Back

Supervised Learning of Universal Sentence Representations from Natural Language Inference Data

Alexis ConneauDouwe KielaHolger SchwenkLoïc BarraultAntoine Bordes

article2017EMNLP2,235 citationsOutstanding Paper Award

Demonstrates that training sentence encoders on supervised natural language inference data yields universal embeddings that consistently outperform unsupervised baselines across diverse transfer tasks, establishing inference as an effective pretraining objective for natural language processing.

Listen

Natural language processing systems have successfully used pre-trained word embeddings for various applications, but creating universal sentence representations that generalize across diverse tasks has remained a difficult challenge. Most existing sentence encoders rely on unsupervised learning over large text corpora, which often fail to achieve sufficient transfer performance or require prohibitive training resources. The article set out to evaluate whether supervised training on natural language inference tasks can produce superior, general-purpose sentence representations compared to unsupervised methods, and to identify the most effective neural network architecture for this purpose.

The authors conducted a comprehensive empirical evaluation comparing seven neural network encoder architectures—including recurrent models, attention networks, and convolutional designs—trained primarily on the Stanford Natural Language Inference dataset consisting of 570,000 human-labeled sentence pairs. To evaluate generalizability, the generated sentence representations were kept fixed and used as features across twelve diverse transfer tasks. These benchmarks encompassed binary and multi-class classification, semantic relatedness, textual similarity, paraphrase detection, and multimodal image-caption retrieval.

The analysis produced several significant findings. First, supervised training on natural language inference data consistently outperformed leading unsupervised models, such as SkipThought vectors, across the transfer benchmarks despite using significantly less data (570,000 sentence pairs versus 64 million unsupervised sentences). Second, training efficiency improved dramatically: the proposed supervised model trained in less than a day on a single graphics processing unit, compared to roughly a month for the best unsupervised baseline. Third, among the tested architectures, a bidirectional Long Short-Term Memory network utilizing max pooling achieved the best overall transfer performance. Fourth, semantic textual similarity correlation increased substantially, achieving a Pearson score of 0.68 on standard benchmark data compared to 0.44 for regularized SkipThought vectors. Finally, expanding the training corpus with the multi-genre MultiNLI dataset further improved transfer performance across the board without degrading domain-specific capabilities.

These findings demonstrate that supervised reasoning tasks like natural language inference force models to learn rich, universal semantic representations rather than superficial task biases. For organizations building natural language processing pipelines, adopting supervised sentence embeddings provides a path to higher model accuracy while drastically reducing the computational costs, training timelines, and infrastructure risks associated with massive unsupervised models. The results also show that higher accuracy on a primary training objective does not guarantee better transfer performance, highlighting the need to carefully select architectures that retain general semantic features rather than over-specializing.

Organizations should consider using supervised inference-trained sentence encoders—particularly bidirectional recurrent models with max pooling—as standard baseline feature extractors for downstream language tasks. Future work should focus on scaling up the diversity and size of natural language understanding datasets to further advance representation quality. A key limitation noted in the article is that the evaluation focused on English text and established benchmark tasks; practitioners should test these encoders on their specific operational domains before full-scale deployment to ensure robust generalization.

MetaUniversité du Mainecomputer visioninferenceprobabilistic inferenceSentenceLearningNatural Language
arXiv: 1705.02364facebookresearch/InferSent
  • Paper: Skip-Thought Vectors, Ryan Kiros et al. (2015). Reading the foundational Skip-Thought vectors paper clarifies the unsupervised sentence representation baseline that the source paper surpasses using supervised natural language inference data.
  • Paper: A large annotated corpus for learning natural language inference, Samuel R. Bowman et al. (2015). This paper introduces the Stanford Natural Language Inference corpus, which serves as the core training data foundation utilized by the source paper to learn universal sentence representations.
Cover for Supervised Learning of Universal Sentence Representations from Natural Language Inference Data

Abstract

Many modern NLP systems rely on word embeddings, previously trained in an unsupervised manner on large corpora, as base features. Efforts to obtain embeddings for larger chunks of text, such as sentences, have however not been so successful. Several attempts at learning unsupervised representations of sentences have not reached satisfactory enough performance to be widely adopted. In this paper, we show how universal sentence representations trained using the supervised data of the Stanford Natural Language Inference datasets can consistently outperform unsupervised methods like SkipThought vectors on a wide range of transfer tasks. Much like how computer vision uses ImageNet to obtain features, which can then be transferred to other tasks, our work tends to indicate the suitability of natural language inference for transfer learning to other NLP tasks. Our encoder is publicly available.

Topics

computer visioninferenceprobabilistic inferencesupervised machine learningtransfer learningword embeddingunsupervised learning methodnatural language understandingquestion answeringsyntactic structure

Keywords

SentenceLearningNatural LanguageRepresentationsNatural Language InferenceTrainedSupervisedDataTaskTasks

Table of Contents

  • 1 Introduction
  • 2 Related work
  • 3 Approach
  • 3.1 The Natural Language Inference task
  • 3.2 Sentence encoder architectures
  • 3.2.1 LSTM and GRU
  • 3.2.2 BiLSTM with mean/max pooling
  • 3.2.3 Self-attentive network
  • 3.2.4 Hierarchical ConvNet
  • 3.3 Training details
  • 4 Evaluation of sentence representations
  • 5 Empirical results
  • 5.1 Architecture impact
  • 5.2 Task transfer
  • 6 Conclusion
  • References

Knowls

  1. Knowl 1 — Natural Language Inference Siamese Sentence Encoding Framework

    model/method

    The supervised sentence encoding framework (InferSent) trains a universal sentence encoder on a natural language inference (NLI) dataset such as the Stanford Natural Language Inference (SNLI) corpus. The framework employs a shared sentence encoder fθf_\theta to encode a premise sentence into vector u∈Rdu \in \mathbb{R}^d and a hypothesis sentence into vector v∈Rdv \in \mathbb{R}^d.

    To capture semantic relations, the sentence vectors are combined using three vector operations to form a matching representation:

    m=[u;v;∣u−v∣;u∗v]m = [u; v; |u - v|; u * v]

    where [u;v][u; v] denotes vector concatenation, ∣u−v∣|u - v| is the absolute element-wise difference, and u∗vu * v is the element-wise product. The resulting vector m∈R4dm \in \mathbb{R}^{4d} is passed into a 3-class classifier composed of a multi-layer perceptron with one hidden layer of 512 units and a 3-way softmax output corresponding to entailment, contradiction, and neutral classes.

    Input words are initialized with fixed 300-dimensional GloVe embeddings trained on Common Crawl 840B. The network is optimized using stochastic gradient descent (SGD) with an initial learning rate of 0.10.1, a weight decay of 0.990.99, and mini-batches of size 64. The learning rate is divided by 5 each time the development set accuracy drops, and training terminates when the learning rate falls below 10−510^{-5}.

  2. Knowl 2 — Bi-directional LSTM with Max Pooling Sentence Encoder

    model/method

    The BiLSTM with max-pooling encoder processes a sequence of TT words (w1,…,wT)(w_1, \dots, w_T) using a forward and backward Long Short-Term Memory (LSTM) network. At each time step t∈{1,…,T}t \in \{1, \dots, T\}, hidden state vectors are computed in both directions:

    h⃗t=LSTM→t(w1,…,wT)\vec{h}_t = \overrightarrow{\text{LSTM}}_t(w_1, \dots, w_T)

    h←t=LSTM←t(w1,…,wT)\overleftarrow{h}_t = \overleftarrow{\text{LSTM}}_t(w_1, \dots, w_T)

    The directional hidden states are concatenated to form the recurrent hidden state at time step tt:

    ht=[h⃗t,h←t]h_t = [\vec{h}_t, \overleftarrow{h}_t]

    To compress the variable-length sequence of hidden states {ht}t=1T\{h_t\}_{t=1}^T into a fixed-size sentence representation u∈Rdu \in \mathbb{R}^d, a max-pooling operation is applied across the time dimension for each feature component k∈{1,…,d}k \in \{1, \dots, d\}:

    uk=max⁡t∈{1,…,T}ht,ku_k = \max_{t \in \{1, \dots, T\}} h_{t, k}

    This pooling mechanism selects the most salient activation across the entire sequence for each feature dimension.

  3. Knowl 3 — Downstream Transfer Benchmark Results for NLI-Trained Sentence Encoders

    data/table

    Sentence representations produced by the BiLSTM with max pooling encoder trained on NLI data outperform previous unsupervised and supervised sentence representation approaches across a broad suite of 10 transfer tasks, while requiring less training time (under one day on a single GPU compared to weeks or a month for models like SkipThought).

    Model MR CR SUBJ MPQA SST TREC MRPC SICK-R SICK-E STS14
    Unsupervised representation training
    FastSent 70.8 78.4 88.7 80.6 - 76.8 72.2/80.3 - - .63/.64
    FastSent+AE 71.8 76.7 88.8 81.5 - 80.4 71.2/79.1 - - .62/.62
    SkipThought 76.5 80.1 93.6 87.1 82.0 92.2 73.0/82.0 0.858 82.3 .29/.35
    SkipThought-LN 79.4 83.1 93.7 89.3 82.9 88.4 - 0.858 79.5 .44/.45
    GloVe BOW 78.7 78.5 91.6 87.6 79.8 83.6 72.1/80.9 0.800 78.6 .54/.56
    BiLSTM-Max (untrained) 77.5 81.3 89.6 88.7 80.7 85.8 73.2/81.6 0.860 83.4 .39/.48
    Supervised representation training
    DictRep (bow) 76.7 78.7 90.7 87.2 - 81.0 68.4/76.8 - - .67/.70
    NMT En-to-Fr 64.7 70.1 84.9 81.5 - 82.8 - - - .43/.42
    Paragram-phrase - - - - 79.7 - - 0.849 83.1 .71/ -
    BiLSTM-Max (on SST) - 83.7 90.2 89.5 - 86.0 72.7/80.9 0.863 83.1 .55/.54
    BiLSTM-Max (on SNLI) 79.9 84.6 92.1 89.8 83.3 88.7 75.1/82.3 0.885 86.3 .68/.65
    BiLSTM-Max (on AllNLI) 81.1 86.3 92.4 90.2 84.6 88.2 76.2/83.1 0.884 86.3 .70/.67

    MRPC scores are reported as accuracy/F1. SICK-R and STS14 report Pearson (and Spearman for STS14) correlation coefficients. SICK-E and the classification benchmarks (MR, CR, SUBJ, MPQA, SST, TREC) report accuracy.

  4. Knowl 4 — Comparison of Sentence Encoder Architectures on NLI and Transfer Tasks

    data/table

    Different neural sentence encoder architectures trained under the identical NLI objective on the SNLI dataset exhibit distinct transfer capabilities across downstream classification and relatedness benchmarks.

    Model dim NLI dev (%) NLI test (%) Transfer micro (%) Transfer macro (%)
    LSTM 2048 81.9 80.7 79.5 78.6
    GRU 4096 82.4 81.8 81.7 80.9
    BiGRU-last 4096 81.3 80.9 82.9 81.7
    BiLSTM-Mean 4096 79.0 78.2 83.1 81.7
    Inner-attention 4096 82.3 82.5 82.1 81.0
    HConvNet 4096 83.7 83.4 82.0 80.9
    BiLSTM-Max 4096 85.0 84.5 85.2 83.7

    Transfer macro is the unweighted average of validation accuracies on transfer tasks; transfer micro is the accuracy sum weighted by development set size. BiLSTM-Max achieves the best performance both in-domain and on transfer tasks. Notably, performance on the training task does not strictly correlate with transfer performance (e.g., BiLSTM-Mean achieves lower NLI accuracy than Inner-attention or HConvNet but higher transfer scores), showing that models that over-focus on task-specific biases can sacrifice general semantic representation quality.

  5. Knowl 5 — Self-Attentive and Hierarchical Convolutional Sentence Encoder Architectures

    model/method

    Beyond standard recurrent encoders, two structured neural architectures were evaluated for universal sentence representation:

    1. Self-attentive sentence encoder (Inner-attention): Uses an attention mechanism over the hidden vectors {h1,…,hT}\{h_1, \dots, h_T\} produced by a BiLSTM. Keys hˉi\bar{h}_i and attention weights αi\alpha_i are computed with learned parameters WW, bwb_w, and a context query vector uwu_w:

    hˉi=tanh⁡(Whi+bw)\bar{h}_i = \tanh(W h_i + b_w)

    αi=exp⁡(hˉi⊤uw)∑jexp⁡(hˉj⊤uw)\alpha_i = \frac{\exp(\bar{h}_i^\top u_w)}{\sum_j \exp(\bar{h}_j^\top u_w)}

    u=∑t=1Tαthtu = \sum_{t=1}^T \alpha_t h_t

    To capture multiple semantic aspects, four distinct context vectors (uw1,uw2,uw3,uw4)(u_w^1, u_w^2, u_w^3, u_w^4) are used in parallel, and their corresponding output representations are concatenated to form the final sentence vector.

    1. Hierarchical ConvNet (HConvNet): Comprises a sequence of 4 convolutional layers. At each layer i∈{1,2,3,4}i \in \{1, 2, 3, 4\}, a layer representation uiu_i is extracted via max-pooling across the spatial feature maps. The final sentence embedding concatenates representations from all four abstraction levels: u=[u1,u2,u3,u4]u = [u_1, u_2, u_3, u_4].

  6. Knowl 6 — Sentence Representations on Image-Caption and Caption-Image Retrieval

    data/table

    Pre-trained sentence representations evaluate multimodal alignment via the MS COCO caption-image and image-caption retrieval benchmark. Given an image embedding y∈R2048y \in \mathbb{R}^{2048} (from a pre-trained ResNet-101) and caption sentence embedding xx, linear transformations VV and UU project them to a shared embedding space optimized via a pairwise ranking loss:

    Lcir(x,y)=∑y∑kmax⁡(0,α−s(Vy,Ux)+s(Vy,Uxk))+∑x∑k′max⁡(0,α−s(Ux,Vy)+s(Ux,Vyk′))L_{cir}(x, y) = \sum_y \sum_k \max(0, \alpha - s(Vy, Ux) + s(Vy, Ux_k)) + \sum_x \sum_{k'} \max(0, \alpha - s(Ux, Vy) + s(Ux, Vy_{k'}))

    where s(⋅,⋅)s(\cdot, \cdot) is cosine similarity, α=0.2\alpha = 0.2 is the margin, and k,k′k, k' index 30 contrastive negative samples.

    Caption Retrieval Image Retrieval
    Model R@1 R@5 R@10 Med r R@1 R@5 R@10 Med r
    Direct supervision of sentence representations
    m-CNN 38.3 - 81.0 2 27.4 - 79.5 3
    m-CNNENS 42.8 - 84.1 2 32.6 - 82.8 3
    Order-embeddings 46.7 - 88.9 2 37.9 - 85.9 2
    Pre-trained sentence representations
    SkipThought + VGG19 (82k) 33.8 67.7 82.1 3 25.9 60.0 74.6 4
    SkipThought + ResNet101 (113k) 37.9 72.2 84.3 2 30.6 66.2 81.0 3
    BiLSTM-Max (SNLI) + ResNet101 42.4 76.1 87.0 2 33.2 69.7 83.6 3
    BiLSTM-Max (AllNLI) + ResNet101 42.6 75.3 87.3 2 33.9 69.7 83.8 3

    Results are averaged over 5 random splits of 1k test images from MS COCO. Pre-trained BiLSTM-Max representations transfer competitively against systems trained directly end-to-end on image-caption retrieval.

  7. Knowl 7 — MultiGenre NLI Pretraining Enhances Transfer Generalization

    empirical result

    Training the BiLSTM with max-pooling encoder on AllNLI—the combination of the 570k-pair SNLI dataset and the 433k-pair MultiNLI corpus spanning ten distinct genres of spoken and written English—improves transfer performance across downstream benchmarks compared to training on SNLI alone.

    On customer reviews (CR), accuracy improves from 84.6% to 86.3%; on movie reviews (MR), accuracy increases from 79.9% to 81.1%; on SST sentiment classification, accuracy rises from 83.3% to 84.6%; and on the unsupervised semantic similarity STS14 benchmark, the Pearson correlation increases from 0.68 to 0.70.

    Furthermore, adding diverse non-caption genres from MultiNLI does not degrade performance on image-caption retrieval tasks (improving COCO image retrieval Recall@1 from 33.2% to 33.9%), demonstrating that genre diversity broadens representation robustness without diluting specific semantic domains.

  8. Knowl 8 — SentEval Downstream Sentence Representation Evaluation Protocol

    experimental setup

    SentEval evaluates fixed generic sentence embeddings across 12 downstream NLP transfer tasks without updating the sentence encoder parameters:

    1. Binary and multi-class classification: For single-sentence classification tasks (MR, CR, SUBJ, MPQA, SST, TREC), a logistic regression classifier is trained on top of the fixed sentence representations. Optimization is conducted using Adam with a batch size of 64, and the L2L_2 regularization penalty is selected via grid search on the validation set.
    2. Sentence pair classification and semantic relatedness: For pair tasks (SICK-E, SICK-R, MRPC), the matching feature vector m=[u;v;∣u−v∣;u∗v]m = [u; v; |u - v|; u * v] is formed from the representations uu and vv of the two sentences. A logistic regression classifier is trained for classification tasks (SICK-E, MRPC), and a model predicting the relatedness probability distribution is used for SICK-R to compute Pearson correlation.
    3. Unsupervised Semantic Textual Similarity (STS14): Cosine similarities between sentence embeddings are evaluated directly against human similarity ratings across 6 sub-tasks using Pearson and Spearman correlation without any task-specific parameter training.

  9. Knowl 9 — Impact of Optimization Choice and Embedding Dimensionality on Generalization

    empirical result

    The transferability of sentence representations is influenced by both the training optimizer and the embedding dimensionality:

    1. Optimizer Sensitivity: When training the BiLSTM-Max encoder on SNLI using Adam instead of SGD, the network converged in fewer epochs (5 epochs versus 10 epochs), but yielded worse generalization on downstream transfer tasks. SGD avoids the excessive over-specialization to training set biases that occurs with Adam.
    2. Embedding Dimensionality: Increasing the sentence embedding dimension from 512 to 4096 dimensions consistently improves downstream transfer performance across architectures when evaluated with linear classifiers. Architectures such as BiLSTM-Max, Hierarchical ConvNet, and Inner-attention show the largest performance gains as dimensionality increases, indicating that higher-dimensional spaces allow these encoders to preserve broader semantic information that is not strictly required for the NLI objective but beneficial for transfer tasks.

Coverage note — None was omitted; all key contributions including NLI pretraining formulation, all evaluated encoder architectures, transfer benchmark results on SentEval, MS COCO multimodal retrieval, and optimization analyses are fully covered.

References

  1. 1.Eneko Agirre, Carmen Banea, Claire Cardie, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Rada Mihalcea, German Rigau, and Janyce Wiebe. 2014. Semeval-2014 task 10: Multilingual semantic textual similarity. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 81–91, Dublin, Ireland. Association for Computational Linguistics and Dublin City University.
  2. 2.Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. 2015. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425–2433.
  3. 3.Sanjeev Arora, Yingyu Liang, and Tengyu Ma. 2017. A simple but tough-to-beat baseline for sentence embeddings. Proceedings of the 5th International Conference on Learning Representations (ICLR).
  4. 4.Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normalization. Advances in neural information processing systems (NIPS).
  5. 5.Yoshua Bengio, Rejean Ducharme, and Pascal Vincent. 2003. A neural probabilistic language model. Journal of Machine Learning Research, 3:1137–1155.
  6. 6.Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP).
  7. 7.Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. In Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation (SSST-8).
  8. 8.Ronan Collobert and Jason Weston. 2008. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th international conference on Machine learning, pages 160–167. ACM.
  9. 9.Ronan Collobert, Jason Weston, Leon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. 2011. Natural language processing (almost) from scratch. Journal of Machine Learning Research, 12(Aug):2493–2537.
  10. 10.Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 248–255. IEEE.
  11. 11.Juri Ganitkevitch, Benjamin Van Durme, and Chris Callison-Burch. 2013. PPDB: The paraphrase database. In Proceedings of NAACL-HLT, pages 758–764, Atlanta, Georgia. Association for Computational Linguistics.
  12. 12.Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), page 8. IEEE.
  13. 13.Felix Hill, Kyunghyun Cho, and Anna Korhonen. 2016. Learning distributed representations of sentences from unlabelled data. arXiv preprint arXiv:1602.03483.
  14. 14.Sepp Hochreiter and Jurgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780.
  15. 15.Miach Hodosh, Peter Young, and Julia Hockenmaier. 2013. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research, 47:853–899.
  16. 16.Yangfeng Ji and Jacob Eisenstein. 2013. Discriminative improvements to distributional sentence similarity. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP).
  17. 17.Andrej Karpathy and Li Fei-Fei. 2015. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3128–3137.
  18. 18.Diederik P. Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR).
  19. 19.Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Skip-thought vectors. In Advances in neural information processing systems (NIPS), pages 3294–3302.
  20. 20.Alice Lai and Julia Hockenmaier. 2014. Illinois-lh: A denotational and distributional approach to semantics. Proc. SemEval, 2:5.
  21. 21.Quoc V Le and Tomas Mikolov. 2014. Distributed representations of sentences and documents. In Proceedings of the 31st International Conference on Machine Learning, volume 14, pages 1188–1196.
  22. 22.Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740–755. Springer International Publishing.
  23. 23.Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, and Yoshua Bengio. 2017. A structured self-attentive sentence embedding. International Conference on Learning Representations (ICLR).
  24. 24.Etai Littwin and Lior Wolf. 2016. The multiverse loss for robust transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3957–3966.
  25. 25.Yang Liu, Chengjie Sun, Lei Lin, and Xiaolong Wang. 2016. Learning natural language inference using bidirectional lstm model and inner-attention. arXiv preprint arXiv:1605.09090.
  26. 26.Lin Ma, Zhengdong Lu, Lifeng Shang, and Hang Li. 2015. Multimodal convolutional neural networks for matching image and sentence. In Proceedings of the IEEE International Conference on Computer Vision, pages 2623–2631.
  27. 27.Marco Marelli, Stefano Menini, Marco Baroni, Luisa Bentivogli, Raffaella Bernardi, and Roberto Zamparelli. 2014. A sick cure for the evaluation of compositional distributional semantic models. In Proceedings of the 9th International Conference on Language Resources and Evaluation (LREC), pages 216–223.
  28. 28.Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems (NIPS), pages 3111–3119.
  29. 29.Lili Mou, Zhao Meng, Rui Yan, Ge Li, Yan Xu, Lu Zhang, and Zhi Jin. 2016. How transferable are neural networks in nlp applications? Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP).
  30. 30.Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), volume 14, pages 1532–1543.
  31. 31.Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. 2014. Cnn features off-the-shelf: an astounding baseline for recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 806–813.
  32. 32.Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pages 3104–3112.
  33. 33.Kai Sheng Tai, Richard Socher, and Christopher D Manning. 2015. Improved semantic representations from tree-structured long short-term memory networks. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics (ACL).
  34. 34.Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. 2014. Deepface: Closing the gap to human-level performance in face verification. In Conference on Computer Vision and Pattern Recognition (CVPR), page 8. IEEE.
  35. 35.Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. 2016. Order-embeddings of images and language. International Conference on Learning Representations (ICLR).
  36. 36.John Wieting, Mohit Bansal, Kevin Gimpel, and Karen Livescu. 2016a. Charagram: Embedding words and sentences via character n-grams. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP).
  37. 37.John Wieting, Mohit Bansal, Kevin Gimpel, and Karen Livescu. 2016b. Towards universal paraphrastic sentence embeddings. International Conference on Learning Representations (ICLR).
  38. 38.Adina Williams, Nikita Nangia, and Samuel R Bowman. 2017. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426.
  39. 39.Han Zhao, Zhengdong Lu, and Pascal Poupart. 2015. Self-adaptive hierarchical sentence model. In Proceedings of the 24th International Conference on Artificial Intelligence, IJCAI'15, pages 4069–4076. AAAI Press.
  40. 40.Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pages 19–27.

Citation

MLA
Conneau, A., et al. “Supervised Learning of Universal Sentence Representations from Natural Language Inference Data”. Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, 2017, pp. 670–80, https://doi.org/10.18653/v1/D17-1070.
APA
Conneau, A., Kiela, D., Schwenk, H., Barrault, L., & Bordes, A. (2017). Supervised Learning of Universal Sentence Representations from Natural Language Inference Data. Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, 670–680. https://doi.org/10.18653/v1/D17-1070
Chicago
Conneau, A., D. Kiela, H. Schwenk, L. Barrault, and A. Bordes. 2017. “Supervised Learning of Universal Sentence Representations from Natural Language Inference Data”. Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, 670–80. https://doi.org/10.18653/v1/D17-1070.
Harvard
Conneau, A. et al. (2017) “Supervised Learning of Universal Sentence Representations from Natural Language Inference Data”, Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp. 670–680. Available at: https://doi.org/10.18653/v1/D17-1070.
Vancouver
1. Conneau A, Kiela D, Schwenk H, Barrault L, Bordes A (2017) Supervised Learning of Universal Sentence Representations from Natural Language Inference Data. In: Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp 670–680

BibTeX

@inproceedings{Conneau_2017, title={Supervised Learning of Universal Sentence Representations from
            Natural Language Inference Data}, url={http://dx.doi.org/10.18653/v1/D17-1070}, DOI={10.18653/v1/d17-1070}, booktitle={Proceedings of the 2017 Conference on Empirical Methods in Natural
          Language Processing}, publisher={Association for Computational Linguistics}, author={Conneau, Alexis and Kiela, Douwe and Schwenk, Holger and Barrault, Loïc and Bordes, Antoine}, year={2017}, pages={670–680} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/