paper stringlengths 0 839 | paper_id stringlengths 1 12 | table_caption stringlengths 3 2.35k | table_column_names large_stringlengths 13 1.76k | table_content_values large_stringlengths 2 11.9k | text large_stringlengths 69 2.82k |
|---|---|---|---|---|---|
Neural Aspect and Opinion Term Extraction with Mined Rules as Weak Supervision | 1907.03750 | Table 2: Aspect and opinion term extraction performance of different approaches. F1 score is reported. IHS_RD, DLIREC, Elixa and WDEmb* use manually designed features. For different versions of RINANTE, “Shared” and “Double” means shared BiLSTM model and double BiLSTM model, respectively; “Alt” and “Pre” means the firs... | ['Approach', 'SE14-R Aspect', 'SE14-R Opinion', 'SE14-L Aspect', 'SE14-L Opinion', 'SE15-R Aspect', 'SE15-R Opinion'] | [['DP Qiu et\xa0al. ( 2011 )', '38.72', '65.94', '19.19', '55.29', '27.32', '46.31'], ['IHS_RD Chernyshevich ( 2014 )', '79.62', '-', '74.55', '-', '-', '-'], ['DLIREC Toh and Wang ( 2014 )', '84.01', '-', '73.78', '-', '-', '-'], ['Elixa Vicente et\xa0al. ( 2017 )', '-', '-', '-', '-', '[BOLD] 70.04', '-'], ['WDEmb Yi... | From the results, we can see that the mined rules alone do not perform well. However, by learning from the data automatically labeled by these rules, all four versions of RINANTE achieves better performances than RINANTE (no rule). This verifies that we can indeed use the results of the mined rules to improve the perfo... |
Exploring Models and Data for Image Question Answering | 1505.02074 | Table 1: COCO-QA question type break-down | ['Category Object', 'Train 54992', '% 69.84%', 'Test 27206', '% 69.85%'] | [['Number', '5885', '7.47%', '2755', '7.07%'], ['Color', '13059', '16.59%', '6509', '16.71%'], ['Location', '4800', '6.10%', '2478', '6.36%'], ['Total', '78736', '100.00%', '38948', '100.00%']] | It should be noted that since we applied the QA pair rejection process, mode-guessing performs very poorly on COCO-QA. However, COCO-QA questions are actually easier to answer than DAQUAR from a human point of view. This encourages the model to exploit salient object relations instead of exhaustively searching all poss... |
DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter | 1910.01108 | Table 1: DistilBERT retains 97% of BERT performance. Comparison on the dev sets of the GLUE benchmark. ELMo results as reported by the authors. BERT and DistilBERT results are the medians of 5 runs with different seeds. | ['Model', '[BOLD] Score', 'CoLA', 'MNLI', 'MRPC', 'QNLI', 'QQP', 'RTE', 'SST-2', 'STS-B', 'WNLI'] | [['ELMo', '68.7', '44.1', '68.6', '76.6', '71.1', '86.2', '53.4', '91.5', '70.4', '56.3'], ['BERT-base', '79.5', '56.3', '86.7', '88.6', '91.8', '89.6', '69.3', '92.7', '89.0', '53.5'], ['DistilBERT', '77.0', '51.3', '82.2', '87.5', '89.2', '88.5', '59.9', '91.3', '86.9', '56.3']] | Among the 9 tasks, DistilBERT is always on par or improving over the ELMo baseline (up to 19 points of accuracy on STS-B). DistilBERT also compares surprisingly well to BERT, retaining 97% of the performance with 40% fewer parameters. |
DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter | 1910.01108 | Table 2: DistilBERT yields to comparable performance on downstream tasks. Comparison on downstream tasks: IMDb (test accuracy) and SQuAD 1.1 (EM/F1 on dev set). D: with a second step of distillation during fine-tuning. | ['Model', 'IMDb', 'SQuAD'] | [['[EMPTY]', '(acc.)', '(EM/F1)'], ['BERT-base', '93.46', '81.2/88.5'], ['DistilBERT', '92.82', '77.7/85.8'], ['DistilBERT (D)', '-', '79.1/86.9']] | On SQuAD, DistilBERT is within 3.9 points of the full BERT. the number of parameters of each model along with the inference time needed to do a full pass on the STS-B development set on CPU (Intel Xeon E5-2690 v3 Haswell @2.9GHz) using a batch size of 1. DistilBERT has 40% fewer parameters than BERT and is 60% faster t... |
DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter | 1910.01108 | Table 2: DistilBERT yields to comparable performance on downstream tasks. Comparison on downstream tasks: IMDb (test accuracy) and SQuAD 1.1 (EM/F1 on dev set). D: with a second step of distillation during fine-tuning. | ['Model', '# param.', 'Inf. time'] | [['[EMPTY]', '(Millions)', '(seconds)'], ['ELMo', '180', '895'], ['BERT-base', '110', '668'], ['DistilBERT', '66', '410']] | On SQuAD, DistilBERT is within 3.9 points of the full BERT. the number of parameters of each model along with the inference time needed to do a full pass on the STS-B development set on CPU (Intel Xeon E5-2690 v3 Haswell @2.9GHz) using a batch size of 1. DistilBERT has 40% fewer parameters than BERT and is 60% faster t... |
Updating Pre-trained Word Vectors and Text Classifiers using Monolingual Alignment | 1910.06241 | Table 3: Comparison between the accuracy of our aligned model (RCSLS+Fine.), the accuracy of models trained on each split separately (Train on …) and two baselines. We observe that a model obtained using our procedure matches the performance of the Vote baseline while not requiring to store two separate models. | ['[EMPTY]', 'Sogou 3k', 'Sogou 30k', 'Sogou Full', 'Amazon 3k', 'Amazon 30k', 'Amazon Full', 'Yelp 3k', 'Yelp 30k', 'Yelp Full'] | [['Train on [ITALIC] S0', '91.3', '94.5', '96.2', '46.1', '53.2', '59.5', '50.7', '58.7', '62.8'], ['Train on [ITALIC] S1', '91.4', '94.3', '96.1', '47.2', '53.8', '59.6', '50.2', '59.0', '62.8'], ['Train on [ITALIC] S0∪ [ITALIC] S1', '92.3', '95.1', '96.7', '48.6', '54.9', '60.2', '54.2', '60.1', '63.7'], ['Fine-tu... | We report the performance of a model trained on each of the two splits alone, the two baselines, and a topline obtained by training on S0∪S1. First of all, we observe that our approach reduces the gap between training on a single set with the topline. This effect is especially true on small versions of the datasets (3k... |
Updating Pre-trained Word Vectors and Text Classifiers using Monolingual Alignment | 1910.06241 | Table 1: Performance of updated word vectors on the word analogy task. We split the English word analogy datasets (Mikolov et al., 2013a) into Out of vocab and In vocab questions. Out of vocab questions are composed of words that are out of vocabulary for S0, hence the null accuracy. | ['[EMPTY]', 'Out of vocab', 'In vocab'] | [['Train on [ITALIC] S0', '00.0', '70.1'], ['Train on [ITALIC] S1', '66.9', '66.1'], ['Train on [ITALIC] S0∪ [ITALIC] S1', '68.6', '71.2'], ['Fine-tune', '67.7', '66.8'], ['Subwords', '37.0', '71.6'], ['RCSLS+Fine.', '[BOLD] 67.9', '[BOLD] 72.1']] | S4SS1SSS0Px4 Results. First, we see that fine-tuning X on S1 leads to decent performance on the Out of vocab questions, but saps the accuracy on In vocab questions. As mentionned before, learning vectors initialized with X on S1 may lead to a loss of important statistics learnt on S0: the total accuracy on In vocab que... |
Updating Pre-trained Word Vectors and Text Classifiers using Monolingual Alignment | 1910.06241 | Table 2: Classification accuracy on Yelp reviews from 2018. We compare models trained on 1.2M reviews from 2013-2014 (S0) and a those trained on a smaller sample of reviews from 2018 (S1). We vary the size of S1 from 10k to 500k samples, while keeping S0 of fixed size. | ['[EMPTY]', 'Size of [ITALIC] S1 10k', 'Size of [ITALIC] S1 30k', 'Size of [ITALIC] S1 100k', 'Size of [ITALIC] S1 500k'] | [['Train on [ITALIC] S0', '74.9', '74.9', '74.9', '74.9'], ['Train on [ITALIC] S1', '70.8', '72.7', '74.4', '76.2'], ['Train on [ITALIC] S0∪ [ITALIC] S1', '75.1', '75.1', '75.3', '76.2'], ['Fine-tune', '72.9', '73.8', '75.0', '[BOLD] 76.3'], ['RCSLS+Fine.', '[BOLD] 75.1', '[BOLD] 75.3', '[BOLD] 75.8', '[BOLD] 76.3']... | First of all, we observe that the performance of models trained on S0 and S1 strongly depend on the size of S1. When the two datasets are of the same size (500k), the best performing model is the one trained on S1, as in that case there is no train/test distribution discrepancy. However, when S1 is small (10k or 30k), ... |
A Novel Cascade Binary Tagging Framework for Relational Triple Extraction | 1909.03227 | Table 2: Results of different methods on NYT and WebNLG datasets. Our re-implementation is marked by *. | ['Method', 'NYT [ITALIC] Prec.', 'NYT [ITALIC] Rec.', 'NYT [ITALIC] F1', 'WebNLG [ITALIC] Prec.', 'WebNLG [ITALIC] Rec.', 'WebNLG [ITALIC] F1'] | [['NovelTagging \xa0zheng2017Joint', '62.4', '31.7', '42.0', '52.5', '19.3', '28.3'], ['CopyR [ITALIC] OneDecoder \xa0zeng2018Extracting', '59.4', '53.1', '56.0', '32.2', '28.9', '30.5'], ['CopyR [ITALIC] MultiDecoder \xa0zeng2018Extracting', '61.0', '56.6', '58.7', '37.7', '36.4', '37.1'], ['GraphRel1 [ITALIC] p \xa0f... | The CasRel model overwhelmingly outperforms all the baselines in terms of all three evaluation metrics and achieves encouraging 17.5% and 30.2% improvements in F1-score over the best state-of-the-art method zeng2019Learning on NYT and WebNLG datasets respectively. Even without taking advantage of the pre-trained BERT, ... |
A Novel Cascade Binary Tagging Framework for Relational Triple Extraction | 1909.03227 | Table 3: F1-score of extracting relational triples from sentences with different number (denoted as N) of triples. | ['Method', 'NYT [ITALIC] N=1', 'NYT [ITALIC] N=2', 'NYT [ITALIC] N=3', 'NYT [ITALIC] N=4', 'NYT [ITALIC] N≥5', 'WebNLG [ITALIC] N=1', 'WebNLG [ITALIC] N=2', 'WebNLG [ITALIC] N=3', 'WebNLG [ITALIC] N=4', 'WebNLG [ITALIC] N≥5'] | [['CopyR [ITALIC] OneDecoder', '66.6', '52.6', '49.7', '48.7', '20.3', '65.2', '33.0', '22.2', '14.2', '13.2'], ['CopyR [ITALIC] MultiDecoder', '67.1', '58.6', '52.0', '53.6', '30.0', '59.2', '42.5', '31.7', '24.2', '30.0'], ['GraphRel1 [ITALIC] p', '69.1', '59.5', '54.4', '53.9', '37.5', '63.8', '46.3', '34.7', '30.8'... | It can be seen that the performance of most baselines on Normal, EPO and SEO presents a decreasing trend, reflecting the increasing difficulty of extracting relational triples from sentences with different overlapping patterns. That is, among the three overlapping patterns, Normal class is the easiest pattern while EPO... |
A Novel Cascade Binary Tagging Framework for Relational Triple Extraction | 1909.03227 | Table 4: Results on relational triple elements. | ['Element', 'NYT [ITALIC] Prec.', 'NYT [ITALIC] Rec.', 'NYT [ITALIC] F1', 'WebNLG [ITALIC] Prec.', 'WebNLG [ITALIC] Rec.', 'WebNLG [ITALIC] F1'] | [['[ITALIC] E1', '94.6', '92.4', '93.5', '98.7', '92.8', '95.7'], ['[ITALIC] E2', '94.1', '93.0', '93.5', '97.7', '93.0', '95.3'], ['[ITALIC] R', '96.0', '93.8', '94.9', '96.6', '91.5', '94.0'], ['[ITALIC] (E1, R)', '93.6', '90.9', '92.2', '94.8', '90.3', '92.5'], ['[ITALIC] (R, E2)', '93.1', '91.3', '92.2', '95.4', '9... | For NYT, the performance on E1 and E2 are consistent with that on (E1, R) and (R, E2), demonstrating the effectiveness of our proposed framework in identifying both subject and object entity mentions. We also find that there is only a trivial gap between the F1-score on (E1, E2) and (E1, R, E2), but an obvious gap betw... |
Market Trend Prediction using Sentiment Analysis:Lessons Learned and Paths Forward | 1903.05440 | Table 8. Market trend prediction using main technical indicators — the baseline model. | ['Type', 'Method', '3-day ahead Acc', '3-day ahead [ITALIC] Fup1', '3-day ahead [ITALIC] Fdown1', '5-day ahead Acc', '5-day ahead [ITALIC] Fup1', '5-day ahead [ITALIC] Fdown1'] | [['DJIA', 'SVM', '0.616', '0.738', '0.282', '[BOLD] 0.700', '[BOLD] 0.754', '[BOLD] 0.615'], ['DJIA', 'LSTM', '0.559', '0.706', '0.120', '0.585', '0.728', '0.127'], ['AAPL', 'SVM', '0.577', '0.676', '0.391', '[BOLD] 0.685', '[BOLD] 0.723', '[BOLD] 0.634'], ['AAPL', 'LSTM', '0.547', '0.693', '0.138', '0.521', '0.641', '... | The F1 scores suggest that LSTM often favoured the positive class over the negative class and produced unbalanced results. The reason could be that the size of the dataset is relatively small: there are 670 data points in the analysed time period 2011-2015. Contrary to LSTM, SVM always yielded balanced and stable resul... |
Market Trend Prediction using Sentiment Analysis:Lessons Learned and Paths Forward | 1903.05440 | Table 2. Sentiment attitudes Granger-causality on the FT I dataset. | ['Stock', 'Model', 'Lag', 'Attitude', 'Price⇒'] | [['Stock', 'Model', 'Lag', '⇒Price', 'Attitude'], ['S&P 500', 'Standard', '1', '0.1929', '0.1105'], ['S&P 500', 'Standard', '2', '0.2611', '[BOLD] 0.0780'], ['S&P 500', 'Temporal', '1', '0.2689', '[BOLD] 0.0495'], ['S&P 500', 'Temporal', '2', '0.1692', '[BOLD] 0.0940'], ['APPL', 'Standard', '1', '0.7351', '0.4253'], ['... | In all the experiments, we failed to discover any sign that sentiment attitudes Granger-cause stock price changes, which would suggest that in general sentiment attitudes probably cannot be useful for the prediction of stock price movements. However, in many cases, we found that the opposite was true — stock price chan... |
Market Trend Prediction using Sentiment Analysis:Lessons Learned and Paths Forward | 1903.05440 | Table 9. Market trend prediction using FT news articles and RWNC headlines (2011-2015). | ['Type', 'Baseline Acc', 'Baseline [ITALIC] Fup1', 'Baseline [ITALIC] Fdown1', 'Financial Times Acc', 'Financial Times [ITALIC] Fup1', 'Financial Times [ITALIC] Fdown1', 'Reddit Headlines Acc', 'Reddit Headlines [ITALIC] Fup1', 'Reddit Headlines [ITALIC] Fdown1'] | [['DJIA', '0.700', '[BOLD] 0.754', '0.615', '[BOLD] 0.706', '0.752', '[BOLD] 0.639', '0.618', '0.716', '0.417'], ['AAPL', '[BOLD] 0.685', '[BOLD] 0.723', '[BOLD] 0.634', '0.652', '0.723', '0.531', '0.624', '0.700', '0.496'], ['JPM', '0.673', '0.733', '0.578', '[BOLD] 0.679', '[BOLD] 0.739', '[BOLD] 0.583', '0.615', '0.... | Sentiment attitudes and emotions were extracted from the FT news articles and the RWNC headlines in the time period from 2011 to 2015. This is consistent with the previous section, in which no correlation or causality link was established between headlines sentiments and stock prices. It might be explained by the fact ... |
Market Trend Prediction using Sentiment Analysis:Lessons Learned and Paths Forward | 1903.05440 | Table 10. Market trend prediction using financial tweets from Twitter (01/04/2014 – 01/04/2015). | ['Type', 'baseline Acc', 'baseline [ITALIC] Fup1', 'baseline [ITALIC] Fdown1', 'all+attitude+emotion Acc', 'all+attitude+emotion [ITALIC] Fup1', 'all+attitude+emotion [ITALIC] Fdown1', 'all+emotion Acc', 'all+emotion [ITALIC] Fup1', 'all+emotion [ITALIC] Fdown1', 'filtering+emotion Acc', 'filtering+emotion [ITAL... | [['DJIA', '[BOLD] 0.810', '[BOLD] 0.854', '0.727', '0.810', '0.846', '[BOLD] 0.750', '0.778', '0.829', '0.682', '-', '-', '-'], ['AAPL', '[BOLD] 0.889', '[BOLD] 0.918', '[BOLD] 0.829', '0.810', '0.860', '0.700', '0.794', '0.847', '0.683', '0.794', '0.831', '0.735'], ['JPM', '0.746', '0.800', '0.652', '0.730', '0.779', ... | Only for the JPM stock we could see noticeable performance improvements in the “filtering+emotion” setting. Our results have also confirmed that sentiment attitudes on their own are probably not very useful for market trend prediction, but at least for some particular stocks sentiment emotions could be exploited to imp... |
Pretrained Semantic Speech Embeddings for End-to-End Spoken Language Understanding via Cross-Modal Teacher-Student Learning | 2007.01836 | Table 5: Effect of objective function | ['Objective', 'Accuracy on Test, %', 'Accuracy on Test, %', 'Accuracy on Test, %', 'Valid value', 'Valid value', 'Valid value'] | [['function', 'SwBD', 'MRDA', 'FSC', 'Cosine', 'L2', 'L1'], ['Cosine', '55.56', '59.64', '89.45', '0.13', '0.08', '0.21'], ['L2', '53.73', '59.91', '88.64', '0.13', '0.07', '0.20'], ['L1', '[BOLD] 56.32', '[BOLD] 60.39', '[BOLD] 89.98', '0.13', '0.07', '0.20']] | Overall, these results indicate that the evaluated objective functions behave similarly in this task, however L1 distance based objective function yields slightly better results. |
Pretrained Semantic Speech Embeddings for End-to-End Spoken Language Understanding via Cross-Modal Teacher-Student Learning | 2007.01836 | Table 3: Effect of layers fine-tuning | ['ASR', 'NLU', 'Accuracy on Test, %', 'Accuracy on Test, %', 'Accuracy on Test, %', 'Validation'] | [['layers', 'layers', 'SwBD', 'MRDA', 'FSC', 'loss'], ['0', '0', '43.76', '56.08', '68.07', '0.26'], ['0', '1', '37.61', '56.47', '85.53', '0.19'], ['1', '0', '52.37', '[BOLD] 60.21', '86.42', '0.16'], ['1', '1', '52.05', '58.32', '[BOLD] 86.82', '0.17'], ['2', '0', '52.93', '59.42', '85.76', '[BOLD] 0.15'], ['3', '0',... | While it is not completely clear how many layers should be fine-tuned, we can conclusively tell that fine-tuning of former ASR encoder layers is more beneficial than former NLU layers. We decide to fine-tune the two top former ASR encoder layers. The results also illustrate that the optimization of SLU model for smalle... |
Pretrained Semantic Speech Embeddings for End-to-End Spoken Language Understanding via Cross-Modal Teacher-Student Learning | 2007.01836 | Table 4: Effect of learning rate schedule | ['Warmup', 'LR', 'Epochs', 'Accuracy on Test, %', 'Accuracy on Test, %', 'Accuracy on Test, %', 'Validation'] | [['steps', 'constant', '[EMPTY]', 'SwBD', 'MRDA', 'FSC', 'loss'], ['200,000', '50', '10', '52.65', '59.90', '83.94', '0.17'], ['300,000', '50', '10', '52.93', '59.42', '85.76', '0.15'], ['400,000', '50', '10', '51.90', '59.43', '85.79', '0.15'], ['600,000', '50', '10', '51.18', '59.95', '86.84', '0.14'], ['600,000', '5... | After deciding which layers to fine-tune, we run a series of experiments to determine the best learning rate schedule. When we increase number of warmup steps, we notice positive effect from slower learning rate ramp up. However, as number number of warmup steps becomes close to the total number of fine-tuning steps, w... |
One-to-Many Multilingual End-to-End Speech Translation | 1910.03320 | Table 3: Results for multilingual direct SLT systems with 6 and 8 target languages. | ['[EMPTY]', 'De', 'Nl', 'Es', 'Fr', 'It', 'Pt', 'Ro', 'Ru'] | [['[EMPTY]', '[BOLD] Baseline', '[BOLD] Baseline', '[BOLD] Baseline', '[BOLD] Baseline', '[BOLD] Baseline', '[BOLD] Baseline', '[BOLD] Baseline', '[BOLD] Baseline'], ['[EMPTY]', '17.3', '18.8', '20.8', '26.9', '16.8', '20.1', '16.5', '10.5'], ['[EMPTY]', '[BOLD] Multilingual', '[BOLD] Multilingual', '[BOLD] Multilingua... | Number of languages. When training a system with all the 6 target languages (De, Nl, Es, Fr, It, Pt) When adding ASR data to the 6 languages, we observe improvements in most languages, and the new system is worse than the baseline only for Spanish and French, although the gap for French has been reduced to −0.9. Howeve... |
One-to-Many Multilingual End-to-End Speech Translation | 1910.03320 | Table 2: Results with concat (C-*) and merge (M-*) target forcing on 6 languages. The baselines are one-to-one systems. All the other results are computed with one multilingual system for En→De,NL and one for En→Es,Fr,It,Pt. | ['[EMPTY]', 'De', 'Nl', 'Es', 'Fr', 'It', 'Pt'] | [['Baseline', '17.3', '18.8', '20.8', '[BOLD] 26.9', '16.8', '20.1'], ['C-Pre', '14.0', '11.6', '13.0', '16.3', '10.7', '14.5'], ['C-Post', '12.0', '13.8', '12.3', '18.0', '9.3', '14.6'], ['C-Final', '14.5', '12.1', '13.6', '16.7', '10.2', '16.2'], ['M-Pre', '17.6', '19.5', '20.5', '26.2', '17.2', '22.3'], ['M-Post', '... | Concat vs Merge. Our first experiment consists in comparing the baselines with the multilingual models based on the target forcing mechanism. By looking at the translations, we found that the cause of the degradation is that many sentences are acceptable translations, but in a wrong language. We first hypothesize that ... |
One-to-Many Multilingual End-to-End Speech Translation | 1910.03320 | Table 4: Comparison of the Baseline and the best multilingual system with the single language cascade (BL-Cascade) and the multilingual cascade (M-Cascade) | ['[EMPTY]', 'De', 'Nl', 'Es', 'Fr', 'It', 'Pt'] | [['Baseline', '17.3', '18.8', '20.8', '26.9', '16.8', '20.1'], ['M-Pre + ASR', '17.7', '20.0', '20.9', '26.5', '18.0', '22.6'], ['BL-Cascade', '18.5', '22.2', '22.5', '27.9', '18.9', '21.5'], ['M-Cascade', '18.6', '22.0', '22.1', '27.3', '18.5', '22.8']] | Comparison with cascade. As expected, our direct SLT baselines are significantly worse than the BL-cascade systems with differences that range from −1.0 for French to −3.4 for Dutch. Comparing the BL-Cascade with the M-Cascade systems, we observe not significant variation for the Germanic languages, but lower results i... |
One-to-Many Multilingual End-to-End Speech Translation | 1910.03320 | Table 5: Percentage of sentences in the correct language computed with langdetect. | ['[EMPTY]', 'De', 'Nl', 'Es', 'Fr', 'It', 'Pt'] | [['M-Pre', '95.7', '98.5', '97.2', '94.6', '95.3', '96.6'], ['M-Pre + ASR', '96.1', '98.7', '97.9', '95.3', '95.4', '95.2']] | Language analysis. Then, when using also ASR data, the percentage of correct language increases slightly in all languages except for Portuguese. However, the improvement in the correct language does not correlate with the improvement in BLEU score. This suggests that the improvement in BLEU score of M-PRE + ASR comes f... |
eRevise: Using Natural Language Processing to Provide Formative Feedback on Text Evidence Usage in Student Writing | 1908.01992 | Table 3: Quadratic Weighted Kappa (QWK) of different AES models. The CO-ATTN model significantly outperforms the Rubric and SG models, respectively (p<0.05). | ['[BOLD] AES Model', '[BOLD] QWK'] | [['Rubric', '0.632'], ['SG', '0.653'], ['CO-ATTN', '0.697']] | We have developed several AES systems for RTA assessment Our first model (denoted by Rubric) A subsequent model (denoted by SG) Most recently, \citeauthorzhang2018co \shortcitezhang2018co developed a neural network model with a co-attention layer (denoted by CO-ATTN) to eliminate human feature engineering. Although the... |
eRevise: Using Natural Language Processing to Provide Formative Feedback on Text Evidence Usage in Student Writing | 1908.01992 | Table 5: Lookup table for feedback selection. | ['[BOLD] Feature [ITALIC] NPE', '[BOLD] Value 0', '[BOLD] Value 0', '[BOLD] Value 0', '[BOLD] Value 1', '[BOLD] Value 2', '[BOLD] Value 3', '[BOLD] Value 4', '[BOLD] Value 1', '[BOLD] Value 1', '[BOLD] Value 2', '[BOLD] Value 2', '[BOLD] Value 3', '[BOLD] Value 4', '[BOLD] Value 3', '[BOLD] Value 4'] | [['[ITALIC] SPClmh', 'L', 'M', 'H', 'L', 'L', 'L', 'L', 'M', 'H', 'H', 'M', 'M', 'M', 'H', 'H'], ['[BOLD] Feedback Messages', '1,2', '1,2', '1,2', '1,2', '1,2', '1,2', '1,2', '1,2', '1,2', '1,2', '2,3', '2,3', '2,3', '3,4', '3,4']] | SPCAWE=3, and SPClmh =M. We are about to begin the next deployment of eRevise, which will extend our work in two ways. First, to better determine the benefit of using AES to adaptively guide revision, we have added a control condition where eRevise will display the same generic feedback message to all students: “MAKE Y... |
Graph-Based Decoding for Event Sequencing and Coreference Resolution | 1806.05099 | Table 2: Test Results for Event Coreference with the Singleton and Matching baselines. | ['[EMPTY]', '[ITALIC] B3', 'CEAF-E', 'MUC', 'BLANC', 'AVG.'] | [['ALL', '81.97', '74.80', '76.33', '76.07', '77.29'], ['-Distance', '81.92', '74.48', '76.02', '77.55', '77.50'], ['-Frame', '82.14', '75.01', '76.28', '77.74', '77.79'], ['-Syntactic', '81.87', '74.89', '75.79', '76.22', '77.19']] | Comparing to the top 3 coreference systems in TAC-KBP 2015, we outperform the best system by about 2 points absolute F-score on average. Our system is also competitive on individual metrics. Our model performs the best based on B3 and CEAF-E, and is comparable to the top performing systems on MUC and BLANC. Note that w... |
Graph-Based Decoding for Event Sequencing and Coreference Resolution | 1806.05099 | Table 2: Test Results for Event Coreference with the Singleton and Matching baselines. | ['[EMPTY]', '[ITALIC] B3', 'CEAF-E', 'MUC', 'BLANC', 'AVG.'] | [['Singleton', '78.10', '68.98', '0.00', '48.88', '52.01'], ['Matching', '78.40', '65.82', '[BOLD] 69.83', '76.29', '71.94'], ['LCC', '82.85', '74.66', '68.50', '[BOLD] 77.61', '75.69'], ['UI-CCG', '83.75', '75.81', '63.78', '73.99', '74.28'], ['LTI', '82.27', '75.15', '60.93', '71.57', '72.60'], ['This work', '[BOLD] ... | Comparing to the top 3 coreference systems in TAC-KBP 2015, we outperform the best system by about 2 points absolute F-score on average. Our system is also competitive on individual metrics. Our model performs the best based on B3 and CEAF-E, and is comparable to the top performing systems on MUC and BLANC. Note that w... |
Graph-Based Decoding for Event Sequencing and Coreference Resolution | 1806.05099 | Table 4: Test Results for event sequencing. The Oracle Cluster+Temporal system is using Caevo’s result on the Oracle Clusters. | ['[EMPTY]', 'Prec.', 'Recall', 'F-Score'] | [['Oracle Cluster+Temporal', '[BOLD] 46.21', '8.72', '14.68'], ['Our Model', '18.28', '[BOLD] 16.91', '[BOLD] 17.57']] | Because the baseline system has access to the oracle script clusters, it produces high precision. However, the low recall value shows that it fails to produce enough After links. Our analysis shows that a lot of After relations are not indicated by clear temporal clues, but can only be solved with script knowledge. How... |
Graph-Based Decoding for Event Sequencing and Coreference Resolution | 1806.05099 | Table 5: Ablation Study for Event Sequencing. | ['[EMPTY]', 'Prec.', 'Recall', 'F-Score', 'Δ'] | [['Full', '37.92', '36.79', '36.36', '[EMPTY]'], ['- Mention Type', '32.78', '29.81', '30.07', '6.29'], ['- Sentence', '33.90', '30.75', '31.00', '5.36'], ['- Temporal', '37.21', '36.53', '35.81', '0.55'], ['- Dependency', '38.18', '36.44', '36.23', '0.13'], ['- Function words', '38.08', '36.51', '36.18', '0.18']] | While most of the features only affect the performance by less than 1 absolute F1 score, the feature sets after removing mention or sentences show a significant drop in both precision and recall. This shows that discourse proximity is the most significant ones among these features. In addition, the mention feature set ... |
Impact of Batch Size on Stopping Active Learning for Text Classification | 1801.07887 | TABLE I: Stopping method results on 20Newsgroups for different batch sizes using various window sizes. The top number in each row shows the number of annotations at the stopping point and the bottom number shows the F-Measure at the stopping point. | ['[width=15em]Stopping MethodBatch Percent', '1%', '5%', '10%'] | [['Oracle Method', '1514.20', '2490.40', '3901.95'], ['Oracle Method', '76.17', '75.55', '75.49'], ['BV2009 (Window Size = 3)', '1299.50', '3877.10', '6446.70'], ['BV2009 (Window Size = 3)', '74.44', '75.17', '75.19'], ['BV2009 (Window Size = 1)', '1101.75', '3141.30', '5089.50'], ['BV2009 (Window Size = 1)', '74.40', ... | We considered different batch sizes in our experiments, based on percentages of the entire set of training data. We ran BV2009 with smaller window sizes for each of our different batch sizes. When using a window size of one, BV2009 is able to stop with a smaller number of annotations than when using a window size of th... |
Visually grounded cross-lingual keyword spotting in speech | 1806.05030 | Table 3: Analysis of errors by a human annotator of the top ten retrievals on development data. Percentages (%) indicate the absolute drop in P@10 due to that error type. | ['Error type', 'XVisionSpeech Count', 'XVisionSpeech %', 'XBoWCNN Count', 'XBoWCNN %'] | [['(1) Correct (exact)', '032', '08.2', '45', '11.5'], ['(2) Semantically related', '086', '22.1', '13', '03.3'], ['(3) Incorrect retrieval', '035', '09.0', '19', '04.9'], ['Total', '153', '39.3', '77', '19.7']] | Transcriptions of the top English utterances retrieved using XVisionSpeechCNN for a selection of German keywords. Errors from both XVisionSpeechCNN and XBoWCNN were presented to the annotator in shuffled order. For both models, around 10% of the retrievals marked as errors are actually correct. The bulk of errors from ... |
Visually grounded cross-lingual keyword spotting in speech | 1806.05030 | Table 2: Cross-lingual keyword spotting results (%) on test data. | ['Model', '[ITALIC] P@10', '[ITALIC] P@ [ITALIC] N', 'EER', 'AP'] | [['DETextPrior', '07.2', '06.3', '50', '10.4'], ['DEVisionCNN', '41.5', '32.9', '25.9', '29.7'], ['XVisionSpeechCNN', '58.2', '40.4', '23.5', '40.0'], ['XBoWCNN', '80.8', '54.3', '19.1', '54.3']] | Without seeing any speech transcriptions or translated text, XVisionSpeechCNN achieves a P@10 of 58%, with XBoWCNN the only model to outperform the visually grounded model. By comparing performance to DETextPrior, we see that XVisionSpeechCNN is not just predicting common German words. Interestingly, XVisionSpeechCNN a... |
Visually grounded cross-lingual keyword spotting in speech | 1806.05030 | Table 4: Cross-lingual keyword spotting results (%) for different variants of XVisionSpeechCNN on development data. | ['Model', '[ITALIC] P@10', '[ITALIC] P@ [ITALIC] N', 'EER', 'AP'] | [['XVisionSpeechCNN', '60.8', '39.3', '23.1', '38.0'], ['KeyXVisionSpeechCNN', '60.0', '39.6', '24.5', '36.9'], ['OracleXVisionSpeechCNN', '57.4', '37.6', '24.8', '36.5']] | Variants and ideal supervision. We compare different variants of XVisionSpeechCNN to gain insight into properties of the model. XVisionSpeechCNN produces scores →f(X)∈[0,1]W for all W=1k words in its output vocabulary. But we are actually only interested in those dimensions w corresponding to the test keywords. If we k... |
Flexible End-to-End Dialogue System for Knowledge Grounded Conversation | 1709.04264 | Table 4: Human Evaluation on the MusicConvers dataset. | ['Models', 'Grammar', 'Context Relevance', 'Correctness'] | [['S2SA', '1.76', '0.87', '0.16'], ['GenQA', '1.28', '0.95', '0.41'], ['GenQAD', '1.67', '1.11', '0.51'], ['GenDS-Single', '[BOLD] 2.16', '[BOLD] 1.67', '[BOLD] 1.18'], ['GenDS-Static', '1.97', '1.42', '0.96'], ['GenDS', '2.03', '1.55', '0.89']] | MusicConvers: We compute the mean score of each metric. For automatic evaluation, GenDS shows the best performance on BLEU and entity-accuracy, while GenDS-single achieves the highest entity-recall. Although GenDS does not overwhelm on S2SA in terms of BLEU, it improves entity accuracy and recall by 39% and 14% respect... |
Flexible End-to-End Dialogue System for Knowledge Grounded Conversation | 1709.04264 | Table 3: Automatic Evaluation on the Music dataset | ['Models', 'BLEU', 'Precision', 'Recall'] | [['S2SA', '0.11', '0.01±0.01', '0.004±0.02'], ['GenQA', '0.05', '0.1134±0.14', '0.05±0.1'], ['GenQAD', '0.06', '0.15±0.16', '0.05±0.1'], ['GenDS-Single', '0.108', '0.28±0.19', '[BOLD] 0.19±0.18'], ['GenDS-Static', '0.108', '0.14±0.15', '0.10±0.14'], ['GenDS', '[BOLD] 0.122', '[BOLD] 0.40\xa0±0.25', '0.14±0.16']] | MusicConvers: We compute the mean score of each metric. For automatic evaluation, GenDS shows the best performance on BLEU and entity-accuracy, while GenDS-single achieves the highest entity-recall. Although GenDS does not overwhelm on S2SA in terms of BLEU, it improves entity accuracy and recall by 39% and 14% respect... |
Flexible End-to-End Dialogue System for Knowledge Grounded Conversation | 1709.04264 | Table 5: Automatic Evaluation on the QA dataset | ['Models', 'BLEU', 'Precision', 'Recall'] | [['S2SA', '0.05', '0.08±0.125', '0.07±0.13'], ['GenQA', '0.12', '0.06±0.11', '0.04±0.09'], ['GenQAD', '0.13', '0.25±0.2', '0.34±0.235'], ['GenDS-Single', '0.226', '0.76±0.205', '[BOLD] 0.77±0.21'], ['GenDS-Static', '0.19', '0.64±0.23', '0.66±0.235'], ['GenDS', '[BOLD] 0.227', '[BOLD] 0.77±0.205', '0.76±0.215']] | MusicQA GenDS improves BLEU score, entity-accuracy and entity-recall significantly compared with S2SA, GenQA and GenQAD. GenQA does not obtain comparable performance with the original QA This may be the due to our mitigation of redundancy for the dataset. Unlike MusicConvers, GenDS-static exhibits decent performance on... |
RETURNN as a Generic Flexible Neural Toolkit with Application to Translation and Speech Recognition | 1805.05225 | Table 1: Training speed and memory consumption on WMT 2017 German→English. Train time is for seeing the full train dataset once. Batch size is in words, such that it almost maximizes the GPU memory consumption. The BLEU score is for the converged models, reported for newstest2015 (dev) and newstest2017. The encoder has... | ['toolkit', 'encoder n. layers', 'time [h]', 'batch size', 'BLEU [%] 2015', 'BLEU [%] 2017'] | [['RETURNN', '4', '[BOLD] 11.25', '8500', '28.0', '28.4'], ['Sockeye', '[EMPTY]', '11.45', '3000', '[BOLD] 28.9', '[BOLD] 29.2'], ['RETURNN', '6', '[BOLD] 12.87', '7500', '28.7', '28.7'], ['Sockeye', '[EMPTY]', '14.76', '2500', '[BOLD] 29.4', '[BOLD] 29.1']] | We want to compare different toolkits in training and decoding for a recurrent attention model in terms of speed on a GPU. Here, we try to maximize the batch size such that it still fits into the GPU memory of our reference GPU card, the Nvidia GTX 1080 Ti with 11 GB of memory. We keep the maximum sequence length in a ... |
RETURNN as a Generic Flexible Neural Toolkit with Application to Translation and Speech Recognition | 1805.05225 | Table 3: Comparison on German→English. | ['toolkit', 'BLEU [%] 2015', 'BLEU [%] 2017'] | [['RETURNN', '[BOLD] 31.2', '[BOLD] 31.3'], ['Sockeye', '29.7', '30.2']] | We report the best performing Sockeye model we trained, which has 1 bidirectional and 3 unidirectional encoder layers, 1 pre-attention target recurrent layer, and 1 post-attention decoder layer. We trained with a max sequence length of 75, and used the ‘coverage’ RNN attention type. For Sockeye, the final model is an a... |
RETURNN as a Generic Flexible Neural Toolkit with Application to Translation and Speech Recognition | 1805.05225 | Table 4: Performance comparison on WMT 2017 English→German. The baseline systems (upper half) are trained on the parallel data of the WMT Enlgish→German 2017 task. We downloaded the hypotheses from here.66footnotemark: 6 The WMT 2017 system hypotheses (lower half) are generated using systems having additional back- tra... | ['System', 'BLEU [%]'] | [['[EMPTY]', 'newstest2017'], ['RETURNN', '[BOLD] 26.1'], ['OpenNMT-py', '21.8'], ['OpenNMT-lua', '22.6'], ['Marian', '25.6'], ['Nematus', '23.5'], ['Sockeye', '25.3'], ['WMT 2017 Single Systems + bt data', 'WMT 2017 Single Systems + bt data'], ['LMU', '26.4'], ['+ reranking', '27.0'], ['Systran', '26.5'], ['Edinburgh'... | We observe that our toolkit outperforms all other toolkits. The best result obtained by other toolkits is using Marian (25.5% BLEU). In comparison, RETURNN achieves 26.1%. We also compare RETURNN to the best performing single systems of WMT 2017. In comparison to the fine-tuned evaluation systems that also include back... |
RETURNN as a Generic Flexible Neural Toolkit with Application to Translation and Speech Recognition | 1805.05225 | Table 5: Performance comparison on Switchboard, trained on 300h. hybrid1 is the IBM 2017 ResNet model Saon et al. (2017). hybrid2 trained with Lattice-free MMI Hadian et al. (2018). CTC3 is the Baidu 2014 DeepSpeech model Hannun et al. (2014). Our attention model does not use any language model. | ['model', 'training', 'WER [%] Hub5’00', 'WER [%] Hub5’00', 'WER [%] Hub5’00', 'WER [%] Hub5’01'] | [['[EMPTY]', '[EMPTY]', 'Σ', 'SWB', 'CH', '[EMPTY]'], ['hybrid1', 'frame-wise', '[EMPTY]', '11.2', '[EMPTY]', '[EMPTY]'], ['hybrid2', 'LF-MMI', '15.8', '10.8', '[EMPTY]', '[EMPTY]'], ['CTC3', 'CTC', '25.9', '20.0', '31.8', '[EMPTY]'], ['hybrid', 'frame-wise', '[BOLD] 14.4', '[BOLD] 9.8', '[BOLD] 19.0', '14.7'], ['[EMPT... | We also have preliminary results with recurrent attention models for speech recognition on the Switchboard task, which we trained on the 300h trainset. We report on both the Switchboard (SWB) and the CallHome (CH) part of Hub5’00 and Hub5’01. We also compare to a conventional frame-wise trained hybrid deep bidirectiona... |
RETURNN as a Generic Flexible Neural Toolkit with Application to Translation and Speech Recognition | 1805.05225 | Table 6: Pretraining comparison. | ['encoder num. layers', 'BLEU [%] no pretrain', 'BLEU [%] with pretrain'] | [['2', '29.3', '-'], ['3', '29.9', '-'], ['4', '29.1', '30.3'], ['5', '-', '30.3'], ['6', '-', '30.6'], ['7', '-', '[BOLD] 30.9']] | RETURNN supports very generic and flexible pretraining which iteratively starts with a small model and adds new layers in the process. A similar pretraining scheme for deep bidirectional LSTMs acoustic speech models was presented earlier Zeyer et al. Here, we only study a layer-wise construction of the deep bidirection... |
Assessing Language Proficiency from Eye Movements in Reading | 1804.07329 | Table 1: Pearson’s r of EyeScore for different feature sets with MET (training/development set, 88 participants) and TOEFL (all 53 participants). Fixed denotes the Fixed Text regime in which all the participants read the same sentences, and Any denotes the Any Text regime where different readers read different sentence... | ['[BOLD] Features', '[BOLD] MET Fixed', '[BOLD] MET Any', '[BOLD] TOEFL Fixed', '[BOLD] TOEFL Any'] | [['Reading Speed', '0.28', '0.27', '0.15', '0.13'], ['WP-Coefficients', '0.38', '0.37', '0.21', '0.13'], ['S-Clusters', '0.45', '[BOLD] 0.48', '0.50', '[BOLD] 0.45'], ['Transitions', '0.45', '[EMPTY]', '0.44', '[EMPTY]'], ['WFC', '[BOLD] 0.50', '[EMPTY]', '[BOLD] 0.54', '[EMPTY]']] | We evaluate the ability of EyeScore to capture language proficiency by comparing it against our two external proficiency tests, MET and TOEFL. Similarly to the EyeScore outcomes, the best performance in the Fixed Text regime is obtained using the WFC feature set, with a Pearson’s r of 0.7 and MAE of 3.31 for MET. This ... |
Assessing Language Proficiency from Eye Movements in Reading | 1804.07329 | Table 2: Pearson’s r and Mean Absolute Error (MAE) for prediction of MET scores (test set, 57 participants) and TOEFL scores (leave-one-out cross validation, all 53 participants) from eye movement patterns in reading. We consider two baselines which do not use eyetracking information: (1) the average proficiency score ... | ['[EMPTY]', '[BOLD] MET Fixed', '[BOLD] MET Fixed', '[BOLD] MET Any', '[BOLD] MET Any', '[BOLD] TOEFL Fixed', '[BOLD] TOEFL Fixed', '[BOLD] TOEFL Any', '[BOLD] TOEFL Any'] | [['[BOLD] Features', '[ITALIC] r', 'MAE', '[ITALIC] r', 'MAE', '[ITALIC] r', 'MAE', '[ITALIC] r', 'MAE'], ['Reading Speed', '0.27', '4.58', '0.24', '4.62', '0.09', '7.92', '0.06', '7.96'], ['WP-Coefficients', '0.43', '4.11', '0.44', '4.14', '0.34', '7.76', '0.31', '[BOLD] 7.49'], ['S-Clusters', '0.56', '3.87', '[BOLD] ... | We consider two baselines; the first is assigning all test set participants with the average score of the training participants. This baseline yields an MAE of 4.82 on MET and 8.29 on TOEFL. The second baseline uses reading speed as the sole feature for prediction. In all cases, our eyetracking based features outperfor... |
Learning Paraphrastic Sentence Embeddingsfrom Back-Translated Bitext | 1706.01847 | Table 9: Differences in entropy and repetition of unigrams/trigrams in references and translations. Negative values indicate translations have a higher value, so references show consistently higher entropies and lower repetition rates. | ['Lang.', 'Data', 'Ent. (uni)', 'Ent. (tri)', 'Rep. (uni)', 'Rep. (tri)'] | [['[EMPTY]', 'CC', '0.50', '1.13', '-7.57%', '-5.58%'], ['CS', 'EP', '0.14', '0.31', '-0.88%', '-0.11%'], ['[EMPTY]', 'News', '0.16', '0.31', '-0.96%', '-0.16%'], ['[EMPTY]', 'CC', '0.97', '1.40', '-8.50%', '-7.53%'], ['FR', 'EP', '0.51', '0.69', '-1.85%', '-0.58%'], ['FR', 'Giga', '0.97', '1.21', '-5.30%', '-7.74%'], ... | The translated text has lower n-gram entropies and higher rates of repetition. This appears for all datasets, but is strongest for common crawl and French-English 109. |
Learning Paraphrastic Sentence Embeddingsfrom Back-Translated Bitext | 1706.01847 | Table 5: Test correlations for our models when trained on sentences with particular length ranges (averaged over languages and data sources for the NMT rows). Results are on STS datasets (Pearson’s r×100). | ['Data', 'Model', 'Length Range 0-10', 'Length Range 10-20', 'Length Range 20-30', 'Length Range 30-100'] | [['SimpWiki', 'GRAN', '67.4', '67.7', '67.1', '67.3'], ['SimpWiki', 'Avg', '65.9', '65.7', '65.6', '65.9'], ['NMT', 'GRAN', '66.6', '66.5', '66.0', '64.8'], ['NMT', 'Avg', '65.7', '65.6', '65.3', '65.0']] | These results are averages across all language pairs and data sources of training data for each length range shown. We find it best to select NMT data where the translations have between 0 and 10 tokens, with performance dropping as sentence length increases. This is true for both the GRAN and Avg models. We do the sam... |
Learning Paraphrastic Sentence Embeddingsfrom Back-Translated Bitext | 1706.01847 | Table 6: Length filtering test results after tuning length ranges on development data (averaged over languages and data sources for the NMT rows). Results are on STS datasets (Pearson’s r×100). | ['Filtering Method', 'NMT GRAN', 'NMT Avg', 'SimpWiki GRAN', 'SimpWiki Avg'] | [['None (Random)', '66.9', '65.5', '67.2', '65.8'], ['Length', '67.3', '66.0', '67.4', '66.2'], ['Tuned Len. Range', '[0,10]', '[0,10]', '[0,10]', '[0,15]']] | We then tune the length range using our development data, considering the following length ranges: [0,10], [0,15], [0,20], [0,30], [0,100], [10,20], [10,30], [10,100], [15,25], [15,30], [15,100], [20,30], [20,100], [30,100]. We tune over ranges as well as language, data source, and stopping epoch, each time training on... |
Learning Paraphrastic Sentence Embeddingsfrom Back-Translated Bitext | 1706.01847 | Table 7: Quality filtering test results after tuning quality hyperparameters on development data (averaged over languages and data sources for the NMT rows). Results are on STS datasets (Pearson’s r×100). | ['Filtering Method', 'GRAN', 'Avg'] | [['None (Random)', '66.9', '65.5'], ['Translation Cost', '66.6', '65.4'], ['Language Model', '66.7', '65.5'], ['Reference Classification', '67.0', '65.5']] | The translation cost and language model are not helpful for filtering, as random selection outperforms them. Both methods are outperformed by the reference classifier, which slightly outperforms random selection when using the stronger GRAN model. We now discuss further how we trained the reference classifier and the d... |
Learning Paraphrastic Sentence Embeddingsfrom Back-Translated Bitext | 1706.01847 | Table 8: Results of reference/translation classification (accuracy×100). The highest score in each column is in boldface. Final two columns show accuracies of positive (reference) and negative classes, respectively. | ['Model', 'Lang.', 'Data', 'Test Acc.', '+ Acc.', '- Acc.'] | [['[EMPTY]', '[EMPTY]', 'CC', '72.2', '72.2', '72.3'], ['[EMPTY]', 'CS', 'EP', '72.3', '64.3', '80.3'], ['[EMPTY]', '[EMPTY]', 'News', '79.7', '73.2', '86.3'], ['[EMPTY]', '[EMPTY]', 'CC', '80.7', '82.1', '79.3'], ['LSTM', 'FR', 'EP', '79.3', '75.2', '83.4'], ['LSTM', 'FR', 'Giga', '[BOLD] 93.1', '[BOLD] 92.3', '93.8']... | While performance varies greatly across data sources, the LSTM always outperforms the word averaging model. For our translation-reference classification, we note that our results can be further improved. We also trained models on 90,000 examples, essentially doubling the amount of data, and the results improved by abou... |
Learning Paraphrastic Sentence Embeddingsfrom Back-Translated Bitext | 1706.01847 | Table 12: Diversity filtering test results after tuning filtering hyperparameters on development data (averaged over languages and data sources for the NMT rows). Results are on STS datasets (Pearson’s r×100). | ['Filtering Method', 'NMT GRAN', 'NMT Avg', 'SimpWiki GRAN', 'SimpWiki Avg'] | [['Random', '66.9', '65.5', '67.2', '65.8'], ['Unigram Overlap', '66.6', '66.1', '67.8', '67.4'], ['Bigram Overlap', '67.0', '65.5', '68.0', '67.2'], ['Trigram Overlap', '66.9', '65.4', '67.8', '66.6'], ['BLEU Score', '67.1', '65.3', '67.5', '66.5']] | We find that the diversity filtering methods lead to consistent improvements when training on SimpWiki. We believe this is because many of the sentence pairs in SimpWiki are near-duplicates and these filtering methods favor data with more differences. |
Learning Paraphrastic Sentence Embeddingsfrom Back-Translated Bitext | 1706.01847 | Table 13: Test results when using more training data. More data helps both Avg and GRAN, where both models get close to or surpass training on SimpWiki and comfortably surpass the PPDB baseline. The amount of training examples used is in parentheses. | ['Data', 'GRAN', 'Avg'] | [['PPDB', '64.6', '66.3'], ['SimpWiki (100k/168k)', '67.4', '[BOLD] 67.7'], ['CC-CS (24k)', '66.8', '-'], ['CC-DE (24k)', '-', '66.6'], ['CC-CS (100k)', '[BOLD] 68.5', '-'], ['CC-DE (168k)', '-', '67.6']] | PPDB is soundly beaten by both data sources. Moreover, the CC-CS and CC-DE data have greatly improved their performance from training on just 24,000 examples from 66.77 and 66.61 respectively, providing more evidence that this approach does indeed scale. |
1 Introduction | 1602.02068 | Table 3: Accuracies for the natural language inference task. Shown are our implementations of a system without attention, and with logistic, soft, and sparse attentions. | ['[EMPTY]', 'Dev Acc.', 'Test Acc.'] | [['NoAttention', '81.84', '80.99'], ['LogisticAttention', '82.11', '80.84'], ['SoftAttention', '82.86', '82.08'], ['SparseAttention', '82.52', '[BOLD] 82.20']] | We observe that the soft and sparse-activated attention systems perform similarly, the latter being slightly more accurate on the test set, and that both outperform the NoAttention and LogisticAttention systems. |
DuTongChuanDuTongChuan, abbreviation of Baidu simultaneous interpreting in Chinese Pinyin. : Context-aware Translation Model for Simultaneous Interpreting | 1907.12984 | Table 5: Comparison between machine translation and human interpretation. The interpretation reference consists of a collection of interpretations from S, A and B. Our model is trained on the large-scale corpus. | ['Models', 'Translation Reference BLEU', 'Translation Reference Brevity Penalty', 'Interpretation Reference (3-references) BLEU', 'Interpretation Reference (3-references) Brevity Penalty'] | [['[ITALIC] Our Model', '20.93', '1.000', '28.08', '1.000'], ['[ITALIC] S', '16.02', '0.845', '-', '-'], ['[ITALIC] A', '16.38', '0.887', '-', '-'], ['[ITALIC] B', '12.08', '0.893', '-', '-']] | We concatenate the translation of each talk into one big sentence, and then evaluate it by BLEU score. Moreover, the length of interpretations are relatively short, and results in a high length penalty provided by the evaluation script. The result is unsurprising, because human interpreters often deliberately skip non-... |
DuTongChuanDuTongChuan, abbreviation of Baidu simultaneous interpreting in Chinese Pinyin. : Context-aware Translation Model for Simultaneous Interpreting | 1907.12984 | Table 2: The overall results on NIST Chinese-English translation task. | ['Models', 'NIST02', 'NIST03', 'NIST04', 'NIST05', 'NIST08', 'Average'] | [['[ITALIC] baseline', '49.40', '49.71', '50.03', '48.83', '44.38', '40.39'], ['[ITALIC] sub-sentence', '45.41', '45.62', '46.06', '43.63', '43.11', '37.31'], ['[ITALIC] wait-1', '38.37', '36.87', '38.17', '36.09', '35.31', '30.80'], ['[ITALIC] wait-3', '40.75', '39.30', '40.57', '38.18', '38.29', '32.85'], ['[ITALIC] ... | Effectiveness on Translation Quality. On an average the sub-sentence shows weaker performance by a 3.08 drop in BLEU score (40.39 → 37.31). Similarly, the wait-k model also brings an obvious decrease in translation quality, even with the best wait-15 policy, its performance is still worse than the baseline system, with... |
DuTongChuanDuTongChuan, abbreviation of Baidu simultaneous interpreting in Chinese Pinyin. : Context-aware Translation Model for Simultaneous Interpreting | 1907.12984 | Table 3: The comparison between our sequence detector and previous work. The latency represents the words requiring to make an explicit decision. | ['Models', 'Precision (%)', 'Recall (%)', 'F-score (%)', 'Average Latency', 'Max Latency'] | [['[ITALIC] 5-LM', '55.30', '72.63', '62.79', '8.68', '46'], ['[ITALIC] RNN', '67.61', '70.35', '68.95', '9.79', '48'], ['[ITALIC] Our model', '[BOLD] 75.09', '[BOLD] 81.70', '[BOLD] 78.26', '10.49', '39']] | In our context-sensitive model, the dynamic context based information unit boundary detector is essential to determine the IU boundaries in the steaming input. Both of two contrastive models are trained on approximate 2 million monolingual Chinese sentences. This observation indicates that with bidirectional context, t... |
DuTongChuanDuTongChuan, abbreviation of Baidu simultaneous interpreting in Chinese Pinyin. : Context-aware Translation Model for Simultaneous Interpreting | 1907.12984 | Table 4: The overall results on BSTC Chinese-English translation task (Pre-train represents training on the NIST dataset, and fine-tune represents fine-tuning on the BSTC dataset.). Clean input indicates the input is from human annotated transcription, while the ASR input represents the input contains ASR errors. ASR +... | ['Models', 'Clean Input Pre-train', 'Clean Input Fine-tune', 'ASR Input Pre-train', 'ASR Input Fine-tune', 'ASR + Auto IU Pre-train', 'ASR + Auto IU Fine-tune'] | [['[ITALIC] baseline', '[BOLD] 15.85', '[BOLD] 21.98', '[BOLD] 14.60', '[BOLD] 19.91', '[BOLD] 14.41', '[BOLD] 17.35'], ['[ITALIC] sub-sentence', '14.39', '18.61', '13.50', '16.99', '13.76', '16.29'], ['[ITALIC] wait-3', '12.23', '16.74', '11.62', '15.59', '11.75', '14.68'], ['[ITALIC] wait-5', '12.84', '17.70', '11.96... | Due to the relatively lower CER in ASR errors (10.32 %), the distinction between the clean input and the noisy input results in a BLEU score difference smaller than 2 points (15.85 vs. 14.60 for pre-train, and 21.98 vs. 19.91 for fine-tune). Despite the small size of the training data in BSTC, fine-tuning on this data ... |
XGPT: Cross-modal Generative Pre-Training for Image Captioning | 2003.01473 | Table 5: Results of image retrieval task on Flickr30k. | ['[EMPTY]', 'R@1', 'R@5', 'R@10'] | [['ViLBERT ', '58.2', '84.9', '91.5'], ['ViLBERT + augmentation', '[BOLD] 60.4', '[BOLD] 86.4', '[BOLD] 91.9']] | The higher relative gain on R@1 also indicates that the generator can produce high-quality image captions which can help the model better understand images. |
XGPT: Cross-modal Generative Pre-Training for Image Captioning | 2003.01473 | Table 4: Comparison of two masking methods on COCO Captions. | ['[EMPTY]', 'C', 'B@4', 'M', 'S'] | [['multi [MASK]', '117.8', '36.1', '28.2', '21.3'], ['single [MASK]', '[BOLD] 118.1', '[BOLD] 36.4', '[BOLD] 28.3', '[BOLD] 21.3']] | Comparing with one-stage pre-training, we find that each task combination pre-trained after the second stage (Row 6-8) gains approximately +2 on CIDEr. This indicates that two-stage pre-training with in-domain data enables the model to adapt to the downstream data better than only using out-of-domain pre-training. Comb... |
Movement Pruning: Adaptive Sparsity by Fine-Tuning | 2005.07683 | Table 3: Distillation-augmented performances for selected high sparsity levels. All pruning methods benefit from distillation signal further enhancing the ratio Performance VS Model Size. | ['[EMPTY]', 'BERT base fine-tuned', 'Remaining Weights (%)', 'MaP', '[ITALIC] L0 Regu', 'MvP', 'soft MvP'] | [['SQuAD - Dev EM/F1', '80.4/88.1', '10%', '70.2/80.1', '72.4/81.9', '75.6/84.3', '[BOLD] 76.6/ [BOLD] 84.9'], ['SQuAD - Dev EM/F1', '80.4/88.1', '3%', '45.5/59.6', '65.5/75.9', '67.5/78.0', '[BOLD] 72.9/ [BOLD] 82.4'], ['MNLI - Dev acc/MM acc', '84.5/84.9', '10%', '78.3/79.3', '78.7/79.8', '80.1/80.4', '[BOLD] 81.2/ [... | The training objective is a linear combination of the training loss and a knowledge distillation loss on the output distributions. Overall, we observe that the relative comparisons of the pruning methods remain unchanged while the performances are strictly increased. When combined with distillation, soft movement pruni... |
Movement Pruning: Adaptive Sparsity by Fine-Tuning | 2005.07683 | Table 2: Performance at high sparsity levels. (Soft) movement pruning outperforms current state-of-the art pruning methods at different high sparsity levels. | ['[EMPTY]', 'BERT base fine-tuned', 'Remaining Weights (%)', 'MaP', '[ITALIC] L0 Regu', 'MvP', 'soft MvP'] | [['SQuAD - Dev EM/F1', '80.4/88.1', '10%', '67.7/78.5', '69.9/80.1', '[BOLD] 71.9/ [BOLD] 81.7', '71.3/81.5'], ['SQuAD - Dev EM/F1', '80.4/88.1', '3%', '40.1/54.5', '61.6/73.6', '65.2/76.3', '[BOLD] 69.6/ [BOLD] 79.9'], ['MNLI - Dev acc/MM acc', '84.5/84.9', '10%', '77.8/79.0', '77.9/78.5', '79.3/79.5', '[BOLD] 80.7/ [... | Magnitude pruning on SQuAD achieves 54.5 F1 with 3% of the weights compared to 73.6 F1 with L0 regularization, 76.3 F1 for movement pruning, and 79.9 F1 with soft movement pruning. These experiments indicate that in high sparsity regimes, importance scores derived from the movement accumulated during fine-tuning induce... |
Closed-Book Training to Improve Summarization Encoder Memory | 1809.04585 | Table 7: ROUGE F1 and METEOR scores of sanity check ablations, evaluated on CNN/DM validation set. | ['[EMPTY]', 'ROUGE 1', 'ROUGE 2', 'ROUGE L'] | [['pg baseline', '37.73', '16.52', '34.49'], ['pg + ptrdec', '37.66', '16.50', '34.47'], ['pg-2layer', '37.92', '16.48', '34.62'], ['pg-big', '38.03', '16.71', '34.84'], ['pg + cbdec', '[BOLD] 38.87', '[BOLD] 16.93', '[BOLD] 35.38']] | Model Capacity: To validate and sanity-check that the improvements are the result of the inclusion of our closed-book decoder and not due to some trivial effects of having two decoders or larger model capacity (more parameters), we train a variant of our model with two duplicated (initialized to be different) attention... |
Closed-Book Training to Improve Summarization Encoder Memory | 1809.04585 | Table 1: ROUGE F1 and METEOR scores (non-coverage) on CNN/Daily Mail test set of previous works and our models. ‘pg’ is the pointer-generator baseline, and ‘pg + cbdec’ is our 2-decoder model with closed-book decoder(cbdec). The model marked with ⋆ is trained and evaluated on the anonymized version of the data. | ['[EMPTY]', 'ROUGE 1', 'ROUGE 2', 'ROUGE L', 'MTR Full'] | [['previous works', 'previous works', 'previous works', 'previous works', 'previous works'], ['⋆(Nallapati16)', '35.46', '13.30', '32.65', '[EMPTY]'], ['pg (See17)', '36.44', '15.66', '33.42', '16.65'], ['our models', 'our models', 'our models', 'our models', 'our models'], ['pg (baseline)', '36.70', '15.71', '33.74', ... | We first report our evaluation results on CNN/Daily Mail dataset. In the reinforced setting, our 2-decoder model still maintains significant (p<0.001) advantage in all metrics over the pointer-generator baseline. and have a 95% ROUGE-significance interval of at most ±0.25. |
Closed-Book Training to Improve Summarization Encoder Memory | 1809.04585 | Table 2: ROUGE F1 and METEOR scores (with-coverage) on the CNN/Daily Mail test set. Coverage mechanism See et al. (2017) is used in all models except the RL model Paulus et al. (2018). The model marked with ⋆ is trained and evaluated on the anonymized version of the data. | ['[EMPTY]', 'ROUGE 1', 'ROUGE 2', 'ROUGE L', 'MTR Full'] | [['previous works', 'previous works', 'previous works', 'previous works', 'previous works'], ['pg (See17)', '39.53', '17.28', '36.38', '18.72'], ['RL⋆ (Paulus17)', '39.87', '15.82', '36.90', '[EMPTY]'], ['our models', 'our models', 'our models', 'our models', 'our models'], ['pg (baseline)', '39.22', '17.02', '35.95', ... | In the reinforced setting, our 2-decoder model (RL + pg + cbdec) outperforms our strong RL baseline (RL + pg) by a considerable margin (stat. significance of p<0.001). Fig. |
Closed-Book Training to Improve Summarization Encoder Memory | 1809.04585 | Table 6: ROUGE F1 scores of ablation studies, evaluated on CNN/Daily Mail validation set. | ['[EMPTY]', 'ROUGE 1', 'ROUGE 2', 'ROUGE L'] | [['Fixed-encoder ablation', 'Fixed-encoder ablation', 'Fixed-encoder ablation', 'Fixed-encoder ablation'], ['pg baseline’s encoder', '37.59', '16.27', '34.33'], ['2-decoder’s encoder', '[BOLD] 38.44', '[BOLD] 16.85', '[BOLD] 35.17'], ['Gradient-Flow-Cut ablation', 'Gradient-Flow-Cut ablation', 'Gradient-Flow-Cut ablati... | Fixed-Encoder Ablation: Next, we conduct an ablation study in order to prove the qualitative superiority of our 2-decoder model’s encoder to the baseline encoder. To do this, we train two pointer-generators with randomly initialized decoders and word embeddings. For the first model, we restore the pre-trained encoder f... |
Towards Unsupervised Grammatical Error Correction using Statistical Machine Translation with Synthetic Comparable Corpus | 1907.09724 | Table 3: The effect of source languages for comparable corpus creation. These News Crawl corpora are as of 2017 version. The number of sentences in each dataset is approximately 20M, respectively. These results are obtained by USMTforward in iter 1. | ['Src', 'Precision', 'Recall', 'F0.5'] | [['Fi News Crawl', '29.17', '28.52', '29.04'], ['Ru News Crawl', '27.11', '29.84', '27.62'], ['Fr News Crawl', '25.05', '30.27', '25.94'], ['De News Crawl', '23.26', '26.04', '25.04']] | We also examine how source languages of machine translation affect performance. The outputs using Finnish data is the best score among various languages; the more similar to English the source-side data is, the lower the F0.5 score of the output. |
Towards Unsupervised Grammatical Error Correction using Statistical Machine Translation with Synthetic Comparable Corpus | 1907.09724 | Table 2: M2 and GLEU results. The bold scores represent the best score in unsupervised SMT. The underlined scores represent the best overall score. | ['[EMPTY]', 'iter', 'CoNLL-14 (M2) P', 'CoNLL-14 (M2) R', 'CoNLL-14 (M2) F0.5', 'JFLEG GLEU'] | [['No edit', '-', '-', '-', '-', '40.54'], ['Supervised NMT', '-', '53.11', '26.47', '44.21', '54.04'], ['Supervised SMT', '-', '43.02', '33.18', '40.61', '55.93'], ['Unsupervised SMT', '0', '21.82', '[BOLD] 36.75', '23.75', '49.94'], ['w/ forward_refine', '1', '[BOLD] 25.92', '32.65', '[BOLD] 27.04', '[BOLD] 50.65'], ... | The F0.5 score for USMTforward in iter 1 is 13.57 points lower than that of supervised SMT and 17.17 points lower than that of supervised NMT. On JFLEG, the highest score was achieved with USMTforward in iter 1 among the unsupervised SMT models; its GLEU scores are 5.28 points and 3.39 points lower than those of superv... |
Towards Unsupervised Grammatical Error Correction using Statistical Machine Translation with Synthetic Comparable Corpus | 1907.09724 | Table 4: GEC results with W&I+LOCNESS test data. | ['Team', 'TP', 'FP', 'FN', 'P', 'R', 'F0.5'] | [['UEDIN-MS', '2,312', '982', '2,506', '70.19', '47.99', '64.24'], ['Kakao&Brain', '2,412', '1,413', '2,797', '63.06', '46.30', '58.80'], ['LAIX', '1,443', '884', '3,175', '62.01', '31.25', '51.81'], ['CAMB-CUED', '1,814', '1,450', '2,956', '55.58', '38.03', '50.88'], ['UFAL, Charles University, Prague', '1,245', '1,22... | The F0.5 score for our system (TMU) is 28.31; this score is eighth among the nine teams. In particular, the number of false positives of our system is 4,314; this is the worst result of all. |
A Hybrid Retrieval-Generation Neural Conversation Model | 1904.09068 | Table 4. The hyper-parameter settings in the generation-based baselines and the generation module in the proposed hybrid neural conversation model. These settings are the optimized settings tuned with the validation data. | ['Models', 'Seq2Seq', 'Seq2Seq-Facts'] | [['Embedding size', '512', '256'], ['# LSTM layers in encoder/decoder', '2', '2'], ['LSTM hidden state size', '512', '256'], ['Learning rate', '0.0001', '0.001'], ['Learning rate decay', '0.5', '0.5'], ['# Steps between validation', '10000', '5000'], ['Patience of early stopping', '10', '10'], ['Dropout', '0.3', '0.3']... | Parameter Settings Hyper-parameters are tuned with the validation data. For the hyper-parameter settings in the hybrid ranking module, we set the window size of the convolution and pooling kernels as (6,6). The number of convolution kernels is 64. The dropout rate is set to 0.5. The margin in the pairwise-ranking hinge... |
A Hybrid Retrieval-Generation Neural Conversation Model | 1904.09068 | Table 9. The response generation performance when we vary the ratios of positive samples in distant supervision. | ['Model', 'Supervision # Positive', 'BLEU-1 BLEU', 'BLEU-1 ROUGE-L', 'BLEU-2 BLEU', 'BLEU-2 ROUGE-L', 'ROUGE-L BLEU', 'ROUGE-L'] | [['HybridNCM-RS', 'k’=1', '0.9022', '8.9596', '0.7547', '8.8351', '1.0964', '8.9234'], ['HybridNCM-RS', 'k’=2', '1.0649', '9.7241', '1.1099', '9.9168', '1.1019', '9.6216'], ['HybridNCM-RS', 'k’=3', '[BOLD] 1.3450', '[BOLD] 10.4078', '[BOLD] 1.1165', '[BOLD] 10.1584', '[BOLD] 1.1435', '[BOLD] 10.0928'], ['HybridNCM-RSF'... | We further analyze the impact of the ratios of positive/ negative training samples on the response generation performance. The value of k′ is the number of positive response candidates for each conversation context when we train the hybrid ranking module. When k′=1, we select one positive candidate from the ground trut... |
A Hybrid Retrieval-Generation Neural Conversation Model | 1904.09068 | Table 10. The response generation performance when we vary different distant supervision signals. This table shows the results for the setting “k’=3”, where there are 3 positive response candidates for each conversation context. “SentBLEU” denotes using sentence-level BLEU scores as distant supervision signals. | ['Model Supervision', 'HybridNCM-RS BLEU', 'HybridNCM-RS ROUGE-L', 'HybridNCM-RSF BLEU', 'HybridNCM-RSF ROUGE-L'] | [['BLEU-1', '[BOLD] 1.3450', '[BOLD] 10.4078', '[BOLD] 1.3695', '[BOLD] 10.3445'], ['BLEU-2', '1.1165', '10.1584', '0.8239', '9.8575'], ['ROUGE-L', '1.1435', '10.0928', '0.9838', '9.7961'], ['SentBLEU', '0.8326', '9.2887', '1.0631', '9.6338']] | We find that distant supervision signals like BLEU-1 are quite effective for training the hybrid ranking module. The sentence-level BLEU is not a good choice for the distant supervision signal. The reason is that the sentence-level BLEU is computed only based on the n-gram precision statistics for a given sentence pair... |
Contextual Graph Attention for Answering Logical Queries over Incomplete Knowledge Graphs | 1910.00084 | Table 2. Macro-average AUC and APR over test queries with different DAG structures are used to evaluate the performance. All and H-Neg. denote macro-averaged across all query types and query types with hard negative sampling (see Section 3.2.3). | ['Dataset Metric', 'Bio AUC', 'Bio AUC', 'Bio APR', 'Bio APR', 'DB18 AUC', 'DB18 AUC', 'DB18 APR', 'DB18 APR', 'WikiGeo19 AUC', 'WikiGeo19 AUC', 'WikiGeo19 APR', 'WikiGeo19 APR'] | [['[EMPTY]', 'All', 'H-Neg', 'All', 'H-Neg', 'All', 'H-Neg', 'All', 'H-Neg', 'All', 'H-Neg', 'All', 'H-Neg'], ['Billinear[mean_simple]', '81.65', '67.26', '82.39', '70.07', '82.85', '64.44', '85.57', '71.72', '81.82', '60.64', '82.35', '64.22'], ['Billinear[min_simple]', '82.52', '69.06', '83.65', '72.7', '82.96', '64.... | We use the ROC AUC score and average percentile rank (APR) as two evaluation metrics. All evaluation results are macro-averaged across queries with different DAG structures The hyper-parameters for the baseline models GQE are tuned using grid search and the best ones are selected. Then we follow the practice of Hamilto... |
How would you say that? Pictures elicit better NLG data from the crowd. | 1608.00339 | Table 3: Human evaluation of the data collected with each MR (** = p<0.01 and *** = p<0.001 for Pictorial versus Textual conditions). Italics denote averages across all numbers of attributes. | ['[EMPTY]', '[BOLD] Textual MR Mean', '[BOLD] Textual MR StDev', '[BOLD] Pictorial MR Mean', '[BOLD] Pictorial MR StDev'] | [['[ITALIC] Informativeness', '[ITALIC] 4.28**', '[ITALIC] 1.54', '[ITALIC] 4.51**', '[ITALIC] 1.37'], ['3 attributes', '4.02', '1.39', '4.11', '1.32'], ['5 attributes', '4.31', '1.54', '4.46', '1.36'], ['8 attributes', '4.52', '1.65', '4.98', '1.29'], ['[ITALIC] Naturalness', '[ITALIC] 4.09***', '[ITALIC] 1.56', '[ITA... | Informativeness was defined (on the questionnaires) as whether the utterance “provides enough useful information about the venue”. A two-way ANOVA was conducted to examine the effect of MR modality and the number of attributes on the perceived Informativeness. There was no statistically significant interaction between ... |
How would you say that? Pictures elicit better NLG data from the crowd. | 1608.00339 | Table 2: Nature of the data collected with each MR. Italics denote averages across all numbers of attributes. | ['[EMPTY]', '[BOLD] Textual MR Mean', '[BOLD] Textual MR StDev', '[BOLD] Pictorial MR Mean', '[BOLD] Pictorial MR StDev'] | [['[ITALIC] Time, sec', '[ITALIC] 347.18', '[ITALIC] 301.74', '[ITALIC] 352.05', '[ITALIC] 249.34'], ['3 attributes', '283.37', '265.82', '298.97', '272.44'], ['5 attributes', '321.75', '290.89', '355.56', '244.57'], ['8 attributes', '433.41', '325.04', '405.56', '215.43'], ['[ITALIC] Length, char', '[ITALIC] 100.83', ... | A two-way ANOVA was conducted to examine the effect of MR modality and the number of attributes on average task duration. The difference between two modalities was not significant, with p=0.76. There was no statistically significant interaction between the effects of modality and the number of attributes in the MR, on ... |
Feature Assisted bi-directional LSTM Model for Protein-Protein Interaction Identification from Biomedical Texts | 1807.02162 | Table 6: Comparative results of the proposed model (sdpLSTM) with different baselines and state-of-the-art systems. Ref. choi2016extraction ∗ and li2015approach ∗ denote the reimplementation of the systems proposed in choi2016extraction and li2015approach with the authors reported experimental setups. | ['[BOLD] Model', '[BOLD] Approach', '[BOLD] AIMED [BOLD] Precision', '[BOLD] AIMED [BOLD] Recall', '[BOLD] AIMED [BOLD] F1-Score', '[BOLD] BioInfer [BOLD] Precision', '[BOLD] BioInfer [BOLD] Recall', '[BOLD] BioInfer [BOLD] F1-Score'] | [['Baseline 1', 'MLP (SDP+Feature Embedding)', '59.73', '75.93', '66.46', '68.56', '72.05', '70.22'], ['Baseline 2', 'RNN (SDP+Feature Embedding)', '66.23', '74.72', '70.22', '71.89', '74.59', '73.21'], ['Proposed Model', 'sdpLSTM (SDP+Feature Embedding)', '91.10', '82.2', '86.45', '72.40', '83.10', '77.35'], ['hua2016... | We perform 10-fold cross validation on both the datasets. With no official development data set available, cross validation seems to be the most reliable method of evaluating our proposed model. To evaluate the performance of our model, we use standard recall, precision, and F1-score. The obtained results clearly show ... |
Feature Assisted bi-directional LSTM Model for Protein-Protein Interaction Identification from Biomedical Texts | 1807.02162 | Table 5: Comparison of different activation functions with same hyperparameters values | ['[BOLD] Activation Function', '[BOLD] F-Score (AiMed)', '[BOLD] F-Score (BioInfer)'] | [['Sigmoid', '86.45', '77.35'], ['ReLU', '85.77', '76.92'], ['tanh', '84.40', '75.79']] | We setup all the experiments by varying the hyper-parameter values and analyze the behaviors of our model. For AiMed dataset, we observed that addition of LSTM units improves the model performance to a certain extent. Thereafter, it keeps on decreasing gradually. We define an optimal value 64 for the same, via cross-va... |
When and Why is Document-level Context Usefulin Neural Machine Translation? | 1910.00294 | Tabelle 3: Comparison of document-level model architectures and complexity. | ['Approach', 'Context Encoder Architecture', 'Context Encoder #layers', '[BOLD] en-it Bleu [%]', '[BOLD] en-it Ter [%]', '[BOLD] en-de Bleu [%]', '[BOLD] en-de Ter [%]'] | [['Baseline', '⋅', '⋅', '31.4', '56.1', '28.9', '61.8'], ['Single-Encoder', 'Transformer', '6', '31.5', '57.2', '28.9', '61.4'], ['Multi-Encoder (Out.)', 'Transformer', '6', '31.3', '56.1', '29.1', '61.4'], ['Multi-Encoder (Seq.)', 'Transformer', '6', '32.6', '55.2', '29.9', '60.7'], ['Multi-Encoder (Para.)', 'Transfor... | Model Architecture The tested methods are equal or closest to: Single-Encoder: \newciteagrawal2018contextual Integration outside the decoder: \newcitevoita2018context without sharing the encoder hidden layers over current/context sentences Integration inside the decoder Sequential attention : Decoder integration of \ne... |
When and Why is Document-level Context Usefulin Neural Machine Translation? | 1910.00294 | Tabelle 4: Comparison of context word filtering methods. | ['Context sentence', '[BOLD] en-it Bleu [%]', '[BOLD] en-it Ter [%]', '[BOLD] en-de Bleu [%]', '[BOLD] en-de Ter [%]', '#tokens'] | [['None', '31.4', '56.1', '28.9', '61.8', '-'], ['Full sentence', '32.5', '54.8', '30.3', '59.9', '100%'], ['Remove stopwords', '32.2', '55.2', '30.3', '59.9', '63%'], ['Remove most frequent words', '32.1', '55.6', '30.2', '60.2', '51%'], ['Retain only named entities', '32.3', '55.4', '30.3', '60.3', '[BOLD] 13%'], ['R... | All filtering methods shrink the context input drastically without a significant loss of performance. Each method has its own motivation to retain only useful tokens in the context; the results show that they are all reasonable in practice. In particular, using only named entities as context input, we achieve the same ... |
Suicidal Ideation and Mental Disorder Detectionwith Attentive Relation Networks | 2004.07601 | Table 4: Statistical information of SuicideWatch and mental health related subreddits, i.e., SWMH dataset | ['Subreddit', '#/% of train', '#/% of valid.', '#/% of test'] | [['depression', '11,940/34.29', '3,032/34.83', '3,774/34.68'], ['SuicideWatch', '6,550/18.81', '1,614/18.54', '2,018/18.54'], ['Anxiety', '6,136/17.62', '1,508/17.32', '1,911/17.56'], ['offmychest', '5,265/15.12', '1,332/15.30', '1,687/15.50'], ['bipolar', '4,932/14.16', '1,220/14.01', '1,493/13.72']] | As severe mental health issues are very likely to lead to suicidal ideation, we also collect another dataset from some mental health related subreddits in Reddit.com to further the study of mental disorders and suicidal ideation. We name this dataset as Reddit SuicideWatch and Mental Health Collection, or SWMH for shor... |
Suicidal Ideation and Mental Disorder Detectionwith Attentive Relation Networks | 2004.07601 | Table 5: Selected linguistic statistical information of UMD dataset extracted by LIWC | ['Linguistic clues', 'Label -1', 'Label 0', 'Label 1'] | [['positive emotion', '3.30', '3.12', '2.96'], ['negative emotion', '1.56', '2.30', '2.74'], ['anxiety', '0.17', '0.33', '0.41'], ['sadness', '0.28', '0.50', '0.68'], ['family', '0.29', '0.39', '0.47'], ['friend', '0.43', '0.56', '0.54'], ['work', '2.54', '1.92', '1.80'], ['money', '1.13', '0.71', '0.61'], ['death', '0... | We have a brief exploratory analysis on the data. Some selected linguistic statistical information of UMD dataset extracted by Linguistic Inquiry and Word Count software (LIWC) The risk of suicide increases among labels of -1, 0, and 1. The linguistic inquiry results show that negative emotion, anxiety, and sadness are... |
Suicidal Ideation and Mental Disorder Detectionwith Attentive Relation Networks | 2004.07601 | Table 7: Comparison of different models on Reddit SWMH collection, where precision, recall, and F1 score are weighted average. | ['Model', 'Accuracy', 'Precision', 'Recall', 'F1'] | [['fastText', '0.5722', '0.5760', '0.5722', '0.5721'], ['CNN', '0.5657', '0.5925', '0.5657', '0.5556'], ['LSTM', '0.5934', '0.6032', '0.5934', '0.5917'], ['BiLSTM', '0.6196', '0.6204', '0.6196', '0.6190'], ['RCNN', '0.6096', '0.6161', '0.6096', '0.6063'], ['SSA', '0.6214', '0.6249', '0.6214', '0.6226'], ['RN', '[BOLD] ... | 5.2.2 Reddit SWMH Then, we perform experiments on the Reddit SWMH dataset, which contains both suicidal ideation and mental health issues to study the predictive performance of our model. It is a larger dataset with more instances when compared with the UMD dataset. Experiments on this dataset show the relational encod... |
Suicidal Ideation and Mental Disorder Detectionwith Attentive Relation Networks | 2004.07601 | Table 8: Performance comparison on Twitter dataset, where precision, recall, and F1 score are weighted average. | ['Model', 'Accuracy', 'Precision', 'Recall', 'F1'] | [['fastText', '0.7927', '0.7924', '0.7927', '0.7918'], ['CNN', '0.7885', '0.7896', '0.7885', '0.7887'], ['LSTM', '0.8021', '0.8094', '0.8021', '0.8039'], ['BiLSTM', '0.8208', '0.8207', '0.8208', '0.8195'], ['RCNN', '0.8094', '0.8089', '0.8094', '0.8090'], ['SSA', '0.8156', '0.8149', '0.8156', '0.8152'], ['RN', '[BOLD] ... | Twitter Collection Lastly, we conduct experiments on the Twitter dataset with similar settings of previous experiments. Unlike posts in Reddit, tweets in this dataset are short sequences due to the tweet’s length limit of 280 characters. Among these competitive methods, our model gains the best performance on these fou... |
Suicidal Ideation and Mental Disorder Detectionwith Attentive Relation Networks | 2004.07601 | Table 9: Performance on each class of UMD suicidality dataset | ['Label', 'Metrics', 'BiLSTM', 'SSA', 'RN'] | [['-1', 'Precision', '0.62', '0.57', '0.69'], ['-1', 'Recall', '0.77', '[BOLD] 0.92', '0.70'], ['-1', 'F1-score', '0.69', '0.70', '0.69'], ['[EMPTY]', 'Precision', '0.51', '0.57', '0.48'], ['1', 'Recall', '0.55', '0.31', '[BOLD] 0.62'], ['[EMPTY]', 'F1-score', '0.53', '0.41', '0.54'], ['[EMPTY]', 'Precision', '0.15', '... | This section studies the performance on each class of UMD dataset. We select two baselines with better performance for comparison. The results are shown in Fig. The proposed RN-based model is poor on predicting post without suicidality, but good at predicting posts with high suicide risk. Unfortunately, all these three... |
Suicidal Ideation and Mental Disorder Detectionwith Attentive Relation Networks | 2004.07601 | Table 10: Performance of different variants of risk indicator injection | ['Model', 'Accuracy', 'Precision', 'Recall', 'F1'] | [['BiLSTM', '0.8208', '0.8207', '0.8208', '0.8195'], ['BiLSTM+concat', '0.8167', '0.8262', '0.8167', '0.8190'], ['BiLSTM+RN+sentiment', '0.8240', '0.8246', '0.8240', '0.8239'], ['BiLSTM+RN+topic', '0.8198', '0.8177', '0.8198', '0.8183'], ['BiLSTM+RN+sent.+topic', '[BOLD] 0.8385', '[BOLD] 0.8381', '[BOLD] 0.8385', '[BOL... | We then conduct ablation study to explore several variants and compare their performance. We compare our complete framework with three different settings of injecting risk indicators. The BiLSTM+concat model concatenates final hidden state with sentiment and topic features. The BiLSTM+RN+sentiment and BiLSTM+RN+topic u... |
Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes | 1909.08700 | Table 3: Comparison between state-of-the-art models Merity et al. (2017); Yang et al. (2017) and a Simple LSTM, and the same models with Alleviated TOI. The comparison highlights how the addition of Alleviated TOI is able to improve state-of-the-art models, as well as a simple model that does not benefit from extensive... | ['Model', 'test ppl'] | [['AWD-LSTM Merity et al. ( 2017 )', '58.8'], ['AWD-LSTM + Alleviated TOI', '56.46'], ['AWD-LSTM-MoS Yang et al. ( 2017 )', '55.97'], ['AWD-LSTM-MoS + Alleviated TOI', '54.58'], ['Simple-LSTM', '75.36'], ['Simple-LSTM + Alleviated TOI', '74.44']] | Comparison with State of the Art and Simple LSTM. With the MoS model and an Alleviated TOI, we improve the current state of the art without fine tuning for the PTB dataset with 54.58 perplexity on the test set. AWD-LSTM-MoS Yang et al. We compare the results with the same hyper-parameters used on the original papers wi... |
Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes | 1909.08700 | Table 1: Data point repetition with period q for batch size K=20 and Alleviated TOI (P). | ['[ITALIC] P', '[BOLD] Period [ITALIC] q', '[BOLD] Repetitions'] | [['2', '10', '2'], ['5', '4', '5'], ['7', '20', '1'], ['10', '2', '10']] | The worst case is P=10 with 10 repetitions of the same data point within the same batch and the best case is P=7, which avoids any redundancy because the GCD of P and K is 1. |
Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes | 1909.08700 | Table 2: Perplexity score (PPL) comparison of the AWD model, on the three datasets, with batch sizes K=20 (PTB), K=80 (WT2) and K=60 (WT103), with different levels of Token Order Imbalance (TOI). With Alleviated TOI (P), we use a prime batch size of K=19 (PTB), K=79 (WT2) and K=59 (WT103). | ['Experiment', 'PTB', 'WT2', 'WT103'] | [['Extreme TOI', '63.49', '73.52', '36.19'], ['Inter-batch TOI', '64.20', '72.61', '36.39'], ['Standard TOI', '[BOLD] 58.94', '[BOLD] 65.86', '[BOLD] 32.94'], ['Alleviated TOI 2', '57.97', '65.14', '32.98'], ['Alleviated TOI 5', '57.14', '65.11', '33.07'], ['Alleviated TOI 7', '57.16', '64.79', '32.89'], ['Alleviated T... | We use the test perplexity after the same equivalent number of epochs. The different Alleviated TOI (P) experiments use a different number of overlapped sequence: An Alleviated TOI (P) means building and concatenating P overlapped sequences. Our results indicate that an Alleviated TOI (P) is better than the Standard TO... |
Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes | 1909.08700 | Table 4: Perplexity score (PPL) comparison on the PTB dataset and the AWD model. We use two different values for the batch size K --- the original one with K=20, and a prime one with K=19. The results directly corroborate the observation portrayed in Figure 4, where the obtained score is related to the diversity of gra... | ['Experiment', 'K=20', 'K=19'] | [['Alleviated TOI 2', '59.37', '57.97'], ['Alleviated TOI 5', '60.50', '57.14'], ['Alleviated TOI 7', '[BOLD] 56.70', '57.16'], ['Alleviated TOI 10', '65.88', '[BOLD] 56.46']] | We compare the scores of a prime batch size K=19 with the scores of the original batch size K=20 for the AWD model with Alleviated TOI (P). When using a prime batch size, we observe consistent and increasing results as P increases. With the original batch size K=20, we observe a strong performance for P=7, but a low pe... |
Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes | 1909.08700 | Table 5: Token order imbalance (TOI) comparison for the IEMOCAP dataset on a SER task using angry, happy, neutral and sad classes with a simple LSTM model. | ['Experiment', 'WA', 'UA'] | [['Extreme TOI (15k steps)', '0.475', '0.377'], ['Inter-batch TOI (15k steps)', '0.478', '0.386'], ['Standard TOI (15k steps)', '0.486', '0.404'], ['Alleviated TOI (15k steps)', '[BOLD] 0.553', '[BOLD] 0.489'], ['Alleviated TOI (60 epochs)', '[BOLD] 0.591', '[BOLD] 0.523']] | When choosing the Extreme TOI instead of the Standard TOI approach we observe a smaller effect than in text related task: this is due to the different nature of the text datasets (large "continuous" corpuses) and the IEMOCAP one (composed of shorter utterances). The fact that we can still observe improvements on a data... |
Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes | 1909.08700 | Table 6: Token order imbalance (TOI) comparison for the IEMOCAP dataset on a SER task using angry, happy, neutral and sad classes for 60 epochs using the Transformer model. | ['Experiment', 'WA (60 epochs)', 'UA (60 epochs)'] | [['Alleviated TOI 1', '0.591±0.012', '0.543±0.021'], ['Alleviated TOI 2', '0.594±0.007', '0.549±0.016'], ['Alleviated TOI 3', '0.605±0.018', '0.563±0.024'], ['Alleviated TOI 5', '0.608±0.015', '0.562±0.028'], ['Alleviated TOI 10', '[BOLD] 0.617±0.015', '[BOLD] 0.571±0.024'], ['Local attention', '[BOLD] 0.635', '[BOLD] ... | The previously described Transformer model is used in these experiments. the standard deviation computed for different P-values of Alleviated TOI (P). We want to highlight the fact that the goal of these experiments is to show the direct contribution of the Alleviated TOI technique for a different model. For this reaso... |
BP-Transformer: Modelling Long-Range Context via Binary Partitioning | 1911.04070 | Table 5: BLEU score on IWSLT 2015 Zh-En | ['Model', 'BLEU'] | [['Transformer vaswani2017attention', '16.87'], ['Transformer+cache tu2018learning', '17.32'], ['HAN-NMT miculicich-etal-2018-document', '17.78'], ['Transformer (ours, single sentence)', '18.91'], ['BPT (k=4, single sentence)', '19.19'], ['BPT (k=4,l=64)', '19.84']] | BPT with k=4 and context length of 32 could further improve the baseline result by 0.93 in terms of BLEU score, which is a significant margin. |
BP-Transformer: Modelling Long-Range Context via Binary Partitioning | 1911.04070 | Table 1: Test accuracy on SST-5 and IMDB. In BPT, k=2 and k=4 for SST and IMDB respectively. The last model used word embeddings pretrained with translation and additional character-level embeddings. | ['Model', 'SST-5', 'IMDB'] | [['[BOLD] BPT', '52.71(0.32)', '[BOLD] 92.12(0.11)'], ['Star Transformer', '52.9', '90.50'], ['Transformer', '50.4', '89.24'], ['Bi-LSTM (li2015tree)', '49.8', '-'], ['Tree-LSTM (socher2013recursive)', '51.0', '-'], ['QRNN DBLP:conf/iclr/0002MXS17', '-', '91.4'], ['BCN+Char+CoVe mccann2017learned', '53.7', '91.8']] | On SST-5, our model outperforms Transformer and LSTM based models. On IMDB, our proposed model outperforms a bidirectional LSTM initialized with pre-trained character embedding and CoVe embedding mccann2017learned. |
BP-Transformer: Modelling Long-Range Context via Binary Partitioning | 1911.04070 | Table 6: BLEU score vs context length on different models | ['Context length', '0', '32', '64', '128'] | [['Transformer', '18.85', '18.66', '17.59', '15.55'], ['BPT (k=4)', '19.19', '19.84', '19.71', '19.84'], ['BPT (k=8)', '19.13', '19.59', '19.78', '19.60']] | Similar to tu2018learning and miculicich-etal-2018-document , we found a small context length is enough for achieving good performance on IWSLT for Document-Level Translation. However, as we increases context size, the performance of BPT does not get worse as these models and Transformers, suggesting the inductive bias... |
BP-Transformer: Modelling Long-Range Context via Binary Partitioning | 1911.04070 | Table 7: BLEU score on newstest 2014 | ['Model', 'BLEU'] | [['ByteNet\xa0kalchbrenner2016neural', '23.75'], ['GNMT+RL\xa0wu2016google', '24.6'], ['ConvS2S\xa0GehringAGYD17', '25.16'], ['Transformer\xa0vaswani2017attention', '27.3'], ['Transformer (our implementation)', '27.2'], ['BPT (k=1)', '26.9'], ['BPT (k=2)', '27.4'], ['BPT (k=4)', '[BOLD] 27.6'], ['BPT (k=8)', '26.7']] | In the setting of k=2 and k=4, BPT outperforms Vanilla Transformer with the same number of parameters and a sparse attention pattern. |
Toward Computation and Memory Efficient Neural Network Acoustic Models with Binary Weights and Activations | 1706.09453 | Table 1: WERs (%) of binary weight networks on WSJ1. The number of hidden units is 1024 for experiments in this table. | ['Model', 'Input', 'Softmax', 'dev93', 'eval92'] | [['Baseline', '—', '—', '6.8', '3.8'], ['Binary weights ( [ITALIC] p=0)', 'fixed', 'fixed', '7.7', '4.8'], ['Binary weights ( [ITALIC] p=.001)', 'fixed', 'fixed', '8.0', '4.5'], ['Binary weights ( [ITALIC] p=.01)', 'fixed', 'fixed', '8.0', '4.4'], ['Binary weights ( [ITALIC] p=0)', 'real', 'real', '10.4', '6.7'], ['Bin... | In our experiments, training neural networks with binary weights from random initialization usually did not converge, or converged to very poor models. We addressed this problem by initializing our models from a well-trained neural network model with real-valued weights. This approach worked well, and was used in all o... |
Toward Computation and Memory Efficient Neural Network Acoustic Models with Binary Weights and Activations | 1706.09453 | Table 3: WERs (%) of neural networks with binary weights and activations on the WSJ1 dataset. We set p=0 for the system with binary weights, and k=1 for the system with binary activations. | ['Model', 'Size', '[ITALIC] b', 'dev93', 'eval92'] | [['Baseline', '1024', '–', '6.8', '3.8'], ['Baseline', '2048', '–', '6.5', '3.5'], ['Binary weights', '1024', '(−1,+1)', '7.7', '4.8'], ['Binary weights', '1024', '(0,1)', 'Not Converged', 'Not Converged'], ['Binary activations', '1024', '(−1,+1)', '8.2', '4.4'], ['Binary activations', '1024', '(0,1)', '7.2', '4.1'], [... | For all the experiments in this table, the weights in the input and Softmax layer were fixed, and the first and last hidden layers used Sigmoid activations. Using a larger number of hidden units works slightly better for both binary weight and binary activation systems. However, we only managed to train the network wit... |
Toward Computation and Memory Efficient Neural Network Acoustic Models with Binary Weights and Activations | 1706.09453 | Table 4: WERs (%) of neural network with binary weights and activations on the AMI dataset. The number of hidden units is 2048, and b denotes binarization. | ['Model', '[ITALIC] b', 'dev', 'eval'] | [['Baseline', '–', '26.1', '27.5'], ['Binary weights ( [ITALIC] p=0)', '(−1,+1)', '30.3', '32.7'], ['Binary weights ( [ITALIC] p=.001)', '(−1,+1)', '30.0', '32.2'], ['Binary weights ( [ITALIC] p=.01)', '(−1,+1)', '29.6', '31.7'], ['Binary weights ( [ITALIC] p=.05)', '(−1,+1)', '29.6', '31.9'], ['Binary activations ( [I... | For the binary weight systems, we revisited the semi-stochastic binarization approach. While the convergence curves were similar Since the model was initialized from a Sigmoid network, the binary activation system with (0,1) worked much better than its counterpart with (−1,+1) when k=1. Again, we did experiments by tun... |
Generative Adversarial Zero-Shot Relational Learning for Knowledge Graphs | 2001.02332 | Table 3: Quantitative analysis of the generated relation embeddings by our generator. These presented relations are from the NELL test relation set. “# Can. Num.” denotes the number of candidates of test relations. For one relation, “# Cos. Sim.” denotes the mean cosine similarity between the corresponding generated re... | ['Relations', '# Can. Num.', '# Cos. Sim.', 'MRR ZSGAN [ITALIC] KG', 'MRR ZS-DistMult', 'Hits@10 ZSGAN [ITALIC] KG', 'Hits@10 ZS-DistMult'] | [['animalThatFeedOnInsect', '293', '0.8580', '[BOLD] 0.347', '0.302', '[BOLD] 63.4', '61.8'], ['automobileMakerDealersInState', '600', '0.1714', '[BOLD] 0.066', '0.039', '[BOLD] 15.4', '5.1'], ['animalSuchAsInvertebrate', '786', '0.7716', '[BOLD] 0.419', '0.401', '[BOLD] 59.8', '57.6'], ['sportFansInCountry', '2100', '... | Unlike image, our generated data cannot be observed intuitively. Instead, we calculate the cosine similarity between the generated relation embeddings and the cluster center xrc of their corresponding relations. It can be seen that our method indeed generates the plausible relation embeddings for many relations and the... |
Generative Adversarial Zero-Shot Relational Learning for Knowledge Graphs | 2001.02332 | Table 2: Zero-shot link prediction results on the unseen relations. The proposed baselines are shown at the top of the table; Our generative adversarial model is denoted as ZSGANKG and the results are shown at the bottom. Bold numbers denote the best results, and Underline numbers denote the best ones among our ZSGANKG... | ['Model', '[BOLD] NELL-ZS MRR', '[BOLD] NELL-ZS Hits@10', '[BOLD] NELL-ZS Hits@5', '[BOLD] NELL-ZS Hits@1', '[BOLD] Wiki-ZS MRR', '[BOLD] Wiki-ZS Hits@10', '[BOLD] Wiki-ZS Hits@5', '[BOLD] Wiki-ZS Hits@1'] | [['ZS-TransE', '0.097', '20.3', '14.7', '4.3', '0.053', '11.9', '8.1', '1.8'], ['ZS-DistMult', '0.235', '32.6', '28.4', '18.5', '0.189', '23.6', '21.0', '16.1'], ['ZS-ComplEx', '0.216', '31.6', '26.7', '16.0', '0.118', '18.0', '14.4', '8.3'], ['ZSGAN [ITALIC] KG (TransE)', '0.240', '[BOLD] 37.6', '[BOLD] 31.6', '17.1',... | For NELL-ZS dataset, we set the embedding size as 100. For Wiki-ZS, we set the embedding size as 50 for faster training. The proposed generative method uses the pre-trained KG embeddings as input, which are trained on the triples in the training set. For TransE and DistMult, we directly use their 1-D vectors. For featu... |
Cross-lingual Word Analogies using Linear Transformations between Semantic Spaces | 1807.04175 | (i) verb-past-tense | ['[EMPTY]', 'En', 'De', 'Es', 'It', 'Cs', 'Hr'] | [['En', '42.2', '13.1', '18.6', '29.7', '58.0', '41.8'], ['De', '45.1', '33.4', '20.3', '35.1', '69.0', '46.8'], ['Es', '27.0', '0.8', '31.8', '48.6', '35.2', '22.4'], ['It', '32.2', '1.4', '39.3', '48.7', '48.8', '29.8'], ['Cs', '42.5', '13.1', '22.4', '40.0', '80.8', '55.6'], ['Hr', '43.6', '8.7', '33.1', '48.9', '73... | Again, rows represent the source language a and columns the target language b. The results were achieved using B-CCA-cu transformation with dictionaries of size n=20,000. Each language seems to have strengths and weaknesses. |
Cross-lingual Word Analogies using Linear Transformations between Semantic Spaces | 1807.04175 | Table 1: Number of word pairs for each language and each analogy type. | ['Semantic', 'family', 'En 24', 'De 24', 'Es 20', 'It 20', 'Cs 26', 'Hr 41'] | [['Semantic', 'state-currency', '29', '29', '28', '29', '29', '21'], ['Semantic', 'capital-common-countries', '23', '23', '21', '23', '23', '23'], ['Syntactic', 'state-adjective', '41', '41', '40', '41', '41', '41'], ['Syntactic', 'adjective-comparative', '23', '37', '5', '10', '40', '77'], ['Syntactic', 'adjective-sup... | We combine and extend available corpora for monolingual word analogies in English (En ) Mikolov et al. (It) Berardi et al. Czech (Cs) Svoboda and Brychcín Svoboda and Beliga We consider only those analogy types, which exist across all six languages (three semantically oriented and six syntactically oriented analogy typ... |
Cross-lingual Word Analogies using Linear Transformations between Semantic Spaces | 1807.04175 | Table 2: The average accuracies across all combinations of language pairs for different linear transformations and post-processing techniques. The size of bilingual dictionary was set to n=20,000. No trans. denotes the monolingual experiments without transforming the spaces. | ['[EMPTY]', '[EMPTY]', '[BOLD] - [BOLD] Acc@1', '[BOLD] - [BOLD] Acc@5', '[BOLD] -c [BOLD] Acc@1', '[BOLD] -c [BOLD] Acc@5', '[BOLD] -u [BOLD] Acc@1', '[BOLD] -u [BOLD] Acc@5', '[BOLD] -cu [BOLD] Acc@1', '[BOLD] -cu [BOLD] Acc@5'] | [['Monoling.', 'No trans.', '49.6', '63.7', '50.1', '64.6', '50.6', '64.6', '[BOLD] 51.1', '[BOLD] 65.2'], ['Monoling.', 'M-LS', '40.2', '55.3', '40.3', '55.6', '41.3', '56.5', '41.3', '56.6'], ['Monoling.', 'M-OT', '49.6', '63.7', '50.1', '64.6', '50.6', '64.6', '[BOLD] 51.1', '[BOLD] 65.2'], ['Monoling.', 'M-CCA', '4... | The columns represent different post-processing techniques and rows different transformations. The upper part of the table shows the monolingual experiments with original spaces without transformation (No trans.) compared with the unified multilingual space for all six languages. The orthogonal transformation provides ... |
Cross-lingual Word Analogies using Linear Transformations between Semantic Spaces | 1807.04175 | (a) family | ['[EMPTY]', 'En', 'De', 'Es', 'It', 'Cs', 'Hr'] | [['En', '68.8', '52.4', '85.4', '76.0', '41.2', '47.2'], ['De', '65.5', '48.0', '76.3', '66.5', '35.9', '40.9'], ['Es', '70.6', '49.0', '86.8', '74.5', '43.1', '45.2'], ['It', '65.4', '45.6', '81.8', '72.9', '39.2', '45.2'], ['Cs', '61.5', '38.6', '74.0', '65.0', '35.6', '42.0'], ['Hr', '57.4', '33.3', '62.8', '60.0', ... | Again, rows represent the source language a and columns the target language b. The results were achieved using B-CCA-cu transformation with dictionaries of size n=20,000. Each language seems to have strengths and weaknesses. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.