Multi-view learning methods provide the ability to extract information from different views of the data and enable self-supervised learning of useful features for future prediction when annotated data is not available(de Sa, 1993). Minimising the disagreement among multiple views helps the model to learn rich feature representations of the data and, also after training, the ensemble of the feature vectors from multiple views can provide an even stronger generalisation ability.
consolidated this idea by stating that the meaning of a word can be determined by the company it has. The hypothesis has been widely used in machine learning community to learn vector representations of human languages. Models built upon distributional similarity don’t explicitly require human-annotated training data; the supervision comes from the semantic continuity of language data, such as text and speech.
Large quantities of annotated data are usually hard to obtain. Our goal is to propose learning algorithms built upon the ideas of multi-view learning and distributional hypothesis to learn from unlabelled data. We draw inspiration from the lateralisation and asymmetry in information processing of the two hemispheres of the human brain where, for most adults, sequential processing dominates the left hemisphere, and the right hemisphere has a focus on parallel processing (Bryden, 2012), but both hemispheres have been shown to have roles in literal and non-literal language comprehension (Coulson et al., 2005; Coulson & van Petten, 2007).
Our proposed multi-view frameworks aim to leverage the functionality of both RNN-based models, which have been widely applied in sentiment analysis tasks(Yang et al., 2016), and the linear/log-linear models, which have excelled at capturing attributional similarities of words and sentences (Arora et al., 2016, 2017; Hill et al., 2016; Turney & Pantel, 2010) for learning sentence representations. Previous work on unsupervised sentence representation learning based on distributional hypothesis can be roughly categorised into two types:
Generative objective: These models generally follow the encoder-decoder structure. The encoder learns to produce a vector representation for the current input, and the decoder learns to generate sentences in the adjacent context given the produced vector (Kiros et al., 2015; Hill et al., 2016; Gan et al., 2017; Tang et al., 2018). The idea is straightforward, yet its scalability for very large corpora is hindered by the slow decoding process that dominates training time, and also the decoder in each model is discarded after learning as the quality of generated sequences is not the main concern, which is a waste of parameters and learning effort.
Our first multi-view framework has a generative objective and uses an RNN as the encoder and an invertible linear projection as the decoder. The training time is drastically reduced as the decoder is simple, and the decoder is also utilised after learning. A regularisation is applied on the linear decoder to enforce invertibility, so that after learning, the inverse of the decoder can be applied as a linear encoder in addition to the RNN encoder.
In these models, a classifier is learnt on top of the encoders to distinguish adjacent sentences from those that are not(Li & Hovy, 2014; Jernite et al., 2017; Nie et al., 2017; Logeswaran & Lee, 2018); these models make a prediction using a predefined differential similarity function on the representations of the input sentence pairs or triplets.
Our second multi-view framework has a discriminative objective and uses an RNN encoder and a linear encoder; it learns to maximise agreement among adjacent sentences. Compared to earlier work on multi-view learning (de Sa, 1993; Dhillon et al., 2011) that takes data from various sources or splits data into disjoint populations, our framework processes the exact same input data in two distinctive ways. Having two distinctive information processing views encourages the model to encode different aspects of an input sentence, and is beneficial to the future use of the learnt representations.
Our contribution is threefold:
Two multi-view frameworks for learning sentence representations are proposed, in which one framework uses a generative objective and the other one adopts a discriminative objective. Two encoding functions, an RNN and a linear model, are learnt in both frameworks.
The results show that in both frameworks, the ensemble of two views provide better results than each view alone, and especially, aligning two views by the invertible contraint in the generative approach and by maximising the agreement in the discriminative objective improves each view in our efficiently trained multi-view frameworks.
Models trained under our proposed frameworks achieve good performance on the unsupervised tasks, and overall outperform existing unsupervised transfer learning models, and armed with various pooling functions, they also show solid results on supervised tasks, which are either comparable to or better than those of the best unsupervised transfer model.
It is shown in Hill et al. (2016) that the consistency between supervised and unsupervised evaluation tasks is much lower than that within either supervised or unsupervised evaluation tasks alone and that a model that performs well on supervised evaluation tasks may fail on unsupervised tasks. Conneau et al. (2017) subsequently showed that, with a labelled training corpus, such as SNLI (Bowman et al., 2015) and MultiNLI (Williams et al., 2017), the resulting representations of the sentences from the trained model excel in both supervised and unsupervised tasks. Our model is able to achieve good results on both groups of tasks without labelled information.
2 Model Architecture
Our goal is to marry RNN-based sentence encoder and the avg-on-word-vectors sentence encoder into multi-view frameworks with simple training objectives.
The motivation for the idea is that, as mentioned in the prior work, RNN-based encoders process the sentences sequentially, and are able to capture complex syntactic interactions, while the avg-on-word-vectors encoder has been shown to be good at capturing the coarse meaning of a sentence which could be useful for finding paradigmatic parallels (Turney & Pantel, 2010).
We present two multi-view frameworks to learn two different sentence encoders; after learning, the vectors produced from two encoders of the same input sentence are used to compose the sentence representation. The details of our learning frameworks are described as follows:
In our multi-view frameworks, we first introduce two encoders that, after learning, can be used to build sentence representations. One encoder is a bi-directional Gated Recurrent Unit(Chung et al., 2014) , where is the input sentence and is the parameter vector in the GRU. During learning, only hidden state at the last time step is sent to the next stage in learning. The other encoder is a linear avg-on-word-vectors model , which basically transforms word vectors in a sentence by a learnable weight matrix and outputs an averaged vector.
2.2 Generative Objective
Given the finding (Tang et al., 2018) that neither an autoregressive nor an RNN decoder is necessary for learning sentence representations that excel on downstream tasks, our learning framework only learns to predict words in the next sentence. The framework has an RNN encoder , and a linear decoder . Given an input sentence , the encoder produces a vector , and the decoder projects the vector to , which has the same dimension as the word vectors . Negative sampling is applied to calculate the likelihood of generating the -th word in the -th sentence, shown in Eq. 1.
where are pretrained word vectors for , the empirical distribution is the unigram distribution raised to power 0.75 (Mikolov et al., 2013), and is the number of negative samples. The learning objective is to maximise the likelihood for words in all sentences in the training corpus.
Ideally, the inverse of should be easy to compute so that during testing we can set . As is a linear projection, the simplest situation is when
is an orthogonal matrix and its inverse is equal to its transpose. Often, as the dimensionality of vectordoesn’t necessarily need to match that of word vectors , is not a square matrix111As often the dimension of sentence vectors are equal to or large than that of word vectors, has more columns than rows. If it is not the case, then regulariser becomes .. To enforce invertibility on , a row-wise orthonormal regularisation on is applied during training, which leads to , where
is the identity matrix, thus the inverse function is simply, which is easily computed. The regularisation formula is , where is the Frobenius norm. Specifically, the update rule (Cissé et al., 2017) for the regularisation is:
where is set to . After learning, we set , then the inverse of the decoder becomes the encoder . Compared to prior work with generative objective, our framework reuses the decoding function rather than ignoring it for building sentence representations after learning, thus information encoded in the decoder is also utilised.
2.3 Discriminative Objective
Our multi-view framework with discriminative objective learns to maximise the agreement between the representations of a sentence pair across two views if one sentence in the pair is in the neighbourhood of the other one. An RNN encoder and a linear avg-on-word-vectors produce a vector representation and for -th sentence respectively. The agreement between two views of a sentence pair is defined as
. The training objective is to minimise the loss function:
where is the trainable temperature term, which is essential for exaggerating the difference between adjacent sentences and those that are not. The neighbourhood/context window , and the batch size
The choice of cosine similarity based loss is based on the observations inTurney & Pantel (2010) that, of word vectors derived from distributional similarity, vector length tends to correlate with frequency of words, thus angular distance captures more important meaning-related information. Also, since our model is unsupervised/self-supervised, whatever similarity there is between neighbouring sentences is what is learnt as important for meaning.
The postprocessing step proposed in Arora et al. (2017), which removes the top principal component of a batch of representations, is applied on produced representations from and respectively after learning with a final normalisation.
3 Experimental Design
Three unlabelled corpora from different genres are used in our experiments, including BookCorpus (Zhu et al., 2015), UMBC News (Han et al., 2013) and Amazon Book Review222Largest subset of Amazon Review.(McAuley et al., 2015); six models are trained separately on each of three corpora with each of two objectives. The summary statistics of the three corpora can be found in Table 1. Adam optimiser (Kingma & Ba, 2014)2013) are applied for stable training. Pretrained word vectors, fastText (Bojanowski et al., 2017), are used in our frameworks and fixed during learning.
|Name||# of sentences||mean # of words per sentence|
|UMBC News (2)||134.5M||25|
|Amazon Book Review (3)||150.8M||19|
All of our experiments including training and testing are done in PyTorch(Paszke et al., 2017). The modified SentEval (Conneau & Kiela, 2018) package with the step that removes the first principal component is used to evaluate our models on the downstream tasks. Hyperparameters, including negative samples in the framework with generative objective, context window in the one with discriminative objective, are tuned only on the averaged performance on STS14 of the model trained on the BookCorpus; STS14/G1 and STS14/D1 results are thus marked with a in Table 3 and Table 4 to indicate possible overfitting on that dataset/model only. Batch size and dimension in both frameworks are set to be the same for fair comparison. Hyperparameters are summarised in supplementary material.
|Arora et al. (2017);Wieting et al. (2015);Wieting & Gimpel (2018);Conneau et al. (2017);|
|Wieting & Gimpel (2018);Agirre et al. (2012);Agirre et al. (2013);Agirre et al. (2014);|
|Agirre et al. (2015);Agirre et al. (2016);Marelli et al. (2014);Mikolov et al. (2017)|
|Hill et al. (2016),Logeswaran & Lee (2018)|
3.1 Unsupervised Evaluation - Textual Similarity Tasks
Representation: For a given sentence input with words, suggested by (Pennington et al., 2014; Levy et al., 2015), the representation is calculated as , where refers to the post-processed and normalised vector, and is mentioned in Table 2.
Tasks: The unsupervised tasks include five tasks from SemEval Semantic Textual Similarity (STS) in 2012-2016 (Agirre et al., 2015, 2014, 2016, 2012, 2013) and the SemEval2014 Semantic Relatedness task (SICK-R) (Marelli et al., 2014).
Comparison: We compare our models with: Unsupervised transfer learning: We selected models with strong results from related work, including fastText, fastText+WR. Semisupervised transfer learning: The word vectors are pretrained on each task (Wieting et al., 2015) without label information, and word vectors are averaged to serve as the vector representation for a given sentence (Arora et al., 2017). Supervised transfer learning: ParaNMT (Wieting & Gimpel, 2018)
is included as a supervised transfer learning method as the data collection stage requires a neural machine translation system trained in supervised fashion. and the InferSent333The released InferSent (Conneau et al., 2017) model is evaluated with the postprocessing step. (Conneau et al., 2017) trained on SNLI (Bowman et al., 2015) and MultiNLI (Williams et al., 2017) is included as well.
All six models trained with our learning frameworks outperform other unsupervised and semi supervised transfer learning methods, and the model trained on the UMBC News Corpus with discriminative objective gives the best performance likely because the STS tasks contain multiple news- and headlines-related datasets which is well matched by the domain of the UMBC News Corpus.
3.2 Supervised Evaluation
The evaluation on these tasks involves learning a linear model on top of the learnt sentence representations produced by the model. Since a linear model is capable of selecting the most relevant dimensions in the feature vectors to make predictions, it is preferred to concatenate various types of representations to form a richer, and possibly more redundant feature vector, which allows the linear model to explore the combination of different aspects of encoder functions to provide better results.
Representation: Inspired by prior work (McCann et al., 2017; Shen et al., 2018), the representation is calculated by concatenating the outputs from the global mean-, max- and min-pooling on top of the hidden states , and the last hidden state, and is calculated with three pooling functions as well. The post-processing and the normalisation step is applied individually. These two representations are concatenated to form a final sentence representation. Table 2 presents the details.
Tasks: Semantic relatedness (SICK) (Marelli et al., 2014), paraphrase detection (MRPC) (Dolan et al., 2004), question-type classification (TREC) (Li & Roth, 2002), movie review sentiment (MR) (Pang & Lee, 2005), Stanford Sentiment Treebank (SST) (Socher et al., 2013), customer product reviews (CR) (Hu & Liu, 2004), subjectivity/objectivity classification (SUBJ) (Pang & Lee, 2004), opinion polarity (MPQA) (Wiebe et al., 2005). The results are presented in Table 5.
Comparison: Our results as well as related results of supervised task-dependent training models, supervised transfer learning models, and unsupervised transfer learning models are presented in Table 5. Note that, for fair comparison, we collect the results of the best single model (MC-QT) trained on BookCorpus in Logeswaran & Lee (2018).
Six models trained with our learning frameworks either outperform other existing methods, or achieve similar results on some tasks. The model trained on the Amazon Book Review gives the best performance on sentiment analysis tasks, since the corpus conveys strong sentiment information.
Conneau et al. (2017);Arora et al. (2017);Hill et al. (2016); Kiros et al. (2015);Ba et al. (2016);
|Gan et al. (2017);Jernite et al. (2017);Nie et al. (2017);Tai et al. (2015);Zhao et al. (2015)|
|Le & Mikolov (2014);Logeswaran & Lee (2018);Shen et al. (2018); Khodak et al. (2018)|
|Supervised task-dependent training - No transfer learning|
|Supervised training - Transfer learning|
|Unsupervised training with unordered sentences|
|GloVe+WR||-||86.0||84.6||- / -||-||-||-||-||-||82.2|
|Unsupervised training with ordered sentences|
|DiscSent ‡||8||-||-||75.0/ -||87.2||-||-||93.0||-||-|
|DisSent ‡||-||79.1||80.3||- / -||84.6||82.5||80.2||92.4||89.6||82.9|
|Unsupervised tasks||Supervised tasks|
|UMBC||Hrs||Avg of STS tasks||Avg of||Avg of Binary-CLS tasks||MRPC|
|News||(STS12-16, SICK14)||SICK-R, STS-B||(MR, CR, SUBJ, MPQA, SST)|
|Our Multi-view with Generative Objective + Invertible Constraint|
|Generative Objective without Invertible Constraint|
|9||55.7 (10.9)||79.9 (2.1)||86.0 (0.1)||73.2/81.7|
|70.1 (2.3)||82.8 (0.5)||85.0 (0.3)||74.3/82.0|
|67.8 (2.5)||82.9 (0.2)||86.4 (0.7)||74.8/83.2|
|Our Multi-view with Discriminative Objective:|
Multi-view with and :
|Multi-view with and :|
|17||49.7 (17.7)||82.2 (0.8)||86.3 (0.3)||75.9/83.0|
|57.3 (13.3)||81.9 (1.1)||87.1 (0.5)||77.2/83.7|
|2||68.5 (0.7)||80.8 (1.8)||84.2 (1.0)||72.5/82.0|
|69.1 (1.5)||77.0 (6.0)||84.5 (2.1)||73.5/82.3|
|19||67.5 (3.1)||82.3 (0.7)||86.9 (0.3)||76.6/83.8|
|Single-view with only: , Single-view with only:|
|9||57.8 (9.6)||81.6 (1.4)||85.8 (0.8)||74.8/82.3|
|1.5||68.7 (0.5)||81.1 (1.5)||83.3 (1.9)||72.9/81.0|
|10.5||68.6 (2.0)||82.3 (0.7)||86.3 (0.3)||75.4/82.5|
4.1 Ensemble in Multi-view Frameworks
In both frameworks, RNN encoder and linear encoder perform well on all tasks, and generative objective and discriminative objective give similar performance.
In general, an ensemble of the representations generated from two distinct encoding functions performs even better. The two encoding functions, and , have naturally different behaviour. With distributional similarity (Firth, 1957), our multi-view frameworks help the two encoding functions to learn more generalised representations. Therefore, and encode the input sentence with emphasis on different aspects, and the subsequently trained linear model for each of the supervised downstream tasks benefits from this diversity leading to better predictions.
4.2 Generative Objective: Regularisation on Invertibility
The orthonormal regularisation applied on the linear decoder to enforce invertibility in our multi-view framework encourages the vector representations produced by and those by , which is in testing, to align with each other. A direct comparison is to train our multi-view framework without the invertible constraint, and still directly use as an additional encoder in testing. The results of our framework with and without the invertible constraint are presented in Table 6.
The ensemble method of two views, and , on unsupervised evaluation tasks (STS12-16 and SICK14) is averaging, which benefits from aligning representations from and by applying invertible constraint, and the RNN encoder gets improved on unsupervised tasks by learning to align with . On supervised evaluation tasks, as the ensemble method is concatenation and a linear model is applied on top of the concatenated representations, as long as the encoders in two views process sentences distinctively, the linear classifier is capable of picking relevant feature dimensions from both views to make good predictions, thus there is no significant difference between our multi-view framework with and without invertible constraint.
4.3 Discriminative Objective: Multi-view vs. Single-view
In order to determine if the multi-view framework with two different views/encoding functions is helping the learning, we compare our framework with discriminative objective to other reasonable variants, including the multi-view model with two functions of the same type but parametrised independently, either two -s or two -s, and the single-view model with only one or . Table 6 presents the results of the models trained on UMBC Corpus.
In our multi-view learning with and , the two encoding functions improve each other’s view. As illustrated in previous work, and specifically emphasised in Hill et al. (2016), linear/log-linear models, which include in our model, produce better representations for STS tasks than RNN-based models do. The same finding can be observed in Table 6 as well, where consistently provides better results on STS tasks than does. In addition, as we expected, in our multi-view learning with and , improves the performance of on STS tasks. With maximising the agreement between the representations generated from and , we also see in the table that improves on supervised evaluation tasks.
Compared with the ensemble of two multi-view models, each with two encoding functions of the same type, our multi-view framework with and provides slightly better results on STS tasks, and similar results on supervised evaluation tasks, while our model has much higher training efficiency. Compared with the ensemble of two single-view models, each with only one encoding function, the matching between and in our multi-view model produces better results.
We proposed multi-view sentence representation learning frameworks with generative and discriminative objectives; each framework combines an RNN-based encoder and an average-on-word-vectors linear encoder and can be efficiently trained within a few hours on a large unlabelled corpus. The experiments were conducted on three large unlabelled corpora, and meaningful comparisons were made to demonstrate the generalisation ability and the transferability of our learning frameworks, and also to consolidate our claim. The produced sentence representations outperform existing unsupervised transfer methods on unsupervised evaluation tasks, and match the performance of the best unsupervised model on supervised evaluation tasks.
As presented in our experiments, the ensemble of two views leveraged the advantages of both views, and provides rich semantic information of the input sentence, also the multi-view learning helps each view to produce better representations than single-view learning. Meanwhile, our experimental results also support the finding in Hill et al. (2016) that linear/log-linear models ( in our frameworks) tend to work better on the unsupervised tasks, while RNN-based models ( in our frameworks) generally perform better on the supervised tasks. Future work should explore the impact of having various encoding architectures and learning under the multi-view framework.
Our multi-view learning frameworks were inspired by the asymmetric information processing in the two hemispheres of the human brain, in which for most adults, the left hemisphere contributes to sequential processing, including primarily language understanding, and the right one carries out more parallel processing, including visual spatial understanding (Bryden, 2012). The experimental results raise an intriguing hypothesis about how these two types of information processing may complementarily help learning.
We appreciate the gift funding from Adobe Research. Many thanks to Sam Bowman and Andrew Y. Ying for helpful discussion, and to Mengting Wan, Wangcheng Kang, Jianmo Ni and Andrej Zukov-Gregoric for critical comments on the project.
- Agirre et al. (2012) Eneko Agirre, Daniel M. Cer, Mona T. Diab, and Aitor Gonzalez-Agirre. Semeval-2012 task 6: A pilot on semantic textual similarity. In SemEval@NAACL-HLT, 2012.
- Agirre et al. (2013) Eneko Agirre, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez-Agirre, and Weiwei Guo. *sem 2013 shared task: Semantic textual similarity. In *SEM@NAACL-HLT, 2013.
- Agirre et al. (2014) Eneko Agirre, Carmen Banea, Claire Cardie, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Rada Mihalcea, German Rigau, and Janyce Wiebe. Semeval-2014 task 10: Multilingual semantic textual similarity. In SemEval@COLING, 2014.
- Agirre et al. (2015) Eneko Agirre, Carmen Banea, Claire Cardie, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Iñigo Lopez-Gazpio, Montse Maritxalar, Rada Mihalcea, German Rigau, Larraitz Uria, and Janyce Wiebe. Semeval-2015 task 2: Semantic textual similarity, english, spanish and pilot on interpretability. In SemEval@NAACL-HLT, 2015.
- Agirre et al. (2016) Eneko Agirre, Carmen Banea, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez-Agirre, Rada Mihalcea, German Rigau, and Janyce Wiebe. Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In SemEval@NAACL-HLT, 2016.
- Arora et al. (2016) Sanjeev Arora, Yuanzhi Li, Yingyu Liang, Tengyu Ma, and Andrej Risteski. A latent variable model approach to pmi-based word embeddings. TACL, 4:385–399, 2016.
- Arora et al. (2017) Sanjeev Arora, Yingyu Liang, and Tengyu Ma. A simple but tough-to-beat baseline for sentence embeddings. In International Conference on Learning Representations, 2017.
- Ba et al. (2016) Jimmy Ba, Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016.
- Bojanowski et al. (2017) Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. TACL, 5:135–146, 2017.
- Bowman et al. (2015) Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In EMNLP, 2015.
- Bryden (2012) MP Bryden. Laterality functional asymmetry in the intact brain. Elsevier, 2012.
- Chung et al. (2014) Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.
- Cissé et al. (2017) Moustapha Cissé, Piotr Bojanowski, Edouard Grave, Yann Dauphin, and Nicolas Usunier. Parseval networks: Improving robustness to adversarial examples. In ICML, 2017.
- Conneau & Kiela (2018) Alexis Conneau and Douwe Kiela. Senteval: An evaluation toolkit for universal sentence representations. In LREC, 2018.
- Conneau et al. (2017) Alexis Conneau, Douwe Kiela, Holger Schwenk, Loïc Barrault, and Antoine Bordes. Supervised learning of universal sentence representations from natural language inference data. In EMNLP, 2017.
- Coulson & van Petten (2007) Seana Coulson and Cyma van Petten. A special role for the right hemisphere in metaphor comprehension? erp evidence from hemifield presentation. Brain research, 1146:128–45, 2007.
- Coulson et al. (2005) Seana Coulson, Kara D. Federmeier, Cyma van Petten, and Marta Kutas. Right hemisphere sensitivity to word- and sentence-level context: evidence from event-related brain potentials. Journal of experimental psychology. Learning, memory, and cognition, 31 1:129–47, 2005.
- de Sa (1993) Virginia R. de Sa. Learning classification with unlabeled data. In NIPS, pp. 112–119, 1993.
- Dhillon et al. (2011) Paramveer S. Dhillon, Dean P. Foster, and Lyle H. Ungar. Multi-view learning of word embeddings via cca. In NIPS, 2011.
- Dolan et al. (2004) William B. Dolan, Chris Quirk, and Chris Brockett. Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources. In COLING, 2004.
- Firth (1957) J. R. Firth. A synopsis of linguistic theory. 1957.
Gan et al. (2017)
Zhe Gan, Yunchen Pu, Ricardo Henao, Chunyuan Li, Xiaodong He, and Lawrence
Learning generic sentence representations using convolutional neural networks.In EMNLP, 2017.
- Han et al. (2013) Lushan Han, Abhay L Kashyap, Tim Finin, James Mayfield, and Jonathan Weese. Umbc_ebiquity-core: semantic textual similarity systems. In Second Joint Conference on Lexical and Computational Semantics (* SEM), Volume 1: Proceedings of the Main Conference and the Shared Task: Semantic Textual Similarity, volume 1, pp. 44–52, 2013.
- Harris (1954) Zellig S Harris. Distributional structure. Word, 10(2-3):146–162, 1954.
He et al. (2015)
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Delving deep into rectifiers: Surpassing human-level performance on imagenet classification.
2015 IEEE International Conference on Computer Vision (ICCV), pp. 1026–1034, 2015.
Hill et al. (2016)
Felix Hill, Kyunghyun Cho, and Anna Korhonen.
Learning distributed representations of sentences from unlabelled data.In HLT-NAACL, 2016.
- Hu & Liu (2004) Minqing Hu and Bing Liu. Mining and summarizing customer reviews. In KDD, 2004.
- Jernite et al. (2017) Yacine Jernite, Samuel R. Bowman, and David Sontag. Discourse-based objectives for fast unsupervised sentence representation learning. CoRR, abs/1705.00557, 2017.
- Khodak et al. (2018) Mikhail Khodak, Nikunj Saunshi, Yingyu Liang, Tengyu Ma, Brandon Stewart, and Sanjeev Arora. A la carte embedding: Cheap but effective induction of semantic feature vectors. In ACL, 2018.
- Kingma & Ba (2014) Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
- Kiros et al. (2015) Jamie Ryan Kiros, Yukun Zhu, Ruslan Salakhutdinov, Richard S. Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Skip-thought vectors. In NIPS, 2015.
- Le & Mikolov (2014) Quoc V. Le and Tomas Mikolov. Distributed representations of sentences and documents. In ICML, 2014.
- Levy et al. (2015) Omer Levy, Yoav Goldberg, and Ido Dagan. Improving distributional similarity with lessons learned from word embeddings. TACL, 3:211–225, 2015.
- Li & Hovy (2014) Jiwei Li and Eduard H. Hovy. A model of coherence based on distributed sentence representation. In EMNLP, 2014.
- Li & Roth (2002) Xin Li and Dan Roth. Learning question classifiers. In COLING, 2002.
- Logeswaran & Lee (2018) Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. In ICLR, 2018.
- Marelli et al. (2014) Marco Marelli, Stefano Menini, Marco Baroni, Luisa Bentivogli, Raffaella Bernardi, and Roberto Zamparelli. A sick cure for the evaluation of compositional distributional semantic models. In LREC, 2014.
- McAuley et al. (2015) Julian J. McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel. Image-based recommendations on styles and substitutes. In SIGIR, 2015.
- McCann et al. (2017) Bryan McCann, James Bradbury, Caiming Xiong, and Richard Socher. Learned in translation: Contextualized word vectors. In NIPS, 2017.
- Mikolov et al. (2013) Tomas Mikolov, Ilya Sutskever, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. Distributed representations of words and phrases and their compositionality. In NIPS, 2013.
- Mikolov et al. (2017) Tomas Mikolov, Edouard Grave, Piotr Bojanowski, Christian Puhrsch, and Armand Joulin. Advances in pre-training distributed word representations. CoRR, abs/1712.09405, 2017.
- Mises & Pollaczek-Geiringer (1929) RV Mises and Hilda Pollaczek-Geiringer. Praktische verfahren der gleichungsauflösung. ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift für Angewandte Mathematik und Mechanik, 9(1):58–77, 1929.
- Nie et al. (2017) Allen Nie, Erin D. Bennett, and Noah D. Goodman. Dissent: Sentence representation learning from explicit discourse relations. CoRR, abs/1710.04334, 2017.
- Pang & Lee (2004) Bo Pang and Lillian Lee. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In ACL, 2004.
- Pang & Lee (2005) Bo Pang and Lillian Lee. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In ACL, 2005.
- Pascanu et al. (2013) Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In ICML, 2013.
- Paszke et al. (2017) Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017.
- Pennington et al. (2014) Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In EMNLP, 2014.
- Shen et al. (2018) Dinghan Shen, Guoyin Wang, Wenlin Wang, Martin Renqiang Min, Qinliang Su, Yizhe Zhang, Chunyuan Li, Ricardo Henao, and Lawrence Carin. Baseline needs more love: On simple word-embedding-based models and associated pooling mechanisms. In ACL, 2018.
- Socher et al. (2013) Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In EMNLP, 2013.
Tai et al. (2015)
Kai Sheng Tai, Richard Socher, and Christopher D. Manning.
Improved semantic representations from tree-structured long short-term memory networks.In ACL, 2015.
- Tang et al. (2018) Shuai Tang, Hailin Jin, Chen Fang, Zhaowen Wang, and Virginia R. de Sa. Speeding up context-based sentence representation learning with non-autoregressive convolutional decoding. In Rep4NLP@ACL, 2018.
- Turney & Pantel (2010) Peter D. Turney and Patrick Pantel. From frequency to meaning: Vector space models of semantics. J. Artif. Intell. Res., 37:141–188, 2010.
- Wiebe et al. (2005) Janyce Wiebe, Theresa Wilson, and Claire Cardie. Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 39:165–210, 2005.
- Wieting & Gimpel (2018) John Wieting and Kevin Gimpel. Paranmt-50m: Pushing the limits of paraphrastic sentence embeddings with millions of machine translations. In ACL, 2018.
- Wieting et al. (2015) John Wieting, Mohit Bansal, Kevin Gimpel, and Karen Livescu. From paraphrase database to compositional paraphrase model and back. TACL, 3:345–358, 2015.
- Williams et al. (2017) Adina Williams, Nikita Nangia, and Samuel R. Bowman. A broad-coverage challenge corpus for sentence understanding through inference. CoRR, abs/1704.05426, 2017.
- Yang et al. (2016) Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alexander J. Smola, and Eduard H. Hovy. Hierarchical attention networks for document classification. In HLT-NAACL, 2016.
- Zhao et al. (2015) Han Zhao, Zhengdong Lu, and Pascal Poupart. Self-adaptive hierarchical sentence model. In IJCAI, 2015.
- Zhu et al. (2015) Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. ICCV, pp. 19–27, 2015.
1 Evaluation tasks
The details, including size of each dataset and number of classes, about the evaluation tasks are presented below111Provided by https://github.com/facebookresearch/SentEval.
|SICK-R||4.5k||4.9k||semantic textual similarity||6|
|STS-B||5.7k||1.4k||semantic textual similarity||6|
2 Power Iteration
The Power Iteration was proposed in Mises & Pollaczek-Geiringer (1929)
, and it is an efficient algorithm for estimating the top eigenvector of a given covariance matrix. Here, it is used to estimate the top principal component from the representations produced fromand separately. We omit the superscription here, since the same step is applied to both and .
Suppose there is a batch of representations from either or , the Power Iteration method is applied here to estimate the top eigenvector of the covariance matrix222In practice, often is less than , thus we estimate the top eigenvector of .: , and it is described in Algorithm 1:
In our experiments, is set to be .
3 Training & Model Details
The hyperparameters we need to tune include the batch size , the dimension of the GRU encoder , and the context window , and the number of negative samples . The results we presented in this paper is based on the model trained with , . Specifically, in discriminative objective, the context window is set , and in generative objective, the number of negative samples is set . It takes up to 8GB on a GTX 1080Ti GPU.
The initial learning rate is , and we didn’t anneal the learning rate through the training. All weights in the model are initialised using the method proposed in He et al. (2015), and all gates in the bi-GRU are initialised to 1, and all biases in the single-layer neural network are zeroed before training. The word vectors are fixed to be those in the FastText (Bojanowski et al., 2017), and we don’t finetune them. Words that are not in the FastText’s vocabulary are fixed to vectors through training. The temperature term is initialised as , and is tuned by the gradient descent during training.
The temperature term is used to convert the agreementin Eq. 1 in the main paper. In our experiments, is a trainable parameter initialised to that decreased consistently through training. Another model trained with fixed set to the final value performed similarly.