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 |
|---|---|---|---|---|---|
From Word Embeddings to Item Recommendation | 1601.01356 | TABLE I: Comparison of methods | ['Method', 'Precision', 'Ndcg', 'HitRate', 'Coverage'] | [['KNI(skip-gram)', '0.119', '0.169', '0.618', '1.000'], ['NN(skip-gram)', '0.070', '0.117', '0.450', '1.000'], ['KIU(skip-gram)', '0.112', '0.161', '0.599', '1.000'], ['KNI(CBOW)', '0.112', '0.165', '0.586', '1.000'], ['NN(CBOW)', '0.062', '0.100', '0.401', '1.000'], ['KIU(CBOW)', '0.102', '0.150', '0.555', '1.000'], ... | The first method from the recommendation literature is traditional collaborative filtering method (CF-C) which uses the past preferences of the users and their similarities. We implemented this method to observe its performance. The second method is Random, which randomly chooses k-many items (venues) to recommend. We ... |
The Natural Stories Corpus | 1708.05763 | Table 3: Regression coefficients and significance from individual mixed-effects regressions predicting RT for each of the three predictors log frequency, log trigram probability, and word length. | ['Predictor', '^ [ITALIC] β', 'Std. Error', '[ITALIC] t value', '[ITALIC] p value'] | [['Log Frequency', '-2.61', '0.08', '-32.27', '< 0.001'], ['Trigram Surprisal', '-2.19', '0.09', '-23.90', '< 0.001'], ['Word Length', '4.21', '0.12', '35.72', '< 0.001']] | In order to validate our RT data, we checked that basic psycholinguistic effects obtain in it. In particular, we examined whether the well-known effects of frequency, word length, and surprisal To do this, for each of the three predictors log frequency, log trigram probability, and word length, we fit a linear mixed ef... |
Assessing the Stylistic Properties of Neurally Generated Text in Authorship Attribution | 1708.05536 | Table 1: Mean F1, Precision (P) and Recall (R) scores for all classification experiments. | ['Source', 'Experiment', 'F1', 'P', 'R'] | [['Real', '< [ITALIC] α, [ITALIC] ω>', '0.833', '0.818', '0.869'], ['[EMPTY]', '< [ITALIC] ω, [ITALIC] α>', '0.811', '0.795', '0.853'], ['NGLM', '< [ITALIC] α+¯ [ITALIC] α, [ITALIC] ω>', '0.814', '0.809', '0.850'], ['[EMPTY]', ', [ITALIC] ω>', '0.706', '0.744', '0.750'], ['[EMPTY]', '< [ITALIC] ω,¯ [ITALIC] α>', '0.837... | We focus on the macro-averaged F1-scores in our discussion, although one should not forget that the scores vary considerably over individual authors (cf. Fig. With respect to the authentic data, classifying α on the basis of ω is slightly more difficult than the reverse direction, which seems a negligible directionalit... |
Labelled network motifs reveal stylistic subtleties in written texts | 1705.00545 | Table 2: Accuracy rate (%) in discriminating the authorship of books in Dataset 1. The best result obtained with the proposed technique surpasses by 15 percentage points the best performance obtained with traditional features based on the frequency of function words. | ['[BOLD] Features', '| [ITALIC] W|', '[BOLD] J48', '[BOLD] kNN', '[BOLD] SVM', '[BOLD] Bayes'] | [['LMV1', '5', '45.0', '65.0', '62.5', '30.0'], ['LMV1', '10', '37.5', '60.0', '67.5', '27.5'], ['LMV1', '20', '60.0', '65.0', '[BOLD] 75.0', '25.0'], ['LMV2', '5', '55.0', '50 .0', '62.5', '22.5'], ['LMV2', '10', '47.5', '65.0', '77.5', '15.0'], ['LMV2', '20', '45.0', '60.0', '[BOLD] 80.0', '25.0'], ['MFW', '5', '30.0... | The best results were obtained with the SVM, in general. For this reason, the discussion here is focused on the results obtained by this classifier. We note that, when comparing both versions of the proposed technique for the same |W|, the second version yielded best results, which reinforces the importance of function... |
Labelled network motifs reveal stylistic subtleties in written texts | 1705.00545 | Table 3: Accuracy rate (%) in discriminating the authorship of books in Dataset 2. The best performance was obtained when the proposed technique LMV2 was used as attribute to train the SVM classifier. | ['[BOLD] Features', '| [ITALIC] W|', '[BOLD] J48', '[BOLD] kNN', '[BOLD] SVM', '[BOLD] Bayes'] | [['LMV1', '5', '58.7', '65.1', '74.3', '69.2'], ['LMV1', '10', '61.7', '83.7', '91.6', '81.3'], ['LMV1', '20', '66.8', '88.3', '[BOLD] 95.4', '78.7'], ['LMV2', '5', '62.1', '67.4', '82.0', '69.1'], ['LMV2', '10', '65.5', '80.9', '91.6', '75.2'], ['LMV2', '20', '68.1', '88.0', '[BOLD] 96.0', '77.5'], ['MFW', '5', '58.3'... | The authorship attribution task was also evaluated using a different dataset comprising books from 9 authors, henceforth referred to as Dataset 2. The goal of this second experiment was to evaluate the performance of Labelled motifs in characterizing shorter pieces of text. In this dataset, each book was split in sever... |
Labelled network motifs reveal stylistic subtleties in written texts | 1705.00545 | Table 4: Accuracy rate (%) in discriminating the debates from the Canadian Hansard into two classes (Original and Translated). The highest accuracies were obtained with the strategy based on labelled motifs. | ['[BOLD] Language', '[BOLD] Features', '| [ITALIC] W|', '[BOLD] J48', '[BOLD] kNN', '[BOLD] SVM', '[BOLD] Bayes'] | [['[BOLD] English', '[EMPTY]', '[EMPTY]', '[EMPTY]', '[EMPTY]', '[EMPTY]', '[EMPTY]'], ['[EMPTY]', 'LMV1', '20', '90.3', '89.2', '96.5', '90.3'], ['[EMPTY]', 'LMV2', '20', '90.4', '93.5', '[BOLD] 97.5', '88.0'], ['[EMPTY]', 'MFW', '5', '71.8', '75.0', '57.7', '52.7'], ['[EMPTY]', 'MFW', '10', '74.4', '75.6', '60.4', '5... | The accuracies are relatively high for such a simple feature set. The results suggest that labelled motifs are extracting information about French to English (and vice-versa) translation and these features lead to accuracies higher than the ones obtained with only the frequency of the most frequent words. |
Labelled network motifs reveal stylistic subtleties in written texts | 1705.00545 | Table 5: Accuracy rate (%) in discriminating the debates from the European Parliament into two classes (Original and Translated). | ['[BOLD] Language [BOLD] English', '[BOLD] Features', '| [ITALIC] W|', '[BOLD] SVM', '[BOLD] Language [BOLD] Italian', '[BOLD] Features', '| [ITALIC] W|', '[BOLD] SVM'] | [['[EMPTY]', 'LMV1', '20', '90.2', '[EMPTY]', 'LMV1', '20', '93.1'], ['[EMPTY]', 'LMV2', '20', '[BOLD] 92.3', '[EMPTY]', 'LMV2', '20', '[BOLD] 95.6'], ['[EMPTY]', 'MFW', '5', '68.4', '[EMPTY]', 'MFW', '5', '85.2'], ['[EMPTY]', 'MFW', '10', '68.3', '[EMPTY]', 'MFW', '10', '90.1'], ['[EMPTY]', 'MFW', '20', '[BOLD] 78.3',... | To simplify our analysis, we just present the results for the classifier with the best accuracies. They achieved an accuracy of 96.7% using the frequency of 300 function words. However, they did not detect translationese with target languages other than English. Once again, we have found that the characterization by la... |
LANDMARK-BASED CONSONANT VOICING DETECTION ON MULTILINGUAL CORPORA | 1611.03533 | Table 4: relative error rate increment (%) on other languages. | ['[EMPTY]', 'Turkish', 'Spanish'] | [['MFCC(13)(whole utterance)', '59.3', '54.8'], ['MFCC(13)(Landmark Region)', '40.9', '44.5'], ['MFCC(39)(whole utterance)', '59.1', '96.4'], ['MFCC(39)(Landmark Region)', '50.0', '91.7'], ['Acoustic cues', '31.3', '35.2'], ['CNN', '16.2', '19.0']] | CNNs vs. Acoustic Cues: Since the evaluation on two models (CNN+FFT and CNN+FB) results in similar accuracy scores, we consider CNN+FFT as the best CNN model. The last columns in Figs. |
Modeling Label Semantics for Predicting Emotional Reactions | 2006.05489 | Table 1: Comparison Results on ROCStories with Plutchik emotion labels | ['[EMPTY]', '[BOLD] Model', '[BOLD] Precision', '[BOLD] Recall', '[BOLD] F1'] | [['[EMPTY]', 'Rashkin et\xa0al. ( 2018 )', '[EMPTY]', '[EMPTY]', '[EMPTY]'], ['[EMPTY]', 'BiLSTM', '25.31', '33.44', '28.81'], ['[EMPTY]', 'CNN', '24.47', '38.87', '30.04'], ['Baselines', 'REN', '25.30', '37.30', '30.15'], ['[EMPTY]', 'NPN', '24.33', '40.10', '30.29'], ['[EMPTY]', 'Paul and Frank ( 2019 )*', '59.66', '... | Among the baselines, the fine-tuned BERT base model obtains the best results. Adding label embeddings (section 3.1) to the basic BiLSTM via LEAM model provides substantial increase, more than 27 absolute points in F1. We swapped in BERT features instead of GloVe and found a further 3 point improvement. The BERT baselin... |
Semi-Autoregressive Training Improves Mask-Predict Decoding | 2001.08785 | Table 4: Increasing the number of forward passes used to produce each training example in SMART can negatively effect the resulting model. Performance measured on WMT’14 DE-EN (development set BLEU). | ['[BOLD] Forward Passes', '[BOLD] Decoding Iterations [BOLD] 1', '[BOLD] Decoding Iterations [BOLD] 4', '[BOLD] Decoding Iterations [BOLD] 10'] | [['2', '[BOLD] 24.24', '[BOLD] 29.98', '[BOLD] 30.37'], ['3', '23.74', '29.89', '30.28'], ['4', '23.81', '39.66', '30.10']] | We first consider creating our training examples by performing multiple mask-predict iterations during training, instead of just two. worse results. |
Semi-Autoregressive Training Improves Mask-Predict Decoding | 2001.08785 | Table 1: The performance (test set BLEU) of semi-autoregressive training (SMART), compared to the original non-autoregressive training for CMLMs (NART). All models are decoded with mask-predict. | ['[BOLD] Training Mode', '[BOLD] Decoding [BOLD] Iterations', '[BOLD] WMT’14 [BOLD] EN-DE', '[BOLD] WMT’14 [BOLD] DE-EN', '[BOLD] WMT’17 [BOLD] EN-ZH', '[BOLD] WMT’17 [BOLD] ZH-EN'] | [['[ITALIC] NART', '1', '18.05', '21.83', '[BOLD] 24.23', '[BOLD] 13.64'], ['[ITALIC] SMART', '1', '[BOLD] 18.58', '[BOLD] 23.77', '24.15', '13.51'], ['[ITALIC] NART', '4', '25.94', '29.90', '32.63', '21.90'], ['[ITALIC] SMART', '4', '[BOLD] 27.03', '[BOLD] 30.87', '[BOLD] 33.37', '[BOLD] 22.61'], ['[ITALIC] NART', '10... | We first compare SMART to the original CMLM training process (NART). Even with a single decoding iteration (the purely non-autoregressive scenario), SMART produces better models in WMT’14 and falls short of the baseline by a slim margin in WMT’17 (0.08 and 0.13 BLEU). |
Semi-Autoregressive Training Improves Mask-Predict Decoding | 2001.08785 | Table 2: The performance (test set BLEU) of semi-autoregressive training (SMART), compared to the standard (sequential) transformer. Length beam, beam size and length penalty is tuned for each model on validation set. | ['[BOLD] Model', '[BOLD] Decoding [BOLD] Iterations', '[BOLD] WMT’14 [BOLD] EN-DE', '[BOLD] WMT’14 [BOLD] DE-EN', '[BOLD] WMT’17 [BOLD] EN-ZH', '[BOLD] WMT’17 [BOLD] ZH-EN'] | [['[ITALIC] Autoregressive Transformer with Beam Search', '[ITALIC] N', '27.61', '31.38', '34.31', '23.65'], ['[ITALIC] + Knowledge Distillation', '[ITALIC] N', '[BOLD] 27.75', '31.30', '[BOLD] 34.38', '23.91'], ['[ITALIC] SMART CMLM with Mask-Predict', '10', '27.65', '31.27', '34.06', '23.78'], ['[ITALIC] SMART CMLM ... | We also compare between SMART-trained CMLMs with mask-predict decoding and autoregressive transformers with beam search. With the exception of English to Chinese, the performance differences are within the typical random seed variance. Increasing the number of mask-predict iterations to N yields even more balanced resu... |
Semi-Autoregressive Training Improves Mask-Predict Decoding | 2001.08785 | Table 3: The performance (development set BLEU) of SMART-trained models with two flavors of mask-predict: predicting only masked tokens (original version), and predicting all tokens at each iteration. | ['[BOLD] Predicted Tokens', '[BOLD] Decoding [BOLD] Iterations', '[BOLD] WMT’14 [BOLD] EN-DE', '[BOLD] WMT’14 [BOLD] DE-EN'] | [['[ITALIC] Masked Tokens', '4', '25.18', '29.61'], ['[ITALIC] All Tokens', '4', '[BOLD] 25.61', '[BOLD] 29.98'], ['[ITALIC] Masked Tokens', '10', '26.06', '30.29'], ['[ITALIC] All Tokens', '10', '[BOLD] 26.14', '[BOLD] 30.37']] | Repredicting All Tokens Besides SMART, we also augment the mask-predict algorithm to predict all tokens – not only the masked ones – during the predict step We find that predicting all tokens increases performance by 0.40 BLEU on average when using 4 decoding iterations. With 10 decoding iterations, the gains shrink to... |
Card-660: Cambridge Rare Word Dataset – a Reliable Benchmark for Infrequent Word Representation Models | 1808.09308 | Table 4: Pearson r and Spearman ρ correlation percentage performance of mainstream pre-trained word embeddings on the RW and Card-660 datasets. Column |V| shows the size of vocabulary for the corresponding embedding set. | ['[BOLD] Embedding set', '| [ITALIC] V|', '[BOLD] Missed words RW', '[BOLD] Missed words Card', '[BOLD] Missed pairs RW', '[BOLD] Missed pairs Card', '[BOLD] Pearson [ITALIC] r RW', '[BOLD] Pearson [ITALIC] r Card', '[BOLD] Spearman [ITALIC] ρ RW', '[BOLD] Spearman [ITALIC] ρ Card'] | [['Glove Wikipedia-Gigaword (300d)', '400K', '7%', '55%', '12%', '74%', '34.9', '15.1', '34.4', '15.7'], ['Glove Common Crawl - uncased (300d)', '1.9M', '1%', '36%', '1%', '50%', '36.5', '29.2', '37.7', '27.6'], ['Glove Common Crawl - cased (300d)', '2.2M', '1%', '29%', '2%', '44%', '44.0', '33.0', '45.1', '27.3'], ['G... | Pennington et al. Speer et al. In the last two rows of the Table we also report results for two hybrid embeddings constructed by combining the pre-trained Freebase Word2vec, which mostly comprises named entities, with two of the best performing embeddings evaluated on the dataset. Given that the word embeddings are not... |
Orthogonal Relation Transforms with Graph Context Modeling for Knowledge Graph Embedding | 1911.04910 | Table 3: Ablation study on FB15k-237 validation set. | ['Model', '[ITALIC] ds', 'MRR', '@10', '#param'] | [['RotatE-S', '-', '.330', '.515', '5.9'], ['RotatE-L', '-', '.340', '.530', '29.3'], ['OTE', '2', '.327', '.511', '6.1'], ['OTE', '20', '.355', '.540', '7.8'], ['OTE - scalar', '20', '.352', '.535', '7.7'], ['LNE', '20', '.354', '.538', '9.6'], ['GC-RotatE-L', '-', '.354', '.546', '29.3'], ['GC-OTE', '20', '.367', '.5... | Orthogonal transform has many desired properties, for example, the inverse matrix is obtained by simply transposing itself. It also preserves the L2 norm of a vector after the transform. For our work, we are just interested in its property to obtain inverse matrix by simple transposing. We perform the ablation study wi... |
MISA: Modality-Invariant and -Specific Representations for Multimodal Sentiment Analysis | 2005.03545 | Table 6. Final hyper-parameter values in each dataset. | ['Hyper-param', 'MOSI', 'MOSEI', 'UR_FUNNY'] | [['cmd K', '5', '5', '5'], ['activation', 'ReLU', 'LeakyReLU', 'Tanh'], ['batch size', '64', '16', '32'], ['gradient clip', '1.0', '1.0', '1.0'], ['[ITALIC] α', '1.0', '0.7', '0.7'], ['[ITALIC] β', '0.3', '0.3', '1.0'], ['[ITALIC] γ', '1.0', '0.7', '1.0'], ['dropout', '0.5', '0.1', '0.1'], ['[ITALIC] dh', '128', '128',... | We look over finite sets of options for hyper-parameters. These include non-linear activations: Finally, we look at dropout values from {0.1,0.5,0.7}. For optimization, we utilize the Adam optimizer with an exponential decay learning rate scheduler. The training duration of each model is governed by early-stopping stra... |
MISA: Modality-Invariant and -Specific Representations for Multimodal Sentiment Analysis | 2005.03545 | Table 4. Ablation Study. Here, (−) represents removal for the mentioned factors. Model 1 represents the best performing model in each dataset; Model 2,3,4 depicts the effect of individual modalities; Model 5,6,7 presents the effect of regularization; Model 8,9,10,11 presents the variants of MISA as defined in Section 5... | ['[EMPTY]', 'Model', 'MOSI MAE (↓)', 'MOSI Corr (↑)', 'MOSEI MAE (↓)', 'MOSEI Corr (↑)', 'UR_FUNNY Acc-2 (↑)'] | [['1)', 'MISA', '[BOLD] 0.783', '[BOLD] 0.761', '[BOLD] 0.555', '[BOLD] 0.756', '[BOLD] 70.6'], ['2)', '(-) language [ITALIC] l', '1.450', '0.041', '0.801', '0.090', '55.5'], ['3)', '(-) visual [ITALIC] v', '0.798', '0.756', '0.558', '0.753', '69.7'], ['4)', '(-) audio [ITALIC] a', '0.849', '0.732', '0.562', '0.753'... | we remove one modality at a time to observe the effect in performance. Firstly, it is seen that multimodal combination provides the best performance, which indicates that the model is able to learn complementary features. Without this case, the tri-modal combination would not fare better than bi-modal variants such as ... |
Parallax: Visualizing and Understanding the Semantics of Embedding Spaces via Algebraic Formulae | 1905.12099 | Table 1: Two-way ANOVA analyses of Task (Commonality vs. Polarization) and Obfuscation (Obfuscated vs. Non-obfuscated) over Projection (Explicit Formulae vs. t-SNE). | ['Accuracy', 'Factor', '[ITALIC] F(1,91)', 'p-value'] | [['Projection', 'Projection', '46.11', '0.000∗∗∗'], ['×', 'Task', '1.709', '0.194'], ['Task', 'Projection × Task', '3.452', '0.066'], ['Projection', 'Projection', '57.73', '0.000∗∗∗'], ['×', 'Obfuscation', '23.93', '0.000∗∗∗'], ['Obfuscation', 'Projection × Obf', '5.731', '0.019∗']] | that the proposed Explicit Formulae method outperforms t-SNE in terms of accuracy in both Commonality and Polarization tasks. We also observed significant differences in Obfuscation: subjects tend to have better accuracy when the words are not obfuscated. We run post-hoc t-tests that confirmed how the accuracy of Expli... |
Modeling Coherency in Generated Emails by Leveraging Deep Neural Learners | 2007.07403 | Table 4: Human evaluation results on the automated and true emails. Scores for Syntax (Syn), Coherency (Coh) and Fluency (Flu) vary between [1, 5]. Detection Rate has been reported as a percentage | ['[BOLD] Email Content', '[BOLD] Scores [BOLD] Syn', '[BOLD] Scores [BOLD] Coh', '[BOLD] Scores [BOLD] Flu', '[BOLD] Detection Rate', '¯¯¯ [BOLD] W', '[ITALIC] SDW', '¯¯¯ [BOLD] S', '[ITALIC] SDS'] | [['Generated [ITALIC] word', '3.63', '2.8', '2.6', '83.33', '12', '2.88', '1.2', '0.45'], ['Generated [ITALIC] sentence', '4.29', '3.67', '3.62', '38.09', '13', '6.21', '2', '0.81'], ['Generated [ITALIC] all', '4.01', '3.31', '3.19', '56.94', '12', '5', '2', '0.78'], ['Truth', '4.42', '4.56', '4.47', '75', '24', '14.8'... | We report the scores for syntax, coherency and fluency, averaged across all participants, on the combined set of system generated emails - Generatedall. We also report the same on the subset of emails generated by each of the baseline and deep generation models - Generatorword and Generatedsentence respectively, to com... |
Trace norm regularization and faster inference for embedded speech recognition RNNs | 1710.09026 | Table 3: Performance of completely split versus partially joint factorization of recurrent weights. | ['SVD threshold', 'Completely split Parameters (M)', 'Completely split CER', 'Partially joint Parameters (M)', 'Partially joint CER'] | [['0.50', '6.3', '10.3', '5.5', '10.3'], ['0.60', '8.7', '10.5', '7.5', '10.2'], ['0.70', '12.0', '10.3', '10.2', '9.9'], ['0.80', '16.4', '10.1', '13.7', '9.7']] | Finally, we compared the partially joint factorization to the completely split factorization and found that the former indeed led to better accuracy versus number of parameters trade-offs. |
Auxiliary Interference Speaker Loss for Target-Speaker Speech Recognition | 1906.10876 | Table 1: WERs (%) for the two-speaker mixed LibriSpeech corpus in various SIR conditions. Note that for clean single-speaker speech, a clean AM achieved WERs of 4.88% and 5.54% for dev-clean and test-clean, respectively. | ['SIR of the targeted speaker’s speech', 'dev-clean (two spkeakers mixed) 10', 'dev-clean (two spkeakers mixed) 5', 'dev-clean (two spkeakers mixed) 0', 'dev-clean (two spkeakers mixed) -5', 'dev-clean (two spkeakers mixed) -10', 'dev-clean (two spkeakers mixed) Avg.', 'test-clean (two spkeakers mixed) 10', 'test-clean... | [['Clean AM', '65.14', '78.72', '88.56', '91.96', '94.31', '83.74', '66.92', '79.45', '89.18', '92.87', '95.15', '84.71'], ['Target-Speaker AM w/o aux. loss', '13.98', '15.02', '16.80', '18.42', '21.54', '17.15', '15.59', '16.13', '17.65', '19.17', '21.78', '18.06'], ['Target-Speaker AM w/ aux. loss', '[BOLD] 13.51', '... | Note that WERs for clean speech with the clean AM was 4.88% and 5.54% for dev-clean and test-clean, respectively. As shown in the table, WERs were severely degraded by mixing the two speakers’ speech, and the clean AM produced WER averages of 83.74% and 84.71% for dev-clean and test-clean, respectively. This model dram... |
Auxiliary Interference Speaker Loss for Target-Speaker Speech Recognition | 1906.10876 | Table 2: WERs (%) for two-speaker-mixed test-clean. Main output branch was used for the target speaker’s ASR and auxiliary output branch was used for the interference speaker’s ASR. | ['SIR of target spk.', 'SIR of interference spk.', 'WER (%) target spk.', 'WER (%) interference spk.'] | [['10', '-10', '14.59', '26.22'], ['5', '-5', '15.06', '19.90'], ['0', '0', '16.46', '16.23'], ['-5', '5', '17.86', '15.05'], ['-10', '10', '20.38', '14.50'], ['Avg.', 'Avg.', '16.87', '18.38']] | Therefore, we evaluated the ability of an auxiliary network for the secondary ASR. In this evaluation, we provided the target speaker’s embeddings for the network, and evaluated the WERs of the ASR results using the output of the auxiliary output branch. From this table, we found that the auxiliary output branch worked... |
Auxiliary Interference Speaker Loss for Target-Speaker Speech Recognition | 1906.10876 | Table 3: WERs (%) for WSJ corpus with clean AM and target-speaker (TS) AMs. | ['Model', 'Dev93', 'Eval92', 'wsj0-2mix'] | [['Clean AM', '77.51', '78.03', '79.81'], ['TS-AM w/o aux. loss', '12.24', '[BOLD] 11.32', '16.78'], ['TS-AM w/ aux. loss', '[BOLD] 11.31', '11.38', '[BOLD] 16.50']] | We observed a significant improvement for Dev93 and moderate improvement for wsj0-2mix, while a marginal degradation of WER was observed for EVAL92. Note that the results by our AM were fairly good thanks to the state-of-the-art accuracy given by the LF-MMI. |
Fluency-Guided Cross-Lingual Image Captioning | 1708.04390 | Table 5. Automated evaluation of six approaches to cross-lingual image captioning. Rejection sampling and weighted loss are comparable to ‘Without fluency’ which learns from the full set of machine-translated sentences. | ['[BOLD] Approach', '[BOLD] Flickr8k-cn B-4', '[BOLD] Flickr8k-cn ROUGE', '[BOLD] Flickr8k-cn CIDEr', '[BOLD] Flickr30k-cn B-4', '[BOLD] Flickr30k-cn ROUGE', '[BOLD] Flickr30k-cn CIDEr'] | [['Late translation', '17.3', '39.3', '33.7', '15.3', '38.5', '27.1'], ['Late translation rerank', '17.5', '40.2', '34.2', '14.3', '38.5', '27.5'], ['Without fluency', '[BOLD] 24.1', '[BOLD] 45.9', '[BOLD] 47.6', '17.8', '[BOLD] 40.8', '32.5'], ['[ITALIC] Fluency-only', '20.7', '41.1', '35.2', '14.5', '35.9', '25.1'], ... | The reranking strategy improves over ‘Late translation’ showing the benefit of fluency modeling. Nevertheless, both ‘Late translation’ and ‘Late translation rerank’ perform worse than the ‘Without fluency’ run. Fluency-only is inferior to other proposed approaches as this model is trained on much less amounts of data, ... |
Fluency-Guided Cross-Lingual Image Captioning | 1708.04390 | Table 1. Datasets for sentence fluency estimation. The relatively low rate of fluency (less than 30%) in machine-translated sentences indicates the importance of fluency-guided learning for cross-lingual image captioning. | ['[EMPTY]', '[BOLD] training', '[BOLD] validation', '[BOLD] test'] | [['# fluent', '1,240', '291', '294'], ['# not fluent', '3,353', '709', '706']] | Setup. In order to train the four-way sentence fluency classifier, a number of paired bilingual sentences labeled as fluent / not fluent are a prerequisite. We aim to select a representative and diverse set of sentences for manual verification, meanwhile keeping the manual annotation affordable. To this end we sample a... |
Fluency-Guided Cross-Lingual Image Captioning | 1708.04390 | Table 4. Two datasets used in our image captioning experiments. Besides Flickr8k-cn (Li et al., 2016a), we construct Flickr30k-cn, a bilingual version of Flickr30k (Young et al., 2014) obtained by English-to-Chinese machine translation of its train / val sets and human translation of its test set. | ['[EMPTY]', '[BOLD] Flickr8k-cn\xa0(Li et\xa0al., 2016a ) train', '[BOLD] Flickr8k-cn\xa0(Li et\xa0al., 2016a ) val', '[BOLD] Flickr8k-cn\xa0(Li et\xa0al., 2016a ) test', '[BOLD] Flickr30k-cn (this work) train', '[BOLD] Flickr30k-cn (this work) val', '[BOLD] Flickr30k-cn (this work) test'] | [['Images', '6,000', '1,000', '1,000', '29,783', '1,000', '1,000'], ['Machine-translated Chinese sentences', '30,000', '5,000', '–', '148,915', '5,000', '–'], ['Human-translated Chinese sentences', '–', '–', '5,000', '–', '–', '5,000'], ['Human-annotated Chinese sentences', '30,000', '5,000', '5,000', '–', '–', '–']] | Setup. While we target at learning from machine-translated corpus, manually written sentences are needed to evaluate the effectiveness of the proposed framework. Each test image in Flickr8k-cn is associated with five Chinese sentences, obtained by manually translating the corresponding five English sentences from Flick... |
Character-level Chinese-English Translation through ASCII Encoding | 1805.03330 | Table 6: Word counts of the outputs of the char2char models (mean and standard deviation). | ['[BOLD] Model', '[BOLD] Word Count'] | [['[BOLD] wb2en', '25.01±10.95'], ['[BOLD] cn2en', '25.80±11.72'], ['[BOLD] en2wb', '21.61±9.68'], ['[BOLD] en2cn', '22.19±10.11']] | In overall, the Wubi-based outputs appear to be visibly better than the raw Chinese-based outputs, in both directions. |
Abstractive Summarization Improved by WordNet-based Extractive Sentences | 1808.01426 | Table 2: ROUGE F1 scores on CNN/Daily Mail non-anonymized testing dataset for all the controlled experiment models mentioned above. According to the official ROUGE usage description, all our ROUGE scores have a 95% confidence interval of at most ±0.25. PGN, Cov, ML, RL are abbreviations for pointer-generator, coverage,... | ['Models', 'ROUGE [BOLD] F1 scores 1', 'ROUGE [BOLD] F1 scores 2', 'ROUGE [BOLD] F1 scores L'] | [['Seq2seq + Attn', '31.50', '11.95', '28.85'], ['Seq2seq + Attn (150k)', '30.67', '11.32', '28.11'], ['Seq2seq + Attn + PGN', '36.58', '15.76', '33.33'], ['Seq2seq + Attn + PGN + Cov', '[BOLD] 39.16', '[BOLD] 16.98', '[BOLD] 35.81'], ['Lead-3 + Dual-attn + PGN', '37.26', '16.12', '33.87'], ['WordNet + Dual-attn + PGN'... | It counts the number of overlapping basic units including n-grams, longest common subsequences (LCS). We carry out the experiments based on original dataset, i.e., non-anonymized version of data. Attn + PGN, which is 0.09 points lower than the former result. |
A Joint Framework for Inductive Representation Learning and Explainable Reasoning in Knowledge Graphs | 2005.00637 | Table 5: The MRR for the test triples in inductive setting with to-Many and to-1 relation types. The % columns show the percentage of test triples for each relation type. | ['Dataset', 'to-Many %', 'to-Many MRR', 'to-1 %', 'to-1 MRR'] | [['FB15k-237-Inductive', '77.4', '31.6', '22.6', '75.5'], ['WN18RR-Inductive', '48.1', '60.8', '51.9', '30.1'], ['NELL-995-Inductive', '7.6', '41.4', '92.4', '78.5']] | Following \newciteBordes:2013, we categorize the relations in the seen snapshot of the knowledge graph into Many-to-1 and 1-to-Many relations. The categorization is done based on the ratio of the cardinality of the target answer entities to the source entities. If the ratio is greater than 1.5, we categorize the relati... |
Multi-Document Abstractive Summarization Using ILP based Multi-Sentence Compression | 1609.07034 | Table 1: Comparison of ROUGE scores on the DUC 2004 and 2005 datasets: Baselines, state-of-the-arts, our proposed methods and abstractive summarization system using MSC. “†” denotes the differences between [ILPSumm] and the baselines on the ROUGE scores are statistically significant for p<0.05. We limit ROUGE evaluatio... | ['[BOLD] DUC-2004', '[BOLD] DUC-2004 [BOLD] ROUGE-2', '[BOLD] DUC-2004 [BOLD] ROUGE-SU4', '[BOLD] DUC-2005', '[BOLD] DUC-2005 [BOLD] ROUGE-L', '[BOLD] DUC-2005 [BOLD] ROUGE-SU4'] | [['[BOLD] Baselines', '[EMPTY]', '[EMPTY]', '[BOLD] Baselines', '[EMPTY]', '[EMPTY]'], ['GreedyKL', '0.08658', '0.13253', 'Random', '0.26395', '0.09066'], ['FreqSum', '0.08218', '0.12448', 'Centroid', '0.32562', '0.11007'], ['Centroid', '0.08139', '0.12642', 'LexRank', '0.33179', '0.12021'], ['TsSum', '0.08068', '0.122... | Hence, we have six different systems in total. To the best of our knowledge, no publicly available abstractive summarizers have been used on the DUC dataset. In MSC, the input is a pre-defined cluster of similar sentences. Therefore, we compare our ILP based technique with MSC using the same set of input clusters obtai... |
Multi-Document Abstractive Summarization Using ILP based Multi-Sentence Compression | 1609.07034 | Table 2: Manual evaluation by 10 evaluators on Informativeness (Inf) and Linguistic Quality (LQ) of summaries. Average Log-likelihood scores (Avg.LL) from parser are also shown. | ['[BOLD] Type', '[BOLD] Inf', '[BOLD] LQ', '[BOLD] Avg.LL'] | [['Human written', '4.42', '4.35', '-129.02'], ['Extractive (DPP)', '3.90', '3.81', '-142.70'], ['Abstractive (MSC)', '3.78', '2.83', '-210.02'], ['Abstractive (ILPSumm)', '4.10', '3.63', '-180.76']] | The four summaries provided to the evaluators are human-written summary (one summary collected randomly from four model-summaries per cluster), extractive summary (DPP), abstractive summary generated using MSC (MSC) and abstractive summary generated using our ILP based method (ILPSumm). We asked each evaluator to compl... |
Multi-Domain Neural Machine Translation | 1805.02282 | Table 6: BLEU scores and p-values for test on Wikipedia-only data to compare the effect of Unsupervised clustering (Usup12, Usup5), supervised clustering (Super) and no-clustering approach (Ref). The p-values are shown in respect to the version where the value is -. | ['[BOLD] NClust', '[BOLD] Usup12', '[BOLD] Usup5', '[BOLD] Super', '[BOLD] Ref'] | [['[BOLD] BLEU', '26.0±0.4', '25.2±0.4', '25.8±0.4', '23.6±0.4'], ['[BOLD] pU12', '-', '0.01', '0.1', '0.0001'], ['[BOLD] pU5', '0.01', '-', '0.03', '0.0001'], ['[BOLD] pSup', '0.1', '0.03', '-', '0.0001'], ['[BOLD] pRef', '0.0001', '0.0001', '0.0001', '-']] | In the final experiment, three models were trained: Supervised 5-domain source tag model Unsupervised 5-domain source tag model Unsupervised 12-domain source tag model Regular not domain-tagged model Unsupervised 5-domain model was included to compare the performance of supervised and unsupervised approach with the sam... |
Compositional Approaches for Representing Relations Between Words: A Comparative Study | 1709.01193 | Table 4: Accuracy of the compositional operators for knowledge base completion task. | ['Compositional operator', 'WN18 MeanRank', 'WN18 Hits@10(%)', 'FB15k MeanRank', 'FB15k Hits@10(%)'] | [['PairDiff', '13,198', '11.34', '1,206', '44.4'], ['Concat', '9,896', '2.77', '542', '29.49'], ['Add', '12,178', '1.88', '1,211', '21.7'], ['Mult', '[BOLD] 812', '[BOLD] 54.93', '[BOLD] 256', '[BOLD] 50.66']] | As can be seen from the Table, Mult operator yields the lowest mean rank and the highest Hits@10 accuracy among other operators for the both knowledge bases. On the other hand, PairDiff operator scores test entity pairs by (→h−→t)\T(→h′−→t′). Here, (h′,t′) is an entity pair in the test dataset with the target relation ... |
Compositional Approaches for Representing Relations Between Words: A Comparative Study | 1709.01193 | Table 3: Accuracy of the compositional operators for relational similarity prediction and relational classification (last right column). | ['Representation model', 'Compositional operator', 'SAT', 'SemEval', 'MSR', 'Google Sem.', 'Google Syn.', 'Google Total', 'DIFFVECS'] | [['CBOW', 'PairDiff', '[BOLD] 41.82', '[BOLD] 44.35', '[BOLD] 30.16', '[BOLD] 24.43', '[BOLD] 32.31', '[BOLD] 28.74', '[BOLD] 87.38'], ['[EMPTY]', 'Concat', '38.07', '41.06', '0.39', '3.01', '1.26', '2.05', '83.74'], ['[EMPTY]', 'Add', '31.1', '36.37', '0.06', '0.16', '0.15', '0.15', '79.27'], ['[EMPTY]', 'Mult', '27.8... | We observe that PairDiff achieves the best results compared with other operators for all the evaluated datasets and all the word representation methods. PairDiff is significantly better than Add or Mult for all embeddings (both prediction- and counting-based) in MSR, Google and DiffVec datasets according to Clopper-Pea... |
Language Identification of Bengali-English Code-Mixed data using Character & Phonetic based LSTM Models | 1803.03859 | Table 5: Confusion matrices. (1-character, 2-phonetic, 3-ensem_stack, 4-ensem_thresh) | ['[BOLD] Confusion Matrices 1', '[BOLD] Confusion Matrices [ITALIC] BN', '[BOLD] Confusion Matrices [ITALIC] EN', '[BOLD] Confusion Matrices 2', '[BOLD] Confusion Matrices [ITALIC] BN', '[BOLD] Confusion Matrices [ITALIC] EN'] | [['BN', '641', '59', 'BN', '644', '56'], ['EN', '57', '643', 'EN', '78', '622'], ['3', '[ITALIC] BN', '[ITALIC] EN', '4', '[ITALIC] BN', '[ITALIC] EN'], ['BN', '623', '77', 'BN', '667', '33'], ['EN', '38', '662', 'EN', '74', '626']] | We started our error analysis by preparing confusion matrices (CM) of the four models. Here, BN correctly predicted was considered as TP and EN was considered as TN. Unlike the character model, we can see from CM 2 (phonetic model) that TP is much more than TN, thus the precision is higher by 0.15% compared to the char... |
Language Identification of Bengali-English Code-Mixed data using Character & Phonetic based LSTM Models | 1803.03859 | Table 2: Character model results (in %). | ['[BOLD] Model', '[BOLD] Acc', '[BOLD] Prec', '[BOLD] Rec', '[BOLD] F1'] | [['dev_round', '91.50', '87.84', '96.33', '91.48'], ['dev_thresh', '92.16', '91.47', '93.00', '92.16'], ['test_thresh', '91.71', '91.59', '91.85', '91.71']] | Simple round up on the fuzzy output gave an accuracy of 91.50%. Using brute force, the threshold where the accuracy is peaking on the development data was found to be at θ ≤ 0.92. This showed a slight improvement in accuracy by 0.66%. Finally, on test data, this tuned model gave an accuracy of 91.71%. |
Language Identification of Bengali-English Code-Mixed data using Character & Phonetic based LSTM Models | 1803.03859 | Table 3: Phonetic model results (in %). | ['[BOLD] Model', '[BOLD] Acc', '[BOLD] Prec', '[BOLD] Rec', '[BOLD] F1'] | [['dev_round', '82.50', '75.06', '97.33', '82.10'], ['dev_thresh', '88.66', '87.66', '90.00', '88.66'], ['test_thresh', '90.42', '91.74', '88.85', '90.42']] | On simple roundup on development data, the accuracy was not very good (82.5%), but using a similar threshold technique as used in the character model, the accuracy improved significantly (6.16%) and reached 88.66%. The threshold in this case was calculated to be θ ≤ 0.95. Using threshold, the accuracy achieved on the t... |
Using Linguistic Features to Estimate Suicide Probability of Chinese Microblog Users | 1411.0861 | Table 2: Comparison of LIWC and topic features within predictive models of SPS score using RMSE as metric | ['Number', 'Trained', 'Inferred', 'LIWC+Trained', 'LIWC+Inferred'] | [['10', '11.84', '11.74', '11.79', '11.83'], ['20', '11.75', '11.83', '11.79', '11.79'], ['30', '11.81', '11.74', '11.78', '11.84'], ['40', '11.74', '11.91', '11.76', '11.86'], ['50', '11.96', '11.98', '11.82', '11.79'], ['60', '11.79', '11.79', '11.80', '11.86'], ['70', '11.86', '11.68', '11.88', '11.85'], ['80', '12.... | When we combine LIWC and trained topics as features, predict model was improved with regard to big number of topics compared with using trained topics only. When we combine LIWC and inferred topics as features, predict model was not improved. In general, only inferred topics can achieve highest performance no matter wi... |
Augmenting Visual Question Answering with Semantic Frame Information in a Multitask Learning Approach | 2001.11673 | TABLE VIII: Performance evaluation grouped by performance intervals showing verb frequency and role frequency in each group. | ['Accuracy', 'Verb', 'Role'] | [['Difference', 'Frequency', 'Frequency'], ['Range', '[EMPTY]', '[EMPTY]'], ['(-40%,-30%]', '[EMPTY]', '3'], ['(-30%,-20%]', '2', '3'], ['(-20%,-10%]', '10', '5'], ['(-10%,0%)', '67', '24'], ['0%', '27', '32'], ['(0%,10%]', '269', '68'], ['(10%,20%]', '100', '24'], ['(20%,30%]', '15', '13'], ['(30%,40%]', '6', '[EMPTY]... | Fine-grained evaluation. In order to perform a fine-grained analysis of the results, performance per question, per verb and per role are computed. Multi-task CNN-LSTM performs better for who (4%), what (8%) and where (5%) when compared to CNN-LSTM. Exploring performance per verb, we can see for example cooking improves... |
Interpreting Transformed-based Models by Concept Classification | 2005.07647 | Table 5: Performance of the considered models on various downstream tasks, as reported in the reference papers. Not all models report performance on all tasks. | ['Model', 'BERT-B', 'BERT-L', 'Distilbert', 'RoBERTa-L', 'XLM'] | [['Model size', '110M', '330M', '66M', '355M', '667M'], ['GLUE Score', '78.3', '80.5', '76.8', '88.5', '83.1'], ['CoLA', '52.1', '60.5', '49.1', '67.8', '62.9'], ['SST-2', '93.5', '94.9', '92.7', '96.7', '95.6'], ['MRPC (acc)', '88.9', '89.3', '90.2', '92.3', '90.7'], ['MRPC (F1)', '84.8', '85.4', '89.8', '87.1', '[EMP... | Finding an optimal γ⋆ value The choice of γ is important to compute the concept expertise Xγ in Eq. The goal is to obtain an optimal γ⋆ that produces an expertise representative of the generalization power of TMs. As a measure of generalization, we use the average performance of each model on typical downstream tasks: ... |
Compositional Vector Space Models for Knowledge Base Completion | 1504.06662 | Table 4: Results comparing the zero-shot model with supervised RNN and a random baseline on 10 types. RNN is the fully supervised model described in section 3 while zero-shot is the model described in section 4. The zero-shot model without explicitly training for the target relation types achieves impressive results by... | ['[EMPTY]', 'train with top 1000 paths', 'train with all paths'] | [['Method', 'MAP', 'MAP'], ['RNN', '43.82', '50.10'], ['zero-shot', '19.28', '20.61'], ['Random', '7.59', '[EMPTY]']] | We evaluate on randomly selected 10 (out of 46) relation types, hence for the fully supervised version we train 10 RNNs, one for each relation type. For evaluating the zero-shot model, we randomly split the relations into two sets of equal size and train a zero-shot model on one set and test on the other set. So, in th... |
Human acceptability judgements for extractive sentence compression | 1902.00489 | Table 3: Test set accuracy, Fleiss’ κ and ROC AUC scores for six models, trained on the single-prune dataset (§4), as well as scores for a model trained on the CoLA dataset Warstadt et al. (2018). The simplest model uses only language modeling (LM) features. We add dependency type (+ dependencies) and worker ID (+ work... | ['Model', 'Hard classification ( [ITALIC] t=0.5) Accuracy', 'Hard classification ( [ITALIC] t=0.5) Fleiss [ITALIC] κ', 'Ranking ROC AUC ( [ITALIC] p)'] | [['CoLA', '0.622', '-0.210', '0.590 ('], ['language model (LM)', '0.623', '-0.232', '0.583 ('], ['+ dependencies', '0.664', '0.124', '0.646 ('], ['+ worker ID', '0.695', '0.232', '0.746 ('], ['full ≜ [ITALIC] p( [ITALIC] Y=1| [ITALIC] W, [ITALIC] x)', '[BOLD] 0.742', '[BOLD] 0.400', '[BOLD] 0.807'], ['- dependencies', ... | We use Norm LP as a part of two features in our approach. One real-valued feature records the probability of a compression computed by Norm LP(c). Another binary feature computes Norm LP(s) - Norm LP(c) > 0. The test set performance of these language model The appendix further describes our implementation of Norm LP. L... |
Incremental Active Opinion Learning Over a Stream of Opinionated Documents | 1509.01288 | Table 1: Requested labels per method and experiment: numbers in percentage regarding the length of the stream including the documents to train the classifier, i.e. the size of the seed. (IG=Information Gain), (U=Uncertainty) | ['Experiment + Dataset', 'ACOSTREAM\xa0(IG)', 'ACOSTREAM\xa0(U)', 'IncrementalMNB', 'StaticMNB', 'Random'] | [['fixed [ITALIC] V: StreamJi', '44', '40', '100', '1', '40'], ['fixed [ITALIC] V: TwitterSentiment', '40', '47', '100', '1', '42'], ['evolving [ITALIC] V: StreamJi', '60', '59', '100', '1', '60'], ['evolving [ITALIC] V: TwitterSentiment', '52', '88', '100', '2', '31']] | In this section, we compare ACOSTREAM using information gain and uncertainty sampling strategies against the IncrementalMNB, the StaticMNB as well as the random sampling based on the performance of kappa over time. This would, however, lead to an unfair comparison as the budget would be spent differently among the stra... |
Encoders Help You Disambiguate Word Senses in Neural Machine Translation | 1908.11771 | Table 2: BLEU scores of NMT models, and WSD accuracy on the test set using word embeddings or hidden states to represent ambiguous nouns. The hidden states are from the highest layer.555For encoders in RNNS2Ss, this is the last backward RNN. RNN. and Trans. denote RNNS2S and Transformer models, respectively. | ['[EMPTY]', 'DE→EN [ITALIC] RNN.', 'DE→EN [ITALIC] Trans.', 'DE→FR [ITALIC] RNN.', 'DE→FR [ITALIC] Trans.'] | [['BLEU', '29.1', '32.6', '17.0', '19.3'], ['[ITALIC] Embedding', '63.1', '63.2', '68.7', '68.9'], ['[ITALIC] ENC', '94.2', '97.2', '91.7', '95.6'], ['[ITALIC] DEC', '97.5', '98.3', '95.1', '96.9']] | ENC denotes encoder hidden states; DEC means decoder hidden states. We also concatenate the hidden states from both forward and backward RNNs and get higher accuracy, 96.8% in DE→EN and 95.7% in DE→FR. The WSD accuracy of using bidirectional hidden states are competitive to using hidden states from Transformer models. ... |
On the coexistence of competing languages | 2003.04748 | Table 1: Mean fraction ⟨K⟩/N of surviving languages for a uniform distribution of attractivenesses. Comparison between numerically measured values for N=10 and N=40 (see Figure 5) and the asymptotic analytical prediction (39), for four values of W. | ['[ITALIC] W', '[ITALIC] N=10', '[ITALIC] N=40', 'Eq.\xa0( 39 )'] | [['2', '0.750', '0.718', '0.70711'], ['4', '0.541', '0.510', '0.5'], ['10', '0.356', '0.326', '0.31623'], ['30', '0.222', '0.192', '0.18257']] | Nw. (38) Each dataset is the outcome of 107 draws of the attractiveness profile. The widths of the distributions pK are observed to shrink as N is increased, in agreement with the expected 1/√N behavior stemming from the law of large numbers. |
FAQ-based Question Answering via Word Alignment | 1507.02628 | Table 1: Evaluation on the test set | ['[EMPTY]', 'English', 'Spanish', 'Japanese'] | [['BagOfWord', '31.63', '36.23', '55.71'], ['IDF-VSM', '37.76', '37.68', '58.29'], ['Similarity', '41.84', '40.58', '67.43'], ['Dense', '45.92', '44.90', '67.14'], ['Sparse', '51.02', '50.72', '69.43'], ['SparseHidden', '52.04', '59.42', '70.29']] | In this section, we evaluated our systems on the test sets. We tested three systems: (1) “Dense” takes the dense features, (2) “Sparse” takes both dense and sparse features, and (3) “SparseHidden” adds 300 hidden neurons to the second system. We also designed three baseline systems: (1) “BagOfWord” calculates question ... |
FAQ-based Question Answering via Word Alignment | 1507.02628 | Table 2: Evaluation on Answer Sentence Selection | ['[EMPTY]', 'MAP', 'MRR'] | [['wang2007jeopardy', '0.603', '0.685'], ['heilman2010tree', '0.609', '0.692'], ['yao2013answer', '0.631', '0.748'], ['severyn2013automatic', '0.678', '0.736'], ['yih2013question', '0.709', '0.770'], ['yu2014deep', '0.711', '0.785'], ['Our Method', '0.746', '0.820']] | The task is to rank candidate answers for each question, which is very similar to our FAQ-based QA task. We used the same experimental setup as \newcitewang2007jeopardy, and evaluated the result with Mean Average Precision (MAP) and Mean Reciprocal Rank (MRR). We observe our system get a significant improvement than th... |
What Do Recurrent Neural Network Grammars Learn About Syntax? | 1611.05774 | Table 1: Phrase-structure parsing performance on PTB §23. All results are reported using single-model performance and without any additional data. | ['[BOLD] Model', '[ITALIC] F1'] | [['vinyals:2015 – PTB only', '88.3'], ['Discriminative RNNG', '[BOLD] 91.2'], ['choe:2016 – PTB only', '92.6'], ['Generative RNNG', '[BOLD] 93.3']] | \newcitevinyals:2015 directly predict the sequence of nonterminals, “shifts” (which consume a terminal symbol), and parentheses from left to right, conditional on the input terminal sequence x, while \newcitechoe:2016 used a sequential LSTM language model on the same linearized trees to create a generative variant of t... |
Towards Instance-Level Parser Selection for Cross-Lingual Transfer of Dependency Parsers | 2004.07642 | Table 1: Results for single-parser selection models. Results for 42 parsers (an exception is the MSP model which trains a single parser on the concatenation of all training treebanks) on 20 low-resource test languages. Ma & Mi: average performance across 20 languages, macro- and micro-averaged scores, respectively. The... | ['[EMPTY]', 'am', 'be', 'bm', 'br', 'bxr', 'cop', 'fo', 'ga', 'hsb', 'hy', 'kk', 'lt', 'mt', 'myv', 'sme', 'ta', 'te', 'th', 'yo', 'yue', '[BOLD] Ma', '[BOLD] Mi'] | [['[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles', '[ITALIC] Oracles'... | Single-parser selection. We report results (UAS) for all single-parser selection methods (i.e., no ensembles) along with the oracle scores on all 20 test treebanks. Allowing for the selection of more than a single parser in cases in which our ILPS-based predictions warrant so (i.e., when two or more parsers yield simil... |
Zero-Resource Cross-Domain Named Entity Recognition | 2002.05923 | Table 1: F1-scores on the target domain. Models are implemented based on the corresponding embeddings. | ['[BOLD] Model', '[BOLD] BERT [BOLD] Fine-tune', '[BOLD] FastText [BOLD] unfreeze', '[BOLD] FastText [BOLD] freeze'] | [['[ITALIC] Baseline', '[ITALIC] Baseline', '[ITALIC] Baseline', '[ITALIC] Baseline'], ['Concept Tagger', '67.14', '62.34', '66.86'], ['Robust Sequence Tagger', '67.31', '63.66', '68.12'], ['[ITALIC] Zero-Resource', '[ITALIC] Zero-Resource', '[ITALIC] Zero-Resource', '[ITALIC] Zero-Resource'], ['BiLSTM-CRF', '67.55', '... | We conjecture that these two baselines, which utilize slot descriptions or slot examples, are suitable for limited slot names in the slot filling task, while they fail to cope with wide variances of entity names in the NER task across different domains, while our model is more robust to the domain variations. MTL helps... |
Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation | 1806.03661 | Table 2: Left Side: Test-wise results for ”Small” models in Figure 4, Right Side: Test-wise results for ”Large” models in Figure 4 | ['Pair ar-en', 'Agent WUE', 'test12 30.16', 'test13 28.16', 'test14 25.53', 'Agent WUE', 'test12 32.84', 'test13 32.23', 'test14 28.95'] | [['[EMPTY]', '5,2', '29.31', '27.72', '25.21', '7,2', '31.71', '31.46', '28.29'], ['[EMPTY]', 'WIW', '28.06', '25.86', '23.75', 'WIW', '29.48', '28.82', '26.52'], ['[EMPTY]', 'WID', '19.89', '17.24', '15.64', '[EMPTY]', '[EMPTY]', '[EMPTY]', '[EMPTY]'], ['cs-en', 'WUE', '22.95', '25.03', '–', 'WUE', '27.97', '30.50', '... | Here, we present the test-wise results for the interested reader. Missing table values correspond to unavailable test-sets on the IWSLT webpage. we note that even though the WIW agent’s performance is not significantly below our selected STATIC-RW agent, its AP is much higher. When we allow our STATIC-RW agent an AP si... |
Domain specialization: a post-training domain adaptation for Neural Machine Translation | 1612.06141 | Table 2: BLEU score of full trained systems. | ['Models', 'BLEU', 'TER'] | [['[ITALIC] generic', '26.23', '62.47'], ['[ITALIC] generic+emea-0.5K', '26.48', '63.09'], ['[ITALIC] generic+emea-5K', '28.99', '58.98'], ['[ITALIC] generic+emea-50K', '33.76', '53.87'], ['[ITALIC] generic+emea-full', '41.97', '47.07']] | As a baseline, we fully trained five systems, one with the generic data (generic) and the other with generic and various amount of in-domain data: 500 lines (emea-0.5K), 5K lines (emea-5K) and 50K lines (emea-50K). Without surprises, the more the model is trained with in-domain data the more BLEU and TER scores are imp... |
Finding Better Subword Segmentation for Neural Machine Translation | 1807.09639 | Table 2: BLEU scores with frequency-weighted schemes on German-English test set | ['[ITALIC] [BOLD] N', '[BOLD] FRQ′-BPE [BOLD] Greedy', '[BOLD] FRQ′-BPE [BOLD] Beam', '[BOLD] AV′-BPE [BOLD] Greedy', '[BOLD] AV′-BPE [BOLD] Beam', '[BOLD] DLG′-BPE [BOLD] Greedy', '[BOLD] DLG′-BPE [BOLD] Beam'] | [['10 [ITALIC] K', '27.36', '28.86', '27.72', '29.02', '27.56', '29.25'], ['15 [ITALIC] K', '27.42', '28.92', '27.56', '29.25', '27.75', '29.30'], ['20 [ITALIC] K', '27.51', '28.99', '27.65', '[BOLD] 29.46++', '[BOLD] 27.76+', '29.36'], ['25 [ITALIC] K', '27.22', '28.81', '27.69', '29.32', '27.72', '29.34'], ['30 [ITAL... | For greedy search, both DLG′-BPE and AV′-BPE achieve better performance than FRQ′-BPE and the best score is obtained with DLG′-BPE of 20K merge operations. For beam search, AV′-BPE and DLG′-BPE outperform FRQ-BPE′ with improvement of 0.47 and 0.37 BLEU points when N=20K respectively. In general, AV′-BPE and DLG′-BPE ar... |
Finding Better Subword Segmentation for Neural Machine Translation | 1807.09639 | Table 3: BLEU scores on Chinese-English test set (beam size=10) | ['[BOLD] Source', '[BOLD] Target [BOLD] FRQ′-BPE', '[BOLD] Target [BOLD] AV′-BPE', '[BOLD] Target [BOLD] DLG′-BPE'] | [['[BOLD] FRQ′-BPE', '20.07', '20.27', '20.31'], ['[BOLD] AV′-BPE', '19.96', '20.23', '20.13'], ['[BOLD] DLG′-BPE', '20.70', '20.45', '[BOLD] \u200320.72++']] | Here, the subword segmentation learning is performed separately on the source and target sides with 30K merge operations which is tuned the same way as German-English task. We also try different goodness measures on each side at the same time. It can be found that DLG′-BPE on the both sides significantly outperforms FR... |
Very Deep Self-Attention Networks for End-to-End Speech Recognition | 1904.13377 | Table 2: Comparing our best model to other hybrid and end-to-end systems reporting on Hub5’00 test set with 300h SWB training set. | ['[BOLD] Hybrid/End-to-End Models', '[BOLD] Tgt Unit', '[BOLD] SWB', '[BOLD] CH'] | [['TDNN +LFMMI ', 'Phone', '10.0', '20.1'], ['BLSTM +LFMMI ', 'Phone', '[BOLD] 9.6', '19.3'], ['CTC+CharLM ', 'Char', '21.4', '40.2'], ['LSTM w/attention\xa0', 'Char', '15.8', '36.0'], ['Iterated-CTC +LSTM-LM ', 'Char', '14.0', '25.3'], ['Seq2Seq +LSTM-LM ', 'BPE', '11.8', '25.7'], ['Seq2Seq +Speed Perturbation ', 'Cha... | Third, it was revealed that the combination of our regularization techniques (dropout, label-smoothing and stochastic networks) are additive with data augmentation, which further improved our result to 18.1% with the 36−12 setup. Comparing to the best hybrid models with similar data constraints, our models outperformed... |
Very Deep Self-Attention Networks for End-to-End Speech Recognition | 1904.13377 | Table 1: The performance of deep self-attention networks with and without stochastic layers on Hub5’00 test set with 300h SWB training set. | ['[BOLD] Layers', '[BOLD] #Param', '[BOLD] SWB', '[BOLD] CH'] | [['04Enc-04Dec', '21M', '20.8', '33.2'], ['08Enc-08Dec', '42M', '14.8', '25.5'], ['12Enc-12Dec', '63M', '13.0', '23.9'], ['[ITALIC] +Stochastic Layers', '[EMPTY]', '13.1', '23.6'], ['24Enc-24Dec', '126M', '12.1', '23.0'], ['[ITALIC] +Stochastic Layers', '[EMPTY]', '11.7', '21.5'], ['[ITALIC] +Speed Perturbation', '[EMP... | A shallow configuration (i.e 4 layers) is not sufficient for the task, and the WER reduces from 20.8% to 12.1% on the SWB test as we increase the depth from 4 to 24. The improvement is less significant between 12 and 24 (only 5% relative WER), which seems to be a symptom of overfitting. |
Very Deep Self-Attention Networks for End-to-End Speech Recognition | 1904.13377 | Table 3: The Transformer results on the TED-LIUM test set using TED-LIUM 3 training set. | ['[BOLD] Models', 'Test WER'] | [['CTC\xa0', '17.4'], ['CTC/LM + speed perturbation\xa0', '13.7'], ['12Enc-12Dec (Ours)', '14.2'], ['Stc. 12Enc-12Dec (Ours)', '12.4'], ['Stc. 24Enc-24Dec (Ours)', '11.3'], ['Stc. 36Enc-12Dec (Ours)', '[BOLD] 10.6']] | With a similar configuration to the SWB models, we outperformed a strong baseline which uses both an external language model trained on larger data than the available transcription and speed perturbation, using our model with 36 encoder layers and 12 decoder layers. This result continues the trend that these models ben... |
Building a Word Segmenter for Sanskrit Overnight | 1802.06185 | Table 1: Micro-averaged Precision, Recall and F-Score for the competing systems on the test dataset of 4200 strings. | ['[BOLD] Model', '[BOLD] Precision', '[BOLD] Recall', '[BOLD] F-Score'] | [['GraphCRF', '65.20', '66.50', '65.84'], ['SupervisedPCRW', '76.30', '79.47', '77.85'], ['segSeq2Seq', '73.44', '73.04', '73.24'], ['attnsegSeq2Seq', '90.77', '90.3', '90.53']] | We can find that the system ‘attnSegSeq2Seq’ outperforms the current state of the art with a percent increase of 16.29 % in F-Score. The model ‘segSeq2Seq’ falls short of the current state of the art with a percent decrease of 6.29 % in F-Score. It needs to be noted that the systems ‘attnSegSeq2Seq’ and ‘segSeq2Seq’ ar... |
Explainable Prediction of Medical Codes from Clinical Text | 1802.05695 | Table 4: Results on MIMIC-III full, 8922 labels. Here, “Diag” denotes Micro-F1 performance on diagnosis codes only, and “Proc” denotes Micro-F1 performance on procedure codes only. Here and in all tables, (*) by the bold (best) result indicates significantly improved results compared to the next best result, p<0.001. | ['Model', 'AUC Macro', 'AUC Micro', 'F1 Macro', 'F1 Micro', 'F1 Diag', 'F1 Proc', 'P@n 8', 'P@n 15'] | [['Scheurwegs et. al Scheurwegs et\xa0al. ( 2017 )', '–', '–', '–', '–', '0.428', '0.555', '–', '–'], ['Logistic Regression', '0.561', '0.937', '0.011', '0.272', '0.242', '0.398', '0.542', '0.411'], ['CNN', '0.806', '0.969', '0.042', '0.419', '0.402', '0.491', '0.581', '0.443'], ['Bi-GRU', '0.822', '0.971', '0.038', '0... | Our main quantitative evaluation involves predicting the full set of ICD-9 codes based on the text of the MIMIC-III discharge summaries. The CAML model gives the strongest results on all metrics. Attention yields substantial improvements over the “vanilla” convolutional neural network (CNN). The recurrent Bi-GRU archit... |
Enhancing Sentence Relation Modeling with Auxiliary Character-level Embedding | 1603.09405 | Table 3: Results of Tree LSTM vs Sequence LSTM on auxiliary char embedding. | ['Method', 'Accuracy', 'Pearson'] | [['Dep-Tree LSTM', '0.833', '0.849'], ['Dep-Tree LSTM + CNN', '0.798', '0.822'], ['LSTM', '0.812', '0.833'], ['LSTM + CNN', '0.776', '0.810'], ['1-Bidirectional LSTM', '0.834', '0.848'], ['1-Bidirectional LSTM+ CNN', '0.821', '0.846']] | We can see that, if we don’t deploy CNN, simple Tree LSTM yields better result than traditional LSTM, but worse than Bidirectional LSTM. This is reasonable due to the fact that Bidirectional LSTM can enhance sentence representation by concatenating forward and backward representations. We found that adding CNN layer wi... |
Neural Zero-Inflated Quality Estimation ModelFor Automatic Speech Recognition System | 1910.01289 | Table 3: The comparison of different quality estimation models. | ['Method', 'MAE', 'Pearson', 'F1-OK/BAD'] | [['QEBrain', '0.073', '0.7829', '0.4956'], ['speech-BERT', '[BOLD] 0.056', '[BOLD] 0.8187', '[BOLD] 0.5372']] | The second experiment we conduct on our in-domain test dataset is to compare our ASR-QE model as an integrated pipeline with the state-of-the-art quality estimation model Notice that for fair comparison, we have to modify the text encoder of QEBrain to the exactly same speech encoder of ours and the last layer to a zer... |
On Tree-Based Neural Sentence Modeling | 1808.09644 | Table 2: Test results for different encoder architectures trained by a unified encoder-classifier/decoder framework. We report accuracy (×100) for classification tasks, and BLEU score (Papineni et al., 2002; word-level for English targets and char-level for Chinese targets) for generation tasks. Large is better for bot... | ['[BOLD] Model', '[ITALIC] Sentence Classification [BOLD] AGN', '[ITALIC] Sentence Classification [BOLD] ARP', '[ITALIC] Sentence Classification [BOLD] ARF', '[ITALIC] Sentence Classification [BOLD] DBpedia', '[ITALIC] Sentence Classification [BOLD] WSR', '[ITALIC] Sentence Relation [BOLD] NLI', '[ITALIC] Sentenc... | [['[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees', '[ITALIC] Latent Trees'], ['Gumbel', '91.8', '87.1', '48.4', '98.6',... | However, left-branching tree LSTM has a quite similar structure with linear LSTM, and it can be viewed as a linear LSTM-on-LSTM structure. It has the same amounts of parameters as other tree-based encoders, but still falls behind the balance tree encoder on most of the tasks. This indicates that larger quantity of para... |
A Unified MRC Framework for Named Entity Recognition | 1910.11476 | Table 2: Results for nested NER tasks. | ['[BOLD] English ACE 2004 [BOLD] Model', '[BOLD] English ACE 2004 [BOLD] Precision', '[BOLD] English ACE 2004 [BOLD] Rrecall', '[BOLD] English ACE 2004 [BOLD] F1'] | [['Hyper-Graph Katiyar and Cardie ( 2018 )', '73.6', '71.8', '72.7'], ['Seg-Graph Wang and Lu ( 2018 )', '78.0', '72.4', '75.1'], ['Seq2seq-BERT Straková et\xa0al. ( 2019 )', '-', '-', '84.40'], ['Path-BERT Shibuya and Hovy ( 2019 )', '83.73', '81.91', '82.81'], ['DYGIE Luan et\xa0al. ( 2019 )', '-', '-', '84.7'], ['BE... | We observe huge performance boosts on the nested NER datasets over previous state-of-the-art models, achieving F1 scores of 85.98%, 86.88%, 83.75% and 80.97% on ACE04, ACE05, GENIA and KBP-2017 datasets, which are +1.28%, +2.55%, +5.44% and +6.37% over previous SOTA performances, respectively. |
Phonetic-enriched Text Representationfor Chinese Sentiment Analysiswith Reinforcement Learning | 1901.07880 | TABLE X: Performance comparison between learned and random generated phonetic feature. | ['Random phonetic feature (rand)', 'Random phonetic feature (rand)', 'Weibo 53.83', 'It168 56.85', 'Chn2000 55.71', 'Review-4 69.20', 'Review-5 69.77'] | [['Learned phonetic feature', 'Ex0', '66.49', '84.21', '77.82', '81.36', '83.24'], ['Learned phonetic feature', 'Ex04', '67.28', '[BOLD] 84.69', '[BOLD] 78.18', '81.88', '83.38'], ['Learned phonetic feature', 'PO', '64.28', '82.30', '77.09', '83.97', '82.71'], ['Learned phonetic feature', 'PW', '[BOLD] 67.80', '83.73',... | Iv-E1 Validating phonetic feature So far, we have examined the effectiveness of our model as a whole by comparing it with different baselines. In this section, we break down the proposed methods into a reinforcement learning framework and a set of features. First of all, we would like to validate if the performance gai... |
Phonetic-enriched Text Representationfor Chinese Sentiment Analysiswith Reinforcement Learning | 1901.07880 | TABLE VII: Classification accuracy of unimodality in LSTM. | ['GloVe', 'GloVe', 'Weibo [BOLD] 75.39', 'It168 81.82', 'Chn2000 [BOLD] 84.54', 'Review-4 [BOLD] 87.46', 'Review-5 [BOLD] 86.94'] | [['CBOW', 'CBOW', '72.39', '78.75', '81.18', '85.11', '84.71'], ['Skip-gram', 'Skip-gram', '75.05', '80.13', '78.04', '86.23', '86.21'], ['Visual', 'Visual', '61.78', '65.40', '67.21', '78.98', '79.59'], ['charCBOW', 'charCBOW', '71.54', '80.83', '82.82', '86.90', '85.19'], ['charSkipGram', 'charSkipGram', '71.86', '82... | The training procedure of our DISA network is as follows. Firstly, we skip the policy network and directly train the LSTM critic network with the training objective as Eq. Secondly, we fix the parameters of the LSTM critic network and train the policy network with the training objective as Eq. Lastly, we co-train all t... |
Phonetic-enriched Text Representationfor Chinese Sentiment Analysiswith Reinforcement Learning | 1901.07880 | TABLE IX: Cross-domain evaluation. Datasets on the first column are the training sets. Datasets on the first row are the testing sets. The second column represents various baselines and our proposed method. | ['Weibo', 'Hsentic', 'Weibo -', 'It168 66.47', 'Chn2000 61.84', 'Review-4 64.93', 'Review-5 63.71'] | [['Weibo', 'charCBOW', '-', '67.55', '64.08', '62.09', '67.78'], ['Weibo', 'charSkipGram', '-', '65.29', '59.60', '53.22', '49.49'], ['Weibo', 'DISA(T+P)', '-', '[BOLD] 73.68', '[BOLD] 66.55', '[BOLD] 69.16', '[BOLD] 71.01'], ['It168', 'Hsentic', '59.15', '-', '59.30', '69.76', '67.62'], ['It168', 'charCBOW', '57.54', ... | In this section, we examine how our model performs across different domains and datasets in order to validate the generalizability of our proposed method. Particularly for our model, we firstly pretrain the LSTM critic network on the training set. Then we fix the parameters of critic network and train the policy networ... |
Discourse Structurein Machine Translation Evaluation | 1710.01504 | Table 8: Kendall’s (τ) segment level correlation with human judgements on WMT12 obtained by the pairwise preference kernel learning. Results are presented for each language pair and overall. | ['[BOLD] Structure', 'cs-en', 'de-en', 'es-en', 'fr-en', 'Overall'] | [['Syntax', '0.190', '0.244', '0.198', '0.158', '0.198'], ['Discourse', '0.176', '0.235', '0.166', '0.160', '0.184'], ['Syntax+Discourse', '[BOLD] 0.210', '[BOLD] 0.251', '[BOLD] 0.240', '[BOLD] 0.223', '[BOLD] 0.231']] | As we can see, the τ scores of the syntactic and the discourse variants are not very different (with a general advantage for syntax), but when put together there is a sizeable improvement in correlation for all the language pairs and overall. This is clear evidence that the discourse-based features are providing additi... |
SubjQA: A Dataset for Subjectivity and Review Comprehension | 2004.14283 | Table 9: MTL gains/losses over the fine-tuning condition (F1 and Exact match), across subj./fact. QA. | ['[EMPTY]', '[BOLD] Fact. A [BOLD] F1', '[BOLD] Fact. A [BOLD] E', '[BOLD] Subj. A [BOLD] F1', '[BOLD] Subj. A [BOLD] E', '[BOLD] Fact. Q [BOLD] F1', '[BOLD] Fact. Q [BOLD] E', '[BOLD] Subj. Q [BOLD] F1', '[BOLD] Subj. Q [BOLD] E', '[BOLD] Overall [BOLD] F1', '[BOLD] Overall [BOLD] E'] | [['Tripadvisor', '17.50', '20.88', '1.28', '7.43', '18.85', '21.60', '1.16', '7.37', '1.01', '7.42'], ['Restaurants', '10.36', '12.38', '8.37', '11.49', '13.85', '15.77', '8.19', '11.07', '5.71', '8.65'], ['Movies', '14.49', '14.63', '5.17', '8.02', '14.27', '14.41', '5.44', '8.28', '3.08', '5.84'], ['Books', '13.95', ... | After fine-tuning over each domain in the MTL setting, the subjectivity-aware model achieves an average F1 of 76.3% across the different domains, with a minimum of 58.8% and a maximum of 82.0% on any given domain. Under both the F1 and the Exact match metrics, incorporating subjectivity in the model as an auxiliary tas... |
Deep Bayesian Network for Visual Question Generation | 2001.08779 | Table 2: Comparison with state-of-the-art and different combination of Cues. The first block consists of the SOTA methods, second block depicts the models which uses only a single type of information such as Image or Place, third block has models which take one cue along with the Image information, fourth block takes t... | ['[BOLD] Methods', '[BOLD] BLEU1 [BOLD] Max', '[BOLD] BLEU1 [BOLD] Avg', '[BOLD] METEOR [BOLD] Max', '[BOLD] METEOR [BOLD] Avg'] | [['Natural\xa0', '19.2', '-', '19.7', '-'], ['Creative\xa0', '35.6', '-', '19.9', '-'], ['MDN\xa0', '36.0', '-', '[BOLD] 23.4', '-'], ['Img Only (Bernoulli Dropout (BD))', '21.8', '19.57 ±2.5', '13.8', '13.45±1.52'], ['Place Only(BD)', '26.5', '25.36±1.14', '14.5', '13.60±0.40'], ['Cap Only (BD)', '27.8', '26.40 ±1.52'... | The first analysis is considering the various combinations of cues such as caption and place. We use these models as our baseline and compare other variations of our model with the best single cue. The third block takes into consideration one cue along with the Image information, and we see an improvement of around 4% ... |
Improving End-of-Turn Detection in Spoken Dialogues by Detecting Speaker Intentions as a Secondary Task | 1805.06511 | Table 4: Detecting turn-switches and turn-holds for each speaker intention class, as well as the per-class accuracy for detecting intentions by the auxiliary task. | ['[EMPTY]', '[BOLD] Turn-Transitions F1 (switch)', '[BOLD] Turn-Transitions F1 (hold)', '[BOLD] Intentions per-class Acc.'] | [['statement', '51.4', '73.2', '39.6'], ['opinion', '54.5', '70.3', '43.1'], ['agree', '49.7', '66.5', '30.4'], ['abandon', '67.2', '68.3', '56.8'], ['backchannel', '79.1', '51.6', '49.6'], ['question', '72.2', '47.6', '50.3'], ['answer', '61.1', '71.7', '43.9']] | The results show that the model is better able to predict turn-switches when presented with a backchannel or a question, and is better able to predict turn-holds when presented with a statement, opinion, or an answer. This suggests that the performance of the model depends on the context and nature of a dialogue, and t... |
Improving End-of-Turn Detection in Spoken Dialogues by Detecting Speaker Intentions as a Secondary Task | 1805.06511 | Table 3: Performance comparison of different methods. Results shown are macro-averages across turn-switches and turn-holds. | ['[BOLD] Method', 'Rec.', 'Prec.', 'F1', 'AUC'] | [['Random', '50.0', '49.6', '41.7', '45.3'], ['Full model\xa0', '55.8', '56.7', '55.4', '57.8'], ['LSTM', '65.9', '65.6', '65.5', '71.9'], ['MT-LSTM', '[BOLD] 66.4∗', '[BOLD] 66.0', '[BOLD] 65.8', '[BOLD] 72.6∗']] | We attribute this improvements to better feature representations and better sequential modeling abilities of LSTMs. The table shows that a multi-task LSTM, which is trained using a joint loss function, provides consistent improvements over a single-task LSTM (significant improvements under a paired t-test, p<0.05, in t... |
Generalization through Memorization: Nearest Neighbor Language Models | 1911.00172 | Table 3: Experimental results on Wiki-3B. The model trained on 100M tokens is augmented with a datastore that contains about 3B training examples, outperforming the vanilla LM trained on the entire Wiki-3B training set. | ['[BOLD] Training Data', '[BOLD] Datastore', '[BOLD] Perplexity (↓) Dev', '[BOLD] Perplexity (↓) Test'] | [['Wiki-3B', '-', '16.11', '15.17'], ['Wiki-100M', '-', '20.99', '19.59'], ['Wiki-100M', 'Wiki-3B', '14.61', '13.73']] | However, adding nearest neighbors retrieval over those 3B examples to the model trained on 100M tokens improves perplexity from 19.59 to 13.73; i.e. retrieving nearest neighbors from the corpus outperforms training on it. This result suggests that rather than training language models on ever larger datasets, we can use... |
Generalization through Memorization: Nearest Neighbor Language Models | 1911.00172 | Table 5: Wikitext-103 validation results using different states from the final layer of the LM as the representation function f(⋅) for keys and queries. We retrieve k=1024 neighbors and λ is tuned for each. | ['[BOLD] Key Type', '[BOLD] Dev ppl. (↓)'] | [['No datastore', '17.96'], ['Model output', '17.07'], ['Model output layer normalized', '17.01'], ['FFN input after layer norm', '[BOLD] 16.06'], ['FFN input before layer norm', '17.06'], ['MHSA input after layer norm', '16.76'], ['MHSA input before layer norm', '17.14']] | , we extract a representation of context c using an intermediate state of the LM f(c). While all the instantiations of f we tried are helpful, we achieved the largest improvement by using the input to the final layer’s feedforward network. We also observe that normalized representations (i.e. taken immediately after th... |
ACL’20 Temporal Common Sense Acquisition with Minimal Supervision | 2005.04304 | Table 1: Performance on intrinsic evaluations. The “normalized” row is the ratio of the distance to the gold label over the total number of labels in each dimension. Smaller is better. | ['[BOLD] Systems', '[BOLD] RealNews', '[BOLD] RealNews', '[BOLD] RealNews Typical Time', '[BOLD] RealNews Typical Time', '[BOLD] RealNews Typical Time', '[BOLD] RealNews Typical Time', '[BOLD] UDS-T'] | [['[EMPTY]', 'Duration', 'Freq', 'Day', 'Week', 'Month', 'Season', 'Duration'], ['BERT', '1.33', '1.68', '1.75', '1.53', '3.78', '0.87', '1.77'], ['BERT + naive finetune', '1.21', '1.45', '1.47', '1.28', '3.28', '1.08', '2.06'], ['TacoLM\xa0(ours)', '[BOLD] 0.75', '[BOLD] 1.17', '[BOLD] 1.72', '[BOLD] 1.19', '[BOLD] 3.... | We see that our proposed final model is mostly better than other variants, and achieves 19% improvement over BERT on average on the normalized scale. |
Phylogenetic signal in phonotactics1footnote 11footnote 1This article has been submitted but not yet accepted for publication in a book or journal. | 2002.00527 | Table 1: Summary of K analysis for forward and backward transition frequencies between different natural classes. The two rightmost columns indicate the total number of characters analysed and the percentage of those characters with a significant degree of phylogenetic signal according to the randomization procedure. | ['Classes', 'Mean K', 'n characters', 'significant (%)'] | [['Place', '0.63', '132', '100'], ['Major place', '0.64', '98', '83'], ['Manner', '0.63', '97', '80']] | All show highly similar distributions There is no statistically significant difference in the means of K for the three feature types, according to a one-way ANOVA (F(2,324)=0.05,p=0.95). An Anderson-Darling k-sample test, which tests the hypothesis that k independent samples come from a common, unspecified distribution... |
Learning an Executable Neural Semantic Parser | 1711.05066 | Table 9: Distantly supervised experimental results on the Spades dataset. | ['Models', 'F1'] | [['Unsupervised CCG Bisk et\xa0al. ( 2016 )', '24.8'], ['Semi-supervised CCG Bisk et\xa0al. ( 2016 )', '28.4'], ['Supervised CCG Bisk et\xa0al. ( 2016 )', '30.9'], ['Rule-based system Bisk et\xa0al. ( 2016 )', '31.4'], ['Sequence-to-sequence', '28.6'], ['tnsp, soft attention, top-down', '32.4'], ['tnsp, soft structured... | Previous work on this dataset has used a semantic parsing framework where natural language is converted to an intermediate syntactic representation and then grounded to Freebase. Specifically, Bisk et al. As can be seen, tnsp outperforms all CCG variants (from unsupervised to fully supervised) without having access to ... |
Learning an Executable Neural Semantic Parser | 1711.05066 | Table 6: Fully supervised experimental results on the GeoQuery dataset. For Jia and Liang (2016), we include two of their results: one is a standard neural sequence to sequence model; and the other is the same model trained with a data augmentation algorithm on the labeled data (reported in parentheses). | ['Models', 'Accuracy'] | [['Zettlemoyer and Collins ( 2005 )', '79.3'], ['Zettlemoyer and Collins ( 2007 )', '86.1'], ['Kwiatkowksi et\xa0al. ( 2010 )', '87.9'], ['Kwiatkowski et\xa0al. ( 2011 )', '88.6'], ['Kwiatkowski et\xa0al. ( 2013 )', '88.0'], ['Zhao and Huang ( 2015 )', '88.9'], ['Liang, Jordan, and Klein ( 2011 )', '91.1'], ['Dong and ... | The first block contains conventional statistical semantic parsers, previously proposed neural models are presented in the second block, whereas variants of tnsp are shown in the third block. We report accuracy which is defined as the proportion of utterances which correctly parsed to their gold standard logical forms.... |
Learning an Executable Neural Semantic Parser | 1711.05066 | Table 8: Breakdown of questions answered by type for the GraphQuestions. | ['Question type', 'Number', '% Answerable', '% Correctly answered'] | [['relation', '1938', '0.499', '0.213'], ['count', '309', '0.421', '0.032'], ['aggregation', '226', '0.363', '0.075'], ['filter', '135', '0.459', '0.096'], ['All', '2,608', '0.476', '0.173']] | GraphQuestions consists of four types of questions. An example of a relational question is what periodic table block contains oxygen; the second type contains count questions (denoted by count). An example is how many firefighters does the new york city fire department have; the third type includes aggregation question... |
Deep Conversational Recommender in Travel | 1907.00710 | TABLE I: Human evaluation results for different methods. | ['[BOLD] Method', '[BOLD] Fluency', '[BOLD] Informativeness', '[BOLD] Ranking'] | [['HRED', '2.64', '2.34', '3.08'], ['MultiWOZ', '2.74', '2.82', '2.7'], ['Mem2Seq', '3.04', '3.06', '2.3'], ['ReDial', '2.58', '2.62', '2.8'], ['TopicRNN', '3.64', '2.78', '2.66'], ['[BOLD] DCR', '[BOLD] 3.96', '[BOLD] 3.82', '[BOLD] 1.8']] | It directly reflects human perception of the quality of generated responses. The results show that the proposed DCR achieves the best performance across these various metrics, which indicates that the responses generated by it are more fluent and informative. We show that the performance improvements of DCR over the ot... |
Deep Conversational Recommender in Travel | 1907.00710 | TABLE III: Performance comparison of recommenders. | ['[BOLD] Methods', '[BOLD] ReDial', '[BOLD] NCF', '[BOLD] GCN-based'] | [['Top-1 Accuracy', '0.1065', '0.1882', '[BOLD] 0.2420']] | Since often only the top item is leveraged in the dialogs, we report the Top-1 accuracy here. It shows that the GCN-based recommender component achieves better performance as compared to ReDial and NCF methods. For the ReDial recommender, it projects the entity appearance vector v of each dialog session into a smaller ... |
Deep Ranking Based Cost-sensitive Multi-label Learning for Distant Supervision Relation Extraction | 1907.11521 | Table 1: The proportions of NR samples from Riedel’s dataset. | ['[BOLD] Pro. (%)', '[BOLD] Training', '[BOLD] Test'] | [['[BOLD] Riedel', '72.52', '96.26']] | In relation extraction, the dataset will always contain certain negative samples which do not express any relation types and are classified as NR type (no relation). In order to relieve this problem, we adopt cost-sensitive learning to construct the loss function. Based on G[att], the cost-sensitive loss function which... |
Latent Variable Sentiment Grammar Work was done when the first author was visiting Westlake University. The third author is the corresponding author. | 1907.00218 | Table 3: Experimental results with ELMo. BCN(P) is the BCN implemented by Peters et al. (2018). BCN(O) is the BCN implemented by ourselves. | ['Model', 'SST-5 Root', 'SST-5 Phrase', 'SST-2 Root', 'SST-2 Phrase'] | [['BCN(P)', '54.7', '-', '-', '-'], ['BCN(O)', '54.6', '83.3', '91.4', '88.8'], ['BCN+WG', '55.1', '[BOLD] 83.5', '91.5', '90.5'], ['BCN+LVG4', '55.5', '[BOLD] 83.5', '91.7', '91.3'], ['BCN+LVeG', '[BOLD] 56.0', '[BOLD] 83.5', '[BOLD] 92.1', '[BOLD] 91.6']] | There has also been work using large-scale external datasets to improve performances of sentiment classification. Peters et al. combined bi-attentive classification network (BCN, McCann et al. with a pretrained language model with character convolutions on a large-scale corpus (ELMo) and reported an accuracy of 54.7 on... |
Resolving Language and Vision Ambiguities Together: Joint Segmentation & Prepositional Attachment Resolution in Captioned Scenes | 1604.02125 | Table 1: Results on our subset of ABSTRACT-50S. | ['Module', 'Stanford Parser', 'Domain Adaptation', 'Ours', 'oracle'] | [['PPAR', '56.73', '57.23', '[BOLD] 77.39', '97.53']] | We performed a 10-fold cross-validation on the ABSTRACT-50S dataset to pick M (=10) and the weight on the hinge-loss for Mediator (C). This shows a need for diverse hypotheses and reasoning about visual features when picking a sentence parse. oracle denotes the best achievable performance using these 10 hypotheses. |
Resolving Language and Vision Ambiguities Together: Joint Segmentation & Prepositional Attachment Resolution in Captioned Scenes | 1604.02125 | Figure 5: Visualizations for 3 different prepositions (red = high scores, blue = low scores). We can see that our model has implicitly learned spatial arrangements unlike other spatial relation learning (SRL) works. | ['Feature set', 'PASCAL-50S Instance-Level Jaccard Index', 'PASCAL-50S PPAR Acc.', 'PASCAL-Context-50S PPAR Acc.'] | [['All features', '67.58', '80.33', '63.58'], ['Drop all consistency', '66.96', '66.67', '61.47'], ['Drop Euclidean distance', '67.27', '77.33', '63.77'], ['Drop directional distance', '67.12', '78.67', '63.63'], ['Drop word2vec', '67.58', '78.33', '62.72'], ['Drop category presence', '67.48', '79.25', '61.19']] | Visualizing Prepositions. These visualizations show the score obtained by taking the dot product of distance features (Euclidean and directional) between object1 and object2 connected by the preposition with the corresponding learned weights of the model, considering object2 to be at the center of the visualization. Th... |
OpenNMT: Open-Source Toolkit for Neural Machine Translation | 1701.02810 | Table 1: Performance numbers in source tokens per second for the Torch CPU/GPU implementations and for the multi-threaded CPU C implementation. (Run with Intel i7/GTX 1080) | ['Batch', 'Beam', 'GPU', 'CPU', 'CPU/C'] | [['1', '5', '209.0', '24.1', '62.2'], ['1', '1', '166.9', '23.3', '84.9'], ['30', '5', '646.8', '104.0', '116.2'], ['30', '1', '535.1', '128.5', '392.7']] | At deployment, the system is much less complex, and only requires (i) forwarding values through the network and (ii) running a beam search that is much simplified compared to SMT. OpenNMT includes several different translation deployments specialized for different run-time environments: a batched CPU/GPU implementation... |
Calibration of Pre-trained Transformers | 2003.07892 | Table 2: Out-of-the-box calibration results for in-domain (SNLI, QQP, SWAG) and out-of-domain (MNLI, TwitterPPDB, HellaSWAG) datasets using the models described in Table 1. We report accuracy and expected calibration error (ECE), both averaged across 5 runs with random restarts. | ['Model', 'Accuracy ID', 'Accuracy OD', 'ECE ID', 'ECE OD'] | [['[BOLD] Task: SNLI/MNLI', '[BOLD] Task: SNLI/MNLI', '[BOLD] Task: SNLI/MNLI', '[BOLD] Task: SNLI/MNLI', '[BOLD] Task: SNLI/MNLI'], ['DA', '84.63', '57.12', '[BOLD] 1.02', '8.79'], ['ESIM', '88.32', '60.91', '1.33', '12.78'], ['BERT', '90.04', '73.52', '2.54', '7.03'], ['RoBERTa', '[BOLD] 91.23', '[BOLD] 78.79', '1.93... | First, we analyze “out-of-the-box” calibration; that is, the calibration error derived from evaluating a model on a dataset without using post-processing steps like temperature scaling Guo et al. For each task, we train the model on the in-domain training set, and then evaluate its performance on the in-domain and out-... |
Multimodal Storytelling via Generative Adversarial Imitation Learning | 1712.01455 | Table 1: Similarity performance (T./I. denotes Text/Image respectively, T.I. means the combination of T. and I.) | ['Ran.', 'SS', 'PG', 'LSTM'] | [['-3860.02', '33714.67', '34009.27', '338876.34'], ['T.', 'I.', 'T.I.', 'MIL-GAN'], ['34263.59', '12483.20', '36143.34', '[BOLD] 36697.33']] | We evaluate our result with several established alternatives: random, scheduled sampling[bengio2015scheduled] and policy gradient with similarity. Unfortunately, the baselines do not share the same metric or objective function. Instead, they were compared regarding accumulative normalized similarity on the training set... |
Metric Learning for Dynamic Text Classification | 1911.01026 | (a) SENT | ['Model', '5', '10', '50', '500'] | [['MLP (un-tuned)', '29.5', '34.6', '40.3', '42.7'], ['EUC', '56.5', '65.6', '[BOLD] 74.2', '[BOLD] 79.8'], ['EUC (un-tuned)', '53.2', '56.5', '59.6', '60.8'], ['HYP', '[BOLD] 64.8', '[BOLD] 69.7', '72.9', '78.8'], ['HYP (un-tuned)', '60.1', '62.9', '65.4', '66.7']] | The fine-tuned model uses this data for both additional training and for constructing new prototypes. The un-tuned model constructs prototypes using the pretrained model’s representations without additional training. We also construct an un-tuned MLP baseline by fitting a nearest neighbor classifier (KNN, k=5) on the e... |
Metric Learning for Dynamic Text Classification | 1911.01026 | Table 1: Test accuracy for each dataset and method. Columns indicate the number of examples per label nfine used in the fine tuning stage. In all cases, the prototypical models outperform the baseline. The hyperbolic model performs best in the low data regime, but both metrics perform comparably when data is abundant. | ['Dataset', 'Model', '[ITALIC] nfine=5', '[ITALIC] nfine=10', '[ITALIC] nfine=20', '[ITALIC] nfine=100'] | [['[EMPTY]', 'MLP', '37.3±2.9', '43.8±3.5', '45.7±3.8', '57.4±3.5'], ['SENT', 'EUC', '39.6±6.4', '45.5±1.8', '47.7±4.7', '[BOLD] 62.7± [BOLD] 2.1'], ['[EMPTY]', 'HYP', '[BOLD] 42.2± [BOLD] 3.5', '[BOLD] 47.1± [BOLD] 4.8', '[BOLD] 53.0± [BOLD] 2.3', '[BOLD] 62.7± [BOLD] 2.2'], ['[EMPTY]', 'MLP', '49.2±1.0', '55.9±2.5', ... | S4SS0SSS0Px5 Results: The SENT dataset shows performance in the case where completely new labels are added during fine tuning. In the NEWS and WOS datasets new labels originate from the splits of old labels. This is consistent with our hypothesis (and previous work) that hyperbolic geometry is well suited for hierarchi... |
Metric Learning for Dynamic Text Classification | 1911.01026 | (a) SENT | ['Model', '5', '10', '20', '100'] | [['MLP (un-tuned)', '38.2', '46.7', '42.4', '46.3'], ['EUC', '39.6', '45.5', '47.7', '[BOLD] 62.7'], ['EUC (un-tuned)', '43.4', '51.2', '47.6', '55.8'], ['HYP', '42.2', '47.1', '53.0', '[BOLD] 62.7'], ['HYP (un-tuned)', '[BOLD] 45.7', '[BOLD] 52.4', '[BOLD] 53.3', '53.1']] | The fine-tuned model uses this data for both additional training and for constructing new prototypes. The un-tuned model constructs prototypes using the pretrained model’s representations without additional training. We also construct an un-tuned MLP baseline by fitting a nearest neighbor classifier (KNN, k=5) on the e... |
DIALOG-CONTEXT AWARE END-TO-END SPEECH RECOGNITION | 1808.02171 | Table 3: Word Error Rate (WER) on the Switchboard dataset. None of our experiments used any lexicon information or external text data other than the training transcription. The models were trained on 300 hours of Switchboard data only. | ['Train (∼ 300hrs)', 'CH', 'SWB'] | [['Models', 'WER', 'WER'], ['[ITALIC] [BOLD] sentence-level end2end', '[EMPTY]', '[EMPTY]'], ['Seq2Seq A2C ', '40.6', '28.1'], ['CTC A2C ', '31.8', '20.0'], ['CTC A2C ', '32.1', '19.8'], ['CTC A2W(Phone/external-LM init.) ', '23.6', '14.6'], ['[ITALIC] [BOLD] sentence-level end2end', '[EMPTY]', '[EMPTY]'], ['Our base... | Note that CTC A2W(Phone/external-LM init.) Our proposed model (a) performed best on SWB evaluation set showing 4.2% relative improvement over our baseline. Our proposed model (b) performed best on CH evaluation set showing 3.4% relative improvement over our baseline. |
BERT and PALs: Projected Attention Layers for Efficient Adaptation in Multi-Task Learning | 1902.02671 | Table 4: Performance on SQuAD and SWAG, in terms of average score across each task’s development set; this score is exact match and f1 score for SQuAD, and accuracy for SWAG. | ['Method', 'No.\xa0Params', 'New Layers', 'Round Robin'] | [['Shared', '1.00×', '0', '82.75±0.09'], ['Adding within BERT', '[EMPTY]', '[EMPTY]', '[EMPTY]'], ['PALs (204)', '1.13×', '12', '82.774±0.006'], ['Low Rank (100)', '1.13×', '12', '82.74±0.06']] | We tested multi-task learning with the SQuAD and SWAG datasets. We follow all the same experimental settings as before, but we use round robin sampling because of the comparable size of the datasets, and train for 24,000 steps, not 60,000, with an increased maximum sequence length, 256. However all approaches performed... |
BERT and PALs: Projected Attention Layers for Efficient Adaptation in Multi-Task Learning | 1902.02671 | Table 2: GLUE Test results, scored by the GLUE evaluation server. The number below each task denotes the number of training examples. We show F1/accuracy scores for QQP and MRPC, and accuracy on the matched/mismatched test sets for MNLI. The ‘Av.’ column is slightly different than the official GLUE score, since we excl... | ['Method', 'Params', 'MNLI-(m/mm)', 'QQP', 'QNLI', 'SST-2', 'CoLA', 'STS-B', 'MRPC', 'RTE', 'Av.'] | [['[EMPTY]', '[EMPTY]', '392k', '363k', '108k', '67k', '8.5k', '5.7k', '3.5k', '2.5k', '[EMPTY]'], ['BERT-base', '8×', '84.6/83.4', '89.2/71.2', '90.1', '93.5', '52.1', '85.8', '84.8/88.9', '66.4', '79.6'], ['Shared', '1.00×', '84.0/83.4', '88.9/70.8', '89.3', '93.4', '51.2', '83.6', '81.3/86.7', '76.6', '79.9'], ['Top... | Our main comparison is against fine-tuned BERT-base, which in the absence of transfer effects represents an upper bound on our performance, since it involves tuning all BERT-base parameters to perform well on each task individually, therefore requiring approximately 8× as many parameters as our methods. By construction... |
Dialogue Learning With Human-in-the-Loop | 1611.09823 | Table 5: Fully Supervised (Imitation Learning) Results on Human Questions | ['Train data size', '1k', '5k', '10k', '20k', '60k'] | [['Supervised MemN2N', '0.333', '0.429', '0.476', '0.526', '0.599']] | For comparison purposes, we also ran a supervised (imitation learning) MemN2N on different sized training sets of turker authored questions with gold annotated labels (so, there are no numerical rewards or textual feedback, this is a pure supervised setting). They indicate that RBI+FP and even FP alone get close to the... |
Dialogue Learning With Human-in-the-Loop | 1611.09823 | Table 2: Incorporating Feedback From Humans via Mechanical Turk. Textual feedback is provided for 10,000 model predictions (from a model trained with 1k labeled training examples), and additional sparse binary rewards (fraction r of examples have rewards). Forward Prediction and Reward-based Imitation are both useful, ... | ['Model', '[ITALIC] r=0', '[ITALIC] r=0.1', '[ITALIC] r=0.5', '[ITALIC] r=1'] | [['Reward Based Imitation (RBI)', '0.333', '0.340', '0.365', '0.375'], ['Forward Prediction (FP)', '0.358', '0.358', '0.358', '0.358'], ['RBI+FP', '0.431', '0.438', '0.443', '0.441']] | They indicate that both RBI and FP are useful. When rewards are sparse, FP still works via the textual feedback while RBI can only use the initial 1000 examples when r=0. As FP does not use numericalrewards at all, it is invariant to the parameter r. The combination of FP and RBI outperforms either alone. |
Dialogue Learning With Human-in-the-Loop | 1611.09823 | Table 4: Incorporating Feedback From Humans via Mechanical Turk: comparing real human feedback to synthetic feedback. Textual feedback is provided for 10,000 model predictions (from a model trained with 1k labeled training examples), and additional sparse binary rewards (fraction r of examples have rewards). We compare... | ['Model', '[ITALIC] r=0', '[ITALIC] r=0.1', '[ITALIC] r=0.5', '[ITALIC] r=1'] | [['Reward Based Imitation (RBI)', '0.333', '0.340', '0.365', '0.375'], ['Forward Prediction (FP) [real]', '0.358', '0.358', '0.358', '0.358'], ['RBI+FP [real]', '0.431', '0.438', '0.443', '0.441'], ['Forward Prediction (FP) [synthetic Task 2]', '0.188', '0.188', '0.188', '0.188'], ['Forward Prediction (FP) [synthetic T... | In the experiment in Section 5.2 we conducted experiments with real human feedback. Here, we compare this to a form of synthetic feedback, mostly as a sanity check, but also to see how much improvement we can get if the signal is simpler and cleaner (as it is synthetic). We hence constructed synthetic feedback for the ... |
Dialogue Learning With Human-in-the-Loop | 1611.09823 | Table 6: Second Iteration of Feedback Using synthetic textual feedback of synthetic Task2+3 with the RBI+FP method, an additional iteration of data collection of 10k examples, varying sparse binary reward fraction r and exploration ϵ. The performance of the first iteration model was 0.478. | ['[EMPTY]', '[ITALIC] r=0', '[ITALIC] r=0.1', '[ITALIC] r=0.5', '[ITALIC] r=1'] | [['[ITALIC] ϵ=0', '0.499', '0.502', '0.501', '0.502'], ['[ITALIC] ϵ=0.1', '0.494', '0.496', '0.501', '0.502'], ['[ITALIC] ϵ=0.25', '0.493', '0.495', '0.496', '0.499'], ['[ITALIC] ϵ=0.5', '0.501', '0.499', '0.501', '0.504'], ['[ITALIC] ϵ=1', '0.497', '0.497', '0.498', '0.497']] | We conducted experiments with an additional iteration of data collection for the case of binary rewards and textual feedback using the synthetic Task 2+3 mix. Using that model as a predictor, we collected an additional 10,000 training examples. We then continue to train our model using the original 1k+10k training set,... |
Noise Mitigation for Neural Entity Typing and Relation Extraction | 1612.07495 | Table 3: P@1, Micro F1 for all, head and tail entities and MAP results for entity typing. | ['[EMPTY]', '[EMPTY]', '[ITALIC] P@1 all', '[ITALIC] F1 all', '[ITALIC] F1 head', '[ITALIC] F1 tail', 'MAP'] | [['1', 'MLP', '74.3', '69.1', '74.8', '52.5', '42.1'], ['2', 'MLP+MIML-MAX', '74.7', '59.2', '50.7', '46.8', '41.3'], ['3', 'MLP+MIML-AVG', '77.2', '70.6', '74.9', '56.2', '45.0'], ['4', 'MLP+MIML-MAX-AVG', '75.2', '71.2', '76.4', '56.0', '47.1'], ['5', 'MLP+MIML-ATT', '81.0', '72.0', '76.9', '59.1', '48.8'], ['6', 'CN... | EntEmb and FIGMENT baselines. Following \newcitefigment15, we also learn entity embeddings and classify those embeddings to types, i.e., instead of distant supervision, we classify entities based on aggregated information represented in entity embeddings. An MLP with one hidden layer is used as classifier. We call that... |
Document Sub-structure in Neural Machine Translation | 1912.06598 | Table 5: Manual evaluation of 100 sentences per test set. Comparisons are classified in terms of the number of times the section-level translation is better, worse, equal or identical to the document-level output. | ['Lang. pair', 'Orig.', 'Better', 'Worse', 'Equal', 'Identical'] | [['Fr-En', 'En', '15', '14', '22', '49'], ['Fr-En', 'Fr', '13', '16', '33', '38'], ['Bg-En', 'En', '22', '17', '23', '38'], ['Bg-En', 'Bg', '24', '24', '32', '20'], ['Zh-En', 'En', '22', '9', '34', '35'], ['Zh-En', 'Zh', '9', '11', '65', '15']] | There is a further substantial number of sentences for which the two models achieve similar quality. Among the sentences that show a difference in quality between the two models, preference for either the section- or the document-level model depends on the original language of the set. Section-level models do better ac... |
Phoneme Classification in High-Dimensional Linear Feature Domains | 1312.6849 | TABLE II: Absolute reduction in percentage error for each of the classifiers (15)–(18) in quiet conditions. | ['[BOLD] Model', 'Waveform', 'PLP', 'PLP+Δ+ΔΔ'] | [['Model average (A [ITALIC] M)', '1.6', '2.8', '4.4'], ['[ITALIC] f-average (A [ITALIC] R)', '5.6', '6.0', '6.3'], ['Sector sum (A [ITALIC] S)', '6.7', '8.4', '8.7'], ['[ITALIC] f-average + sector sum (A [ITALIC] T)', '9.9', '10.0', '10.4']] | f. The relative benefits of the f-average and the sector sum are clear. The sector sum gives the bigger improvements on its own in all cases compared to only the f-average, but the combination of the two methods is better still throughout. The same qualitative trend holds true in noise. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.