directory stringclasses 105
values | filename stringlengths 13 32 | text stringlengths 75 1.93k | hash stringlengths 64 64 |
|---|---|---|---|
D17 | D17-1065.json | After achieving remarkable successes in Machine Translation (Sutskever et al. , 2014; Cho et al. , 2014), neural networks with the encoder-decoder architectures (a. k. a sequence-to-sequence models, Seq2Seq) have been proven to be a functioning method to model short-text conversations (Vinyals and Le, 2015; Shang et al... | 7ecd8698ea893538ca5194b557e5bbd23bcc55175f891a83f31d61c2b34ebf70 |
D17 | D17-1065.json | However, previous research has indicated that naïve implementations of Seq2Seq based conversation models tend to suffer from the so-called "safe response" problem (Li et al. , 2016a), i. e. such models tend to generate non-informative responses that can be associated to most queries, e. g. "I don't know", "I think so",... | bdd4a0704616ff3e9ca91db3bdbc2570ad035ff1091d7a2c834b4d97fe96606e |
D17 | D17-1065.json | Frequent occurrences of safe responses can dramatically reduce the attractiveness of a chat agent, which therefore should be avoided to the best extent possible when designing the learning algorithms. The pathway to achieve this purpose is to seek a more expressive model with better capacity that can take relevance and... | 63f226588467691bc350800c1c069b7abebfd3ced07f1011f0c598a2cbce4fad |
D17 | D17-1065.json | Generative Adversarial Nets (GANs) (Goodfellow et al. , 2014; Chen et al. , 2016) offers an effective architecture of jointly training a generative model and a discriminative classifier to generate sharp and realistic images. This architecture could also potentially be applied to conversational response generation to r... | 5f6cae90f6aee6d6352c9bec79d9415d883796b8723be3e4d05b9d4d8a3db69f |
D17 | D17-1065.json | To the best of our knowledge, Reinforcement Learning (RL) is first introduced to address the above problem (Li et al. , 2017; , where the score predicted by a discriminator was used as the reinforcement to train the generator, yielding a hybrid model of GAN and RL. But to train the RL phrase, Li et al. (2017) introduce... | 1226c03bdfe39e9772fe88b921ca81ff2bb36ba406f7dd0f16838a0066981931 |
D17 | D17-1065.json | In this paper, we propose a novel variant of GAN for conversational response generation, which introduces an approximate embedding layer to replace the sampling-based decoding phase, such that the entire model is continuous and dif-ferentiable. Empirical experiments are conducted based on two datasets, of which the res... | 2b756926ce345af9a67cfd8b3862c9c8e8207d9f6458c158c67afd309a3361bf |
D17 | D17-1065.json | Inspired by recent advances in Neural Machine Translation (NMT), Ritter et al. (2011) and Vinyals and Le (2015) have shown that singleturn short-text conversations can be modelled as a generative process trained using query-response pairs accumulated on social networks. Earlier works focused on paired word sequences on... | 471034c37e03ba190daa5ac79ee5439e8117766340369349492d932956821b5a |
D17 | D17-1065.json | In addition, attention mechanisms were introduced to Seq2Seq-based models to capture topic and dialog focus information by Shang et al. (2015) and, which had been proven to be helpful for improving query-response relevance. Additional features such as persona information (Li et al. , 2016b) and latent semantics Serban ... | 09cb25a935ff1ea49353310d0fd6f3f30910e5c99be4eb23dcaa9fadd5c79fda |
D17 | D17-1065.json | When compared to previous work, this paper is focused on single-turn conversation modeling, and employs a GAN to yield informative responses. | 41812092a72576efee00bebbf342b054eda6b384d8f7a185e47141cee46dce5b |
D17 | D17-1065.json | Let D = {(q i, r i )} N i=1 be a set of N single- turn human-human conversations, where q i = (w q i, 1, . . . , w q i, t, . . . , w q i, m ) is a query, r i = (w r i, 1, . . . , w r i, t, . | d8ef6b035bf305aa648a1d10ed0c43eb466ead09a3d417f7d5cd42a6a349b878 |
D17 | D17-1065.json | . . , w r i, n ) stands for the response to q i, and w q i, t and w r i, t denote the tth words in q i and r i, respectively. This paper aims to learn a generative model G(r|q) based on a discriminator D that can predict informative responses with good diversity for arbitrary input queries. | 5f876bb1caab7fa5dc7a4d25753fce14e9d87a3f749245fc04ae0d9542b3204d |
D17 | D17-1065.json | … … GRU Real Fake Generator Discriminator [ •! • " ⋯ • $ ⋯ • % ] [! " ⋯ $ ⋯ % ] & ' ℎ $ Fully-Connected | 76774dc51b3d84ed711f330ed13b27e7312b39b7bdc30c5307adb7d54699c18c |
D17 | D17-1065.json | We name the proposed model Generative Adversarial Network with an Approximate Embedding Layer (GAN-AEL), of which Figure 1 illustrates the overall framework. Generally speaking, the whole framework consists of a response generator G, a discriminator D and an embedding approximation layer that connects the G and the D. | a14546ba58c7c39d0a257cae378038f0758ff37cbc2ae94ec00b02ce3b9a04bb |
D17 | D17-1065.json | We explain each of the components in detail as follows. The generator adopts the Gated Recurrent Unit (GRU) (Cho et al. , 2014) based encoderdecoder architecture, where the encoder projects the input query (a discrete word sequence) into a real-valued vector, on which the output will be generated conditionally in the d... | 2666f6ea5659b914246feb52c6e012c7f68497a268dbf0557080af57596f7bcd |
D17 | D17-1065.json | The proposed GAN framework possesses sev-eral advantages over existing conversational response generation models. Firstly, both the generator and the discriminator are conditioned on the input query, which guarantees the relevance of the generated responses. Secondly, the discriminator enforces the generator to produce... | 39d42e2889c84da6b914a53520d93adf8a2996c401b2de5dbe2c78fa13b77ce5 |
D17 | D17-1065.json | In our proposed encoder-decoder framework, both the encoder and the generator (i. e. the decoder) G is composed of GRU (Cho et al. , 2014) units, which is designed to generate responses | 2e0e7d18ad0e6449c7abdecacfa87b220f4e550cb0bcd4066382bbd3c29da98b |
D17 | D17-1065.json | r = {w r, 1, w r, 2, • • •, w r, K } conditioned on an input query q = {w q, 1, w q, 2, • • •, w q, J }. | 3f312487cf206d5d5a5edb9501024e294475ed430fa328ea11087640b7162bb6 |
D17 | D17-1065.json | For a given query-response pair (q, r), the target is to maximum the conditional probability p(r|q) in the generative process. Concretely, in this model, q is firstly encoded into a vector representation q v by the GRU-based encoder as shown in Figure 1, which is actually the last hidden state of the encoder. Then the ... | 728faf26f5221495fa9868acd9266e481ee8f2fb355c76e560fcb76015e88108 |
D17 | D17-1065.json | Taking the logarithm of the probabilities for effective computation, the generator is trained by optimising the Maximum Likelihood Estimation (MLE) objective defined as: | 6da54c4d30405887383330e2fc2d4eae726bb5f4a0aeac4eeb73979b021f79f4 |
D17 | D17-1065.json | (2) Note here, we need to pre-train the generator using Equation 2 as the loss function to guarantee the generator to produce grammatical utterances. Otherwise, the discriminator will tend to learn a rule with ease to distinguish human-produced utterances from those ungrammatical responses generated in the early stages... | 5ac000df6d2a04a379f563578756a245af2e37d459e4604c667790ec93681e2f |
D17 | D17-1065.json | In order to smoothly connect the output layer of the generator to the input layer of the discriminator to yield an end-to-end differentiable GAN, one needs to solve the following critical problem. The output of the generator is a sequence of discrete words, which is usually sampled from the distributions predicted by t... | db5947ac611b0d6a1c294d3ce2e62d6e124d69f8851cc4c28f8166200e513140 |
D17 | D17-1065.json | Since afterward those words will be projected into embedding vectors to feed the CNN-based discriminator, we introduce an embedding approximation layer to merge the generation process of the decoder and the word embedding phrase of the discriminator. This can be done by multiplying the word probabilities in the distrib... | dc4e880553e1f6dcebbb89d248f024bf351d28dcbde70cbe3b88d2089047f660 |
D17 | D17-1065.json | The structure of the approximation layer is illustrated on the right-hand side of Figure 1. Concretely, the approximation layer takes the output h i of the generator and a random noise z i as the input, and reuses the word projection layer (pre-trained in the standard generator) to estimate the probability distribution... | 5691f3a684147eba1f6aec1198aaac1ce0f1a52bbe382374bd180a0d560ed51c |
D17 | D17-1065.json | where W p and b p are the weight and bias parameters of the word projection layer, respectively, and h i is the hidden representation of word w i, from the decoding procedure of the generator G, which is computed as: | 457b0daebaf3f29d3cd6ad1e30151a4e4d71215ae050541b7f04ff68681478dd |
D17 | D17-1065.json | CNN has been proven to be an appropriate classifier for many NLP tasks, such as sentence classification (Kim, 2014) and matching (Hu et al. , 2014). Therefore, in this paper we adopt a CNNbased discriminator as shown in Figure 1. For the convenience of further discussions, we introducer to denote the underlying (distri... | 37af8ad8e37072be28c35d66c8394e02997bae59aa7ebf10fce3f7e6fc04f74a |
D17 | D17-1065.json | Finally, we concatenate A q to A r and Ar separately, and feed the resulting vectors to their respective fully-connected layers, as illustrated in Figure 1. Here, we make the two fully-connected layers share common parameters and predict probabilities D(r|q) and D(r|q), respectively, for r andr being true responses of ... | dd0b953a83095cbbc000102bfbe67c10446706d0f31b2e524a0cc2600b277740 |
D17 | D17-1065.json | In practice, when the Seq2Seq generative network G is pre-trained, we also pre-train the above discriminator D by maximising the following objective function: | 2bce8c64d20a8ee603528a3b6e8ccf1c31f9127834b84814ec22231d2a459210 |
D17 | D17-1065.json | with the parameters of G frozen, before the adversarial training procedure described in Section 3. 6. | 79615fdf89e45edbd8dba95d2d018489746547bf64e012409e0a23ca9afb1f8a |
D17 | D17-1065.json | After the pre-training of the generator G and the discriminator D as described above, the entire network is trained adversarially. Concretely, we iteratively train G and D, where at each iteration, the parameters of the non-training network will be frozen. The following tricks are utilised in the adversarial training p... | 7d4e276f435646181c5b437de8f89920015445bcaef2bb45614ae3526f076cfd |
D17 | D17-1065.json | where θ G denotes the active parameters of the generator G, G loss = A r − Ar and g D, G (•) stands for the inference step of the entire GAN. It can be seen that the feedback signals from D can be propagated to G effectively through the approximate embedding layer, which connects G and D smoothly, and avoids the discre... | 8824f8d1826cc96d0e840c402cfabee59c0e3204863438d9fa008f384c8982cf |
D17 | D17-1065.json | We test our model on two datasets: Baidu Tieba and OpenSubtitles (Lison and Tiedemann, 2016). The Baidu Tieba dataset is composed of single-turn conversations collected from the threads of Baidu Tieba 1, of which the utterance length ranging from 3 to 30 words. The Open-Subtitles dataset contains movie scripts organise... | 73e558a62675f91b705f95caf6ad75a7ef59e3a57d6ba821524f63b4156c6303 |
D17 | D17-1065.json | To illustrate the performance of the proposed model, we introduce three existing approaches as baselines. | b68092e685cf244899983a753a32c3a0fc1bdc5f5a959a3720c48c805edd12b6 |
D17 | D17-1065.json | • Seq2Seq: the standard sequence-to-sequence model (Sutskever et al. , 2014). | 03662964fc9c4aa878c4295ce3cf2edcfdc3e676cd038e6f66066683b38dc9ee |
D17 | D17-1065.json | • MMI-anti: a Seq2Seq model with a Maximum Mutual Information (MMI) criterion (implemented as an anti-language model) (Li et al. , 2016a) in the decoding process, which reduces the probability of generating "safe responses". | 5dff258b91586e2a63ed980bdec2becfbf8841b1d0b10f5333544a0dd25b5365 |
D17 | D17-1065.json | • Adver-REGS: another adversarial strategy proposed by Li et al. (2017) 2, which links the generator and the discriminator together with a reinforcement learning framework, and takes the discriminator's output probability as the reward to train the generator. | 8d38147e8540a46e306d5a4e4e8b2bd88ce84dffafccd46c79d78bd3add6bab5 |
D17 | D17-1065.json | For automatic evaluations, the following commonly accepted metrics are employed. Note here, the goal of our model is to obtain responses not only semantically relevant to the corresponding queries, but also of good diversity and novelty. Therefore, in this work, embedding-based metrics are adopted to evaluate semantic ... | eace696764e0ad167c4daa070e5daf56cc840a140d961cddd7ca9ddea9c2cb02 |
D17 | D17-1065.json | Relevance Metrics: The following three word embedding based metrics 3 are used to compute the semantic relevance of two utterances. The Greedy metric is to greedily match words in two given utterances based on the cosine similarities of their embeddings, and to average the obtained scores (Rus and Lintean, 2012). Alter... | c9699e5168de34e9ce9666d78d425681173417d5872e85f02360822113643ba1 |
D17 | D17-1065.json | Diversity Metrics: To measure the informativeness and diversity of the generated responses, we follow the dist-1 and dist-2 metrics proposed by Li et al. (2016a) and, and introduce a Novelty metric. The dist-1 (dist-2) is defined as the number of unique unigrams (bigrams for dist-2). A common drawback of dist-1 and dis... | 4cc69229e4d83d42bcdc4dce354d515caa8c03daea6074206b652a80cbbcd549 |
D17 | D17-1065.json | Human Evaluation: To evaluate the performance of our model from human perspectives, this paper conducts a human subject experiement by comparing the responses generated by Adver-REGS (which is one of the most competitive existing approaches) with those by the proposed model. Three experienced annotators are invited to ... | 185ac514421ba549f7f4c42ffa8ccacfdbe6608af4b0318782e8e62889da6513 |
D17 | D17-1065.json | Considering that all the annotators use Mandarin as their first language, the above evaluation is only done on the Tieba dataset. | 49d9b21bcf604ae1e23a363bd2456c47cd2de29a70d587acf8e19944bde13008 |
D17 | D17-1065.json | Hyperparameter Settings: The hyperparameters of the networks used in all the experiments below are described as follows. The vocabulary sizes for Tieba and OpenSubtitles are truncated to 100, 000 and 150, 000, respectively. The dimensions of word embedding vectors are set to 100 for Tieba and 300 for OpenSubtitles. The... | ebb5e2380fa1549757765d98ce3685ccc53ea08d5339c81de58b819bf77556ed |
D17 | D17-1065.json | Training Strategies: To train the proposed GAN, the parameters of the generator G are initialised based on the pre-training mentioned in subsection 3. 3, while those of the discriminator D are randomly initialised. starts from pre-training D with the parameters of G fixed. After this, G and D will be trained iterativel... | 55be5ef1137140181bf6cbd82bdf0a60aa8b86743a753a4419f402668444a97c |
D17 | D17-1065.json | From Table 1 and 2, it can be observed that the proposed GAN-AEL model outperforms the baselines on both datasets in all metrics, especially for the diversity oriented scores. The improvements can be explained from the following two angles. a) Since a vanilla Seq2Seq model does not take diversity, novelty or informativ... | 9ebc64fb3d926216ac66fe43b8b1996f63d37a6b38d5239f7e27605531811419 |
D17 | D17-1065.json | b) The proposed approximation layer is an effective way to couple the response generator and the discriminator. Through this differentiable component, the loss of the discriminator is properly propagated to the generator and guide the tuning of the latter's parameters. | fe7d79f698b4fc939bb7b976932002703832b07d4969b79a3230a9142cdd8994 |
D17 | D17-1065.json | It can also be seen from the results that the performance of all the models on the three semantic relevance oriented metrics are comparable to each other. This implies that all the models, including the baseline methods and the proposed model, have the capability to generate responses of reasonable relevance to given q... | e1cf99db7b4ad7f45f914824f42001138669e089317726fc8e89a3315fa7ec8c |
D17 | D17-1065.json | Furthermore, when compared to Adver-REGS, the proposed GAN-AEL gains 30%-60% relative improvement in the dist-1, dist-2 and novelty metrics on both datasets, which indicates that coupling the generator and the discriminator with a differentiable component is a more preferable methodology for text generation tasks, and ... | 6bbc0c1080a25bf21a94ac3654d4b599e4041b4bcbf16137ea699f05fb100cf5 |
D17 | D17-1065.json | In addition, it can be seen that GAN-AEL improves the greedy score to a much greater extent than the average and extreme scores, which further suggests that the responses generated by GAN-AEL are more informative. Concretely, the calculations of the average and extreme scores may be dominated by generic non-informative... | 9435485d311994fa8817c9f3f3494bf87ef0d0f06288a2dade004823c41f0bb0 |
D17 | D17-1065.json | Ties 0. 61 0. 13 0. 26 The discriminator plays an important role in the adversarial training process, which determines whether the GAN model converges to a Nash Equilibrium (Chen et al. , 2016). We conduct a set of experiments to explore the influence of the discriminator's capacities to the adversarial training. Figur... | cfc5e36b2886d454e3804094ecc738798f9e061921024a1caf06b8ee7395c4c3 |
D17 | D17-1065.json | It can be found that the discriminator with "Filter[1, 2]" achieves the best performance. Two facts based on the principle of GAN could be taken to explain this observation: On one hand, a discriminator with too low capacity (such as "Filter[1]") is less capable in distinguishing human responses from generated ones, wh... | 31d1e2a8949c303ce95d81398200f23941f6caaecc2a27fea0d50169c652d8ca |
D17 | D17-1065.json | To demonstrate the intuitive performance of the proposed model in comparison with the naïve Seq2Seq model, we provide some example cases in Figure 3, where for each query the response is the top hypothesis obtained via beam search. Especially, we show that when the vanilla approach generates safe responses such as "I d... | 9deb5526202f048f87b0730f1d95e3e6f2fa5a109928365c69c25648ca42bf27 |
D17 | D17-1065.json | In this paper, we proposed a GAN framework to model single-turn short-text conversations. An approximation embedding layer is introduced to force the entire network differentiable, which significantly overcomes the drawbacks found in | b5eb218e1054d2b72c4d6f5d82989eeac374073f73e65017bcc007dd7247cad7 |
D17 | D17-1065.json | https: //tieba. baidu. com/index. html 2 The codes can be accessed at https: //github. com/jiweil/Neural-Dialogue-Generation/ tree/master/Adversarial | 52997a4fb3f5192e71580397fd0ab13e0a65c1317d8861c6bb6c548dfb0ae390 |
D19 | D19-1360.json | Code-switching is a phenomenon that often happens between multilingual speakers, in which they switch between their two languages in conversations, hence, it is practically useful to recognize this well in spoken language systems (Winata et al. , 2018a). This occurs more often for entities such as organizations or prod... | 4de66f92e62c5eec3050b99caa93053384b99b968173d358ee89cd05dd9ee877 |
D19 | D19-1360.json | • (translation) walking dead (a movie title) takes away the appetite of anyone For this task, previous works have mostly focused on applying pre-trained word embeddings from each language in order to represent noisy mixed-language texts, and combine them with character-level representations (Trivedi et al. , 2018; Wina... | 5bfeb578c3c6bd41d4a589c8cc2c93c3e315f00fd3aea444327bfddc0fb7f983 |
D19 | D19-1360.json | Despite such expected usefulness, there has not been much attention focused around using subword-level features in this task. This is partly because of the non-trivial difficulty of combining different language embeddings in the subword space, which arises from the distinct segmentation into subwords for different lang... | 14fe13c16ba6419c4f60925cd287090f9449c3e691417911c457b43961d2fa2f |
D19 | D19-1360.json | In this paper, we propose Hierarchical Meta-Embeddings (HME) 1 which learns how to combine different pre-trained monolingual embeddings in word, subword, and character-level into a single language-agnostic lexical representation without using specific language identifiers. To address the issue of different segmentation... | aa85be20c7e05ff81c3154488892b1152f6daa592ed34cf2516189a0162fec39 |
D19 | D19-1360.json | Embeddings Previous works have extensively explored different representations such as word (Mikolov et al. , 2013; Pennington et al. , 2014; Grave et al. , 2018; Xu et al. , 2018), subword (Sennrich et al. , 2016; Heinzerling and Strube, 2018), and character (dos Santos and Zadrozny, 2014; Wieting et al. , 2016). Lampl... | f7633ae687083a1afbd074fb66de1600103af805b11e11f214f153591181ef70 |
D19 | D19-1360.json | Meta-embeddings Recently, there are studies on combining multiple word embeddings in pre-processing steps (Yin and Schütze, 2016; Muromägi et al. , 2017; Coates and Bollegala, 2018). Later, introduced a method to dynamically learn word-level meta-embeddings, which can be effectively used in a supervised setting. Winata... | 0ee3d0a608b3d4b06be82a24805fbc46800056cd3d07fc0d583708eccd728a75 |
D19 | D19-1360.json | We propose a method to combine word, subword, and character representations to create a mixture of embeddings. We generate a multilingual metaembeddings of word and subword, and then, we concatenate them with character-level embeddings to generate final word representations, as shown in Figure 1. Let w be a sequence of... | aa6474681b322e90f8d9b1e63333e3ed093a9697db2236fbd064d7ca8e5812d3 |
D19 | D19-1360.json | We generate a meta-representations by taking the vector representation from multiple monolingual pre-trained embeddings in different subunits such as word and subword. We apply a projection matrix W j to transform the dimensions from the original space x i, j ∈ R d to a new shared space x i, j ∈ R d. Then, we calculate... | db4cce13596fc63c9b239b8c96a43905b87f4859591cdf7d2e02bbb2a7f43c09 |
D19 | D19-1360.json | We propose to map subword into word representations and choose byte-pair encodings (BPEs) (Sennrich et al. , 2016 ) since it has a compact vocabulary. First, we apply f to segment words into sets of subwords, and then we extract the pre-trained subword embedding vectors x | e65e3aca7d6eb3ddf88b9679953cd65e18fe69e4aedd09d0055e34cd9db968d8 |
D19 | D19-1360.json | Since, each language has a different f, we replace the projection matrix with Transformer (Vaswani et al. , 2017) to learn and combine important subwords into a single vector representation. Then, we create u | 0b015bcebefa9014e6442822b0f2dd6943dabfce55e4ac603d2a5a8afab59ddc |
D19 | D19-1360.json | To combine character-level representations, we apply an encoder to each character. | de2a07427c3a6997f99ade2a929173d71c1dfe1cb8153441c53d94fb5130378b |
D19 | D19-1360.json | We combine the word-level, subword-level, and character-level representations by concatenation is a character embedding. We randomly initialize the character embedding and keep it trainable. We fix all subword and word pre-trained embeddings during the training. | f5fb56cf291898a104660d6adf02887d87716c9b66d17c6e9b22146aaf89a75e |
D19 | D19-1360.json | u HM E i = (u (w) i, u (s) i, u (c) i ), where u (w) i ∈ R d and u (s) i ∈ R d | d4b120edad1b871da0d9c85db51d2537107509ad04d29061be84d685ff329601 |
D19 | D19-1360.json | To predict the entities, we use Transformer-CRF, a transformer-based encoder followed by a Conditional Random Field (CRF) layer (Lafferty et al. , 2001). The CRF layer is useful to constraint the dependencies between labels. | 48b5f8a2dfa3341190c56058298ac7cbf31a54db531fa8a353fd637decc04cca |
D19 | D19-1360.json | The best output sequence is selected by a forward propagation using the Viterbi algorithm. | 1061780b541d530669c15989f502773cf123df6a1af0c2aaf7a482d443a3f5f5 |
D19 | D19-1360.json | We train our model for solving Named Entity Recognition on English-Spanish code-switching tweets data from Aguilar et al. (2018). There are nine entity labels with IOB format. The training, development, and testing sets contain 50, 757, 832, and 15, 634 tweets, respectively. We use FastText word embeddings trained from... | c0090876fd1f84259a855813d823cf3d629649a186709bfbd883db30e6f8804e |
D19 | D19-1360.json | CONCAT We concatenate word embeddings by merging the dimensions of word representations. This method combines embeddings into a highdimensional input that may cause inefficient com- putation. | 58c8484303436eea4331f62406469cf49c54442605c203c19e3db995b5293cc3 |
D19 | D19-1360.json | LINEAR We sum all word embeddings into a single word vector with equal weight. This method combines embeddings without considering the importance of each of them. | 7c12e77fe452735a6b121c760dfa8c535a4f471c7598b23c9e87ad12d91ab509 |
D19 | D19-1360.json | Random Embeddings We use randomly initialized word embeddings and keep it trainable to calculate the lower-bound performance. | eb5fc74af60d5d4e320cc4767ea11ee0b89091c56900a07197c15cbf3dffb762 |
D19 | D19-1360.json | Aligned Embeddings We align English and Spanish FastText embeddings using CSLS with two scenarios. We set English (en) as the source language and Spanish (es) (en → es) as the target language, and vice versa (es → en). We run MUSE by using the code prepared by the authors of Conneau et al. (2017). 2 | 9865a0349ae0ce8d936788a5030d1e96e9c4bda373043518805994672c0e7a0c |
D19 | D19-1360.json | In general, from Table 1, we can see that wordlevel meta-embeddings even without subword or character-level information, consistently perform better than flat baselines (e. g. , CONCAT and LIN-EAR) in all settings. This is mainly because of the attention layer which does not require additional parameters. Furthermore, ... | 20f5d64702dc64100430dde187fe2afb9fe7a3dbbec7d373c81af8b58b5e6b8b |
D19 | D19-1360.json | On the other hand, adding subword inputs to the model is consistently better than characters. This is due to the transfer of the information from the pre-trained subword embeddings. As shown in Ta- Moreover, we visualize the attention weights of the model in word and subword-level to interpret the model dynamics. From ... | 0e655508f69bb913c324333e340d7997c44119df83c26824ad15e1a6592c4ac1 |
D19 | D19-1360.json | We propose Hierarchical Meta-Embeddings (HME) that learns how to combine multiple monolingual word-level and subword-level embeddings to create language-agnostic representations without specific language information. We achieve the state-of-the-art results on the task of Named Entity Recognition for English-Spanish cod... | 1106172198f129e11b360445e3cc77fd8bbe4c04e51c34e18fc176e059d732d9 |
eamt | 2012.eamt-1.66.json | Lexicalized reordering models are a common component of standard phrase-based machine translation systems. | 37e85e03922a30e98fcb00cd4d09827105abbaacf29c17908a2da25e8dd0dc40 |
eamt | 2012.eamt-1.66.json | In hierarchical phrase-based machine translation, reordering is modeled implicitely as part of the translation model. Hierarchical phrase-based decoders conduct phrase reorderings based on a one-to-one relation between the non-terminals on source and target side within hierarchical translation rules. Non-terminals on s... | 9fa4489573e50e5d8e4dff81a8f1b7c509de114bba2e411b867c2ea8ceb66004 |
eamt | 2012.eamt-1.66.json | The best translation quality is achieved with combinations of the extensions with additional reordering rules and with the discriminative reordering model. The overall improvement over the respective baseline system is +1. 2 %BLEU / -0. 6 %TER absolute in the standard setup and +1. 2 %BLEU / -0. 5 %TER absolute in the ... | 8fbbf29bdd3469ddd7df79be8ef5fc7d64bf726d6d2122440d2ce271b8395513 |
eamt | 2012.eamt-1.66.json | Hierarchical phrase-based translation was proposed by Chiang (2005). Iglesias et al. (2009) and in a later journal publication Gispert et al. (2010) present a way to limit the recursion depth for hierarchical rules by means of a modification to the hierarchical grammar. Their work is of interest to us as a limitation o... | 6b05202f5daa0f59df0013acb98530f78894c647fbea56ce2416d4b1dfe60001 |
eamt | 2012.eamt-1.66.json | Gao et al. (2011). He et al. (2010a) combine an additional BTGstyle swap rule with a maximum entropy based lexicalized reordering model and achieve improvements on the Chinese→English NIST task. Their approach is comparable to ours, but their reordering model requires the training of different classifiers for different... | 3dfa64d2524b1157507d6ea922fb79f6a4dd331dfe07d69d952ede2d4694b494 |
eamt | 2012.eamt-1.66.json | In standard phrase-based systems, lexicalized reordering models are a commonly included component. A widely used variant is the orientation model as implemented in the Moses toolkit (Tillmann, 2004; Koehn et al. , 2007) which distinguishes monotone, swap, and discontinuous phrase orientations. Galley and Manning (2008)... | df8e6758c19786a75f5b6349c295229af0234a826b3f760da24ec53a8470c398 |
eamt | 2012.eamt-1.66.json | 3 Shallow-1 Grammar Gispert et al. (2010) propose a limitation of the recursion depth for hierarchical rules with shallow-n grammars. The main benefit of the limitation is a gain in decoding efficiency. Moreover, the modification of the grammar to a shallow version re-stricts the search space of the decoder and may be ... | d99c2006603696301a831525aac24e4c83e884820dd03c7fd3ab43b2b8d62f3e |
eamt | 2012.eamt-1.66.json | In a shallow-1 grammar, the generic nonterminal X of the standard hierarchical approach is replaced by two distinct non-terminals XH and XP. By changing the left-hand sides of the rules, lexical phrases are allowed to be derived from XP only, hierarchical phrases from XH only. On all right-hand sides of hierarchical ru... | 2a6dcc859b794e5f7cbe59d61bf48c539860e3417356ad2498d0da6cf10473e6 |
eamt | 2012.eamt-1.66.json | In this section we describe three types of reordering extensions to the hierarchical grammar. All of them add specific non-lexicalized reordering rules which facilitate a more flexible arrangement of phrases in the hypotheses. We first present a simple swap rule extension (Section 4. 1), then we suggest two different e... | 84569a8faf349dc61b4e2c4e0c4975385da5f37628fdc37bd52108b42d359e95 |
eamt | 2012.eamt-1.66.json | In a deep grammar, we can bring in more reordering capabilities by adding a single swap rule | 59c16cf1c8c39bf7256c828fa48389ea0d6009f25623c6533b8d2e8898ab4226 |
eamt | 2012.eamt-1.66.json | supplementary to the standard initial rule and glue rule. The swap rule allows adjacent phrases to be transposed. | 4dc430e72611b7754a301f3bdbd2b597a185abf08bb8b617e8087e43046dba09 |
eamt | 2012.eamt-1.66.json | An alternative with a comparable effect would be to remove the standard glue rule and to add two rules instead, one of them being as in Equation (3) and the other a monotonic concatenation rule for the non-terminal X which is symmetric to the swap rule. The latter rule acts as a replacement for the glue rule. This is t... | b6416cc7b799616005c115931b02a6798a4a355353dbf717202650906105772e |
eamt | 2012.eamt-1.66.json | In a shallow grammar, several directions of integrating swaps are possible. We decided to add a swap rule and a monotonic concatenation rule | 7c5f30788a45dbd2eeef94b74029aeb4662ea313f890abd655d7f1e584d8f7d4 |
eamt | 2012.eamt-1.66.json | supplementary to the standard shallow initial rules and glue rules. The swap rule allows adjacent lexical phrases to be transposed, but not hierarchical phrases. Here, we could as well have used XH as the left-hand side of the rules. As we chose XP and thus allow for embedding of subderivations resulting from applicati... | a49edc6e4aba550e143eeec54b96c9fe2908032c36011e46cc2789d8b5edaaf6 |
eamt | 2012.eamt-1.66.json | Instead of employing a swap rule that transposes adjacent phrases, we can adopt more complex extensions to the grammar that implement jumps across blocks of symbols. Our first jump rules variant is inspired by Vilar et al. (2010), but is a generalization that facilitates an arbitrary number of blocks per sentence to be... | 98e1f5463d749afb6278d9d5e8e51b8364f62bc7ac43acce4bd900ee185953d1 |
eamt | 2012.eamt-1.66.json | S → B ∼0 X ∼1, X ∼1 B ∼0 †S → S ∼0 B ∼1 X ∼2, S ∼0 X ∼2 B ∼1 †B → X ∼0, X ∼0 B → B ∼0 X ∼1, B ∼0 X ∼1 ‡ (5) | b0be553dd75e24b52c4e14d0b28fb897da37f8f19336da53d06f5c90c658d642 |
eamt | 2012.eamt-1.66.json | in addition to the standard initial rule and glue rule. The rules marked with †are jump rules that put jumps across blocks (B ) on source side into effect. The rules with B on their left-hand side enable blocks that are skipped by the jump rules to be translated, but without further jumps. Reordering within these wind... | 6d1d6d7169a371161b1e6a91a4a36fa0865adb0c1f73d56934fa073a9fc5f9cb |
eamt | 2012.eamt-1.66.json | A binary jump feature for the two jump rules ( †) may be added to the log-linear model combination of the decoder, as well as a binary feature that fires for the rule that acts analogous to the glue rule, but within blocks that is being jumped across ( ‡ ). A maximum jump width can be established by applying a length ... | 25a783d7991797cc9c110683b32c21e5e6909b420882fbfc003669f6ccfdb6da |
eamt | 2012.eamt-1.66.json | In a shallow grammar, block jumps are realized in the same way as in a deep one, but the number of rules that are required is doubled. | d84b4c195c2ef85d820ed4d126da82ccadc1a4d63507c8e1d5bd60d3e2f145d1 |
eamt | 2012.eamt-1.66.json | S → B ∼0 XP ∼1, XP ∼1 B ∼0 †S → B ∼0 XH ∼1, XH ∼1 B ∼0 †S → S ∼0 B ∼1 XP ∼2, S ∼0 XP ∼2 B ∼1 †S → S ∼0 B ∼1 XH ∼2, S ∼0 XH ∼2 B ∼1 †B → XP ∼0, XP ∼0 B → XH ∼0, XH ∼0 B → B ∼0 XP ∼1, B ∼0 XP ∼1 ‡ B → B ∼0 XH ∼1, B ∼0 XH ∼1 ‡ (6) | 848c5b8769aab5dd9b61f4a186dcae01781a055c73a81aa7ca1c6b987d5a23df |
eamt | 2012.eamt-1.66.json | As a second jump rules variant, we try an approach that follows (Huck et al. , 2011) and allows for very constrained reorderings. At most one contiguous block per sentence can be jumped across in this variant. | 7766335742f4c8836a8897a360fa94a952627742d4006a0b4513fa6334717c5e |
eamt | 2012.eamt-1.66.json | In a deep grammar, to enable constrained block jumps with at most one jump per sentence, we replace the initial and glue rule by the rules given in Equation 7: | 8a642f2d2eababadc6b3f1382b9bc23f0c6e8cda5ae3bb0433b5ca8d76ef651d |
eamt | 2012.eamt-1.66.json | S → M ∼0, M ∼0 S → S ∼0 M ∼1, S ∼0 M ∼1 ‡ S → B ∼0 M ∼1, M ∼1 B ∼0 †M → X ∼0, X ∼0 M → M ∼0 X ∼1, M ∼0 X ∼1 ‡ B → X ∼0, X ∼0 B → B ∼0 X ∼1, B ∼0 X ∼1 ‡ (7) | 2cff83045a49c04c366239103b43db59cecfa10e9dda1139eead01344b0ff85c |
eamt | 2012.eamt-1.66.json | In these rules, the M non-terminal represents a block that will be translated in a monotonic way, and the B is a "back jump". We omit the exposition for shallow grammars as deducing the shallow from the deep version of the rules is straightforward from our previous explanations. | 62c2bf0cc601f6e548569937f7d759b0b7b000223f946c5ed14bb9bd88eaf2b1 |
eamt | 2012.eamt-1.66.json | We add a binary feature that fires for the rules that act analogous to the glue rule ( ‡ ). We further conform to the approach of Huck et al. (2011) by additionally including a distance-based distortion model (dist. feature) that is computed during decoding whenever the back jump rule ( †) is applied. | ffd08d272a0716fdb5d1b93d290ad6d5e885d147d811da9e9e86ddf2c2136a6b |
End of preview. Expand in Data Studio
No dataset card yet
- Downloads last month
- 8