 |
where sigma is a nonlinear activation and the weights W1 in R4dm times dm Wi in Rdm times dm and the bias b1 in Rdm bi in Rdm are learnable parameters and 2le ile 4 The number of layers for the MLP is decided by the grid search The hidden representation of the next token mathbf hs is then 1 emitted out of the d... |
mathbf hk textdropoutmathbf hs in Rdm |
mathbf ho WkTmathbf hkmathbf bk in Rde |
mathbf ps textsoftmaxETmathbf ho in Rdv |
s textargmaxmathbf ps in R |
where the weight Wkin Rdm times de and the bias bkin Rde are learnable parameters Since de is the embedding size and the model parameters are independent of the vocabulary size the CMR decoder can make predictions on a dynamic vocabulary and implicitly supports the generation of unseen words When training the model ... |
K hthshnhm |
hxleftlbrace |
beginarraylcl |
1 OxO1 |
fracxd2xd1 textotherwise |
endarrayright |
where Ox means the Inference Time Complexity ITC of the variable x For a model having an ITC of O1 with respect to the number of slots n and values m the ITM will be a multiplier of 215x while for an ITC of On it will be a multiplier of 251 and 1143 for Omn As a convention the metric of joint goal accuracy is us... |
<s>[INST] Siamese recurrent networks learn first-order logic reasoning and exhibit zero-shot compositional generalization [/INST] Stateoftheart models for almost all popular natural language processing tasks are based on deep neural networks trained on massive amounts of data A key question that has been raised in many... |
<s>[INST] A Simple Method for Commonsense Reasoning [/INST] Although deep neural networks have achieved remarkable successes eg BIBREF1 BIBREF2 BIBREF3 BIBREF4 BIBREF5 BIBREF6 BIBREF7 BIBREF8 BIBREF9 BIBREF10 BIBREF11 BIBREF12 BIBREF13 BIBREF14 their dependence on supervised learning has been challenged a... |
<s>[INST] Counterfactual Data Augmentation for Mitigating Gender Stereotypes in Languages with Rich Morphology [/INST] One of the biggest challenges faced by modern natural language processing NLP systems is the inadvertent replication or amplification of societal biases This is because NLP systems depend on language c... |
<s>[INST] Representation of Constituents in Neural Language Models: Coordination Phrase as a Case Study [/INST] Humans deploy structuresensitive expectations to guide processing during natural language comprehension BIBREF0 While it has been shown that neural language models show similar structuresensitivity in their p... |
<s>[INST] Investigating Linguistic Pattern Ordering in Hierarchical Natural Language Generation [/INST] Spoken dialogue systems that can help users to solve complex tasks have become an emerging research topic in artificial intelligence and natural language processing areas BIBREF0 BIBREF1 BIBREF2 BIBREF3 With a we... |
<s>[INST] Deep Enhanced Representation for Implicit Discourse Relation Recognition [/INST] This work is licenced under a Creative Commons Attribution 40 International Licence Licence details httpcreativecommonsorglicensesby40 Discourse parsing is a fundamental task in natural language processing NLP which determines t... |
mathbf ui mathop maxpoolmathbf C1 cdots mathbf Cj cdots mathbf Cn |
endsplit Finally the K outputs are concatenated |
mathbf u mathbf u1 mathbf u2 cdots mathbf uK in mathbb Rds |
to feed a highway network BIBREF25 mathbf g sigma mathbf Wg mathbf uT mathbf bg in mathbb Rds nonumber |
mathbf eis mathbf g odot mathop ReLUmathbf Wh mathbf uT mathbf bh |
mathbf 1 mathbf g odot mathbf u nonumber |
in mathbb Rds Eq 6 where mathbf g denotes the gate and mathbf Wg in mathbb Rds times ds mathbf bg in mathbb Rds |
mathbf Wh in mathbb Rds times ds mathbf bh in mathbb Rds are parameters odot is elementwise multiplication The above Eq 6 gives the subwordlevel embedding for the i th word ELMo Embeddings from Language Models BIBREF26 is a pretrained contextualized word embeddings involving characterlevel representation It is shown u... |
cdots mathbf hi0 cdots cdots mathbf hi1 cdots |
mathop ELMocdots mathbf ci cdots |
where mathbf hi0 and mathbf hi1 denote the outputs of first and second layers of ELMo encoder for i th word Following Peters2018ELMo we use a selfadjusted weighted average of mathbf hi0 mathbf hi1 beginsplit |
mathbf s mathop softmaxmathbf w in mathbb R2 |
mathbf h gamma sum j01 sj mathbf hij in mathbb Rdcprime |
endsplit where gamma in mathbb R and mathbf w in mathbb R2 are parameters tuned during training and dcprime is the dimension of the ELMo encoders outputs Then the result is fed to a feed forward network to reduce its dimension mathbf eic mathbf Wc mathbf hT mathbf bc in mathbb Rdc Eq 7 mathbf Wc in mathbb Rdcp... |
mathbf zi mathbf Ai odot sigma mathbf Bi in mathbb Rde |
There is also a residual connection Res 1 in the block which means adding the output of mathop GLU and the input of the block as final output so mathbf zi mathbf xi is the output of the block corresponding to the input mathbf xi The output mathbf zi mathbf xi for all i 1 cdots N is sent to both the next layer an... |
mathbf yi mathop biGRUmathbf xi in mathbb R2de |
then this is sent to a feed forword network mathbf zi mathbf Wr mathbf yiT mathbf br in mathbb Rde Eq 10 mathbf Wr in mathbb R2de times de and mathbf br in mathbb Rde are parameters There is also a similar residual connection Res 1 in the block so mathbf zi mathbf xi for all i 1 cdots N is the final output ... |
mathbf Mj mathop FFNmathbf v1j mathbf v2jT |
in mathbb RN times N |
mathop FFN is a feed froward network similar to Eq 10 applied to the last dimension corresponding to the word Then the projected representations are beginsplit |
mathbf w2j mathop softmaxmathbf Mj mathbf v2j in mathbb RN times de |
mathbf w1j mathop softmaxmathbf MjT mathbf v1j in mathbb RN times de |
endsplit where the mathop softmax is applied to each row of the matrix We apply 2max pooling on each projected representation and concatenate them as output of the j th biattention module |
mathbf oj mathop top2mathbf w1j mathop top2mathbf w2j |
in mathbb R4 de |
The number of max pooling operation top2 is selected from experiments and it is a balance of more salient features and less noise The final pair representation is mathbf o mathbf o1 mathbf o2 cdots mathbf ol in mathbb R4 l de Eq 12 Since the output is concatenated from different layers and the outputs of lower ... |
<s>[INST] Detecting Potential Topics In News Using BERT, CRF and Wikipedia [/INST] NamedEntityRecognitionNER approaches can be categorised broadly in three types Detecting NER with predefined dictionaries and rulesBIBREF2 with some statistical approachesBIBREF3 and with deep learning approachesBIBREF4 Stanford CoreNLP ... |
<s>[INST] Gender Bias in Coreference Resolution [/INST] There is a classic riddle A man and his son get into a terrible car crash The father dies and the boy is badly injured In the hospital the surgeon looks at the patient and exclaims I cant operate on this boy hes my son How can this be That a majority of people are... |
<s>[INST] How Far are We from Effective Context Modeling ? An Exploratory Study on Semantic Parsing in Context [/INST] Semantic parsing which translates a natural language sentence into its corresponding executable logic form eg Structured Query Language SQL relieves users from the burden of learning techniques behind ... |
<s>[INST] A Novel Aspect-Guided Deep Transition Model for Aspect Based Sentiment Analysis [/INST] Aspect based sentiment analysis ABSA is a finegrained task in sentiment analysis which can provide important sentiment information for other natural language processing NLP tasks There are two different subtasks in ABSA na... |
<s>[INST] HIBERT: Document Level Pre-training of Hierarchical Bidirectional Transformers for Document Summarization [/INST] Automatic document summarization is the task of rewriting a document into its shorter form while still retaining its important content Over the years many paradigms for document summarization have... |
mathbf Ei mathbf e1i mathbf e2i dots mathbf eSii |
quad quad textwhere quad mathbf eji ewji mathbf pj |
endsplit Eq 6 where ewji and mathbf pj are the word and positional embeddings of wji respectively The word embedding matrix is randomly initialized and we adopt the sinecosine positional embedding BIBREF6 Then the sentence encoder a Transformer transforms mathbf Ei into a list of hidden representations mathbf h1i... |
tildemathbf hj1 textMultiHeadmathbf qj1 mathbf Kj1 mathbf Vj1 |
mathbf qj1 mathbf WQ tildemathbf ej1k |
mathbf Kj1 mathbf WK widetilde mathbf E k1j1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.