A Multi-task Learning Approach for Improving Product Title Compression with User Search Log Data

01/05/2018 ∙ by Jingang Wang, et al. ∙ Beijing Yiyou Network Technology Co., Ltd. East China Normal University 0

It is a challenging and practical research problem to obtain effective compression of lengthy product titles for E-commerce. This is particularly important as more and more users browse mobile E-commerce apps and more merchants make the original product titles redundant and lengthy for Search Engine Optimization. Traditional text summarization approaches often require a large amount of preprocessing costs and do not capture the important issue of conversion rate in E-commerce. This paper proposes a novel multi-task learning approach for improving product title compression with user search log data. In particular, a pointer network-based sequence-to-sequence approach is utilized for title compression with an attentive mechanism as an extractive method and an attentive encoder-decoder approach is utilized for generating user search queries. The encoding parameters (i.e., semantic embedding of original titles) are shared among the two tasks and the attention distributions are jointly optimized. An extensive set of experiments with both human annotated data and online deployment demonstrate the advantage of the proposed research for both compression qualities and online business values.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Mobile Internet are changing our lives profoundly. As depicted in the Statistical Report on China’s Internet development, daily active users of mobile phones in China have overpassed 0.72 billion by June, 2017 111http://www.cnnic.cn/hlwfzyj/hlwxzbg/hlwtjbg/201708/P020170807351923262153.pdf. More online transactions are made on mobile phones instead of on PCs, with the gap still widening. This trend demands mobile E-Commerce platforms to improve user experience on their Apps. The most prominent distinction between smart phones and PCs lies in their screen sizes as a typical screen size of smart phones varies from 4.5 to 5.5 inches only.

It is an important and practical research problem for producing succinct product titles in mobile E-commerce applications. On E-commerce platforms, especially customer to customer (C2C) websites, product titles are often written by the merchants. For the sake of SEO, most of these titles are rather redundant and lengthy. As shown in Figure 0(b), the product title consists of more than 30 Chinese words, seriously hurting users’ browsing experience. Usually, when a customer browses a product on Apps, less than 10 Chinese words could be displayed due to the screen size limit, as shown in Figure 0(a). Amazon’s research also reveals that product titles with less than 80 characters improve the shopping experience by making it easier for customers to find, review, and buy 222https://sellercentral.amazon.com/forums/message.jspa?messageID=2921001.

(a) Search Result Page
(b) Product Detail Page
Figure 1: When a user issues a query “floral-dress long-sleeve women”, the complete title cannot be displayed in the Search Result Page, unless the user proceeds to the detail page further.

In comparison to conventional sentence summarization, which only requires the generated summary preserves important information grammatically, product title compression often has more constraints. The words in long titles are generally elaborated by experienced sellers and proved helpful to transaction in PC era. Many sellers do not want to include words not in their original titles. This is even more serious if the new external words make the purchase conversion rate lower. For this practical reason, we address product title compression as an extractive summarization task, i.e., all the words in the compressed short title are selected from the original product title.

Traditional methods for extractive title compression usually include two steps to perform: (1) word segmentation, fine-grained named entity recognition (NER), and term weighting preprocessing steps, and (2) constrained optimization (e.g., knapsack or integer linear programming (ILP) with pre-defined business rules and target length limit). The preprocessing steps are usually challenging and labor-intensive in E-commerce, because there exists hundreds of product categories (e.g., Clothes, Food and Electronics) on a typical E-commerce website and a single model usually performs sub-optimally across quite a lot of them. Recently emerging seq2seq models can generate sequential texts in a data-driven manner, getting rid of trivial labeling and feature engineering, but they require a large amount of human annotated training data. Furthermore, traditional summarization methods do not consider conversion rate, which is an important issue in E-commerce.

To address the problems of traditional summarization methods, this paper proposes a multi-task learning approach for improving product title compression with user search log data. For many E-commerce websites, both manually compressed titles and a large mount of user search log data exist. In particular, a multi-task framework is designed to include two attention-based neural networks. One network is used to model manually edited short titles and the original product tiles, and the other network models user search queries (with purchase transaction) with the original product titles. The two networks not only share the same encoder embedding (i.e., semantic embedding of original titles) but also are optimized simultaneously to reach an agreement on attention distributions over the original product title words. These properties enable the multi-task framework to generate more effective embedding and attention by making full use of manually edited titles, user queries and the transaction data. An extensive set of experimental results demonstrate that the multi-task framework not only achieves extractive title compression with a higher quality than alternatives but also improves conversion rates for more business values.

Although there are numerous sentence summarization research work, to the best of our knowledge, this is the first work focusing on product title compression in E-commerce. Our main contributions are as follows.

  • We propose a multi-task learning framework for extractive product title compression in E-commerce, outperforming traditional approaches like ILP.

  • Our data-driven approach can save the labeling and feature engineering cost in conventional title compression approaches.

  • The agreement-based setting on attention distributions is helpful to more accurately recognize important words in original titles, which is of great value in various E-commerce scenarios.

2 Related Work

Our work touches on several strands of research within text summarization, sentence compression and neural sequence modeling. Existing text summarization methods can be categorized into extractive and abstractive methods. Extractive methods produce a summary by dropping words from the original sentence, as opposed to abstractive compression which allows more flexible transformation. Traditional extractive methods can be broadly classified into greedy approaches

[Carbonell and Goldstein1998, Knight and Marcu2000], graph-based approaches [Erkan and Radev2004], and constraint optimization-based approaches [McDonald2007, Yao and Wan2017]. Recently neural network based approaches have become popular for sentence summarization, especially abstractive summarization [Chopra et al.2016, Rush, Chopra, and Weston2015]

. In terms of extractive methods, cao2016attsum cao2016attsum propose AttSum to tackle extractive query-focused document summarization. The model learns query relevance and sentence saliency ranking jointly, and an attention mechanism is applied to select sentences when a query is given. nallapati2017summarunner nallapati2017summarunner present a recurrent neural network based sequential model for extractive document summarization, where each sentence is visited sequentially in the original order and a binary decision is made in terms of whether to preserve it in the summary. cheng-lapata:2016:P16-1 cheng-lapata:2016:P16-1 develop a framework composed of a hierarchical document encoder and an attentive extractor for document/sentence extraction. see2017get see2017get propose Pointer-Generator network for summarization, which can copy words from the source text via pointing, achieving a balance between extractive and abstractive summarization.

More related to this work is sentence compression, in particular, compression by extraction. mcdonald2006discriminative mcdonald2006discriminative employs linguistic features including deep syntactic ones to score the decision to extract each single word within its context, and decodes by dynamic programming to achieve an optimal extraction. Similarly, thadani2013sentence thadani2013sentence propose a constrained integer linear programming (ILP) framework to jointly consider n-grams and tree structures as extraction candidates. More recently, filippova2015sentence filippova2015sentence make word extraction decision based on Long Shot Term Memories (LSTMs). miao2016language miao2016language adapt a variational auto-encoder (VAE) with a latent language model, which strives to keep the compression samples to closely resemble natural sentences.

Attention mechanism, which has gained popularity recently in multiple areas, allows the model to learn alignments between modularity [Luong, Pham, and Manning2015, Bahdanau, Cho, and Bengio2015, Ba, Mnih, and Kavukcuoglu2015, Xu et al.2015]

. Cheng2016AgreementBasedJT Cheng2016AgreementBasedJT present agreement-based joint learning for bi-directional attentive neural machine translation, in which encoder-decoder components are trained with identical dataset in reverse directions. In our work, two encoder-decoder components are trained with different data and agreement-based information is integrated into the loss objective. luong2015multi luong2015multi propose multi-task sequence-to-sequence learning (MTL) framework with different settings, including (1) the one-to-many setting sharing the encoder, (2) the many-to-one setting sharing the decoder, and (3) the many-to-many setting sharing multiple encoders and decoders. Our model follows the one-to-many setting, where the encoder parameters are shared between short title generation and query generation. In comparison, klerke2016improving klerke2016improving, dealing with a multi-task but not sequence-to-squence problem, optimize a shared Bi-LSTM network for extraction operation as well as gaze prediction.

3 Data set

Our proposed approach requires two parts of data, including (1) original product titles and their compressed versions, and (2) user search query and purchase log data for these products. Since there doesn’t exist an off-the-shelf benchmark data set for our task, we construct our data set from scratch. We take advantage of realistic data from a well-known C2C website in China as our experimental data set.

In terms of the original title and compressed title pairs, we leverage the human-generated short titles from a product-recommendation channel of the website. Display short titles in this channel are rewritten by professional business analysts in an extractive manner. We crawled all products belonging to the Women’s Clothes category as our experimental products. We exclude the products whose original titles are shorter than 10 Chinese characters because product titles that are shorter than 10 characters can be displayed completely in most scenarios.

In terms of user search queries and purchase log data, we crawled user search queries leading to more than 10 transactions in one month given a product.

The two parts of data are merged and organized as triplets, i.e., (product title, manually generated compressed title, user search query). Therefore, the data set can be represented as , where means products’ original titles, means the handcrafted short titles, and represents the successful transaction-leading search queries. The details of our data set are shown in Table 1, and a triplet example is presented in Figure 2.

Data set (triplets) size 185, 386
Avg. length of original titles 25.1
Avg. length of handcrafted short titles 7.5
Avg. length of search queries 8.3
  • Note: all the lengths are counted by Chinese characters, and each English word is counted as one Chinese character.

Table 1: The statistics of the triplet data set
Figure 2: A triplet example. Both the compressed title and the query can help recognize important information from the original title.

4 Multi-task Learning for Product Title Compression

Figure 3: Multi-task Learning Framework, including two seq2seq components sharing the identical encoder. The main task is a Pointer Network to automatically point (select) the most informative words as compressed title. The auxiliary task is a standard seq2seq model to generate user search query. We utilize the attention distribution generated from user query to encourage the main task to agree on identity words.

Given an input sequence, our goal is to produce a summary, where all the words are extracted from the input sequence. Suppose the input sequence contains words coming from a fixed vocabulary of size . A sequence to sequence (seq2seq) model takes as input and outputs a compressed sequence contains N words () coming from as well. Assume the set as all possible sentences of length , our objective is to find an optimal sequence from this set . Please note that all the words in summary are transferred from the input sequence in this paper. Therefore, we have the objective function as

(1)

where

is a scoring function. We represent it as the conditional log-probability of a summary given the input sequence,

, which can be rewritten as

(2)

where is a set of model parameters and is a partial summary. The parameters of the model are learned by maximizing the Equation 2 for the training set.

A basic seq2seq model consists of two recurrent neural networks (RNNs), named as encoder and decoder respectively [Kalchbrenner and Blunsom2013, Cho et al.2014, Bahdanau, Cho, and Bengio2015]. The encoder processes the input sequence into a sequence of hidden states ,

(3)

where is the hidden state of the -th input word and is a non-linear function. The final hidden state is the new initial state of the decoder. The decoder produces the hidden state by receiving word embeddings of the previous words, and predicts next word according to the current hidden state.

In the basic encoder-decoder framework, the decoder is supposed to generate the output sequence solely based on the last hidden state

from the encoder. It seems unreasonable to assume that the encoder can encode everything into a single state (a vector actually). RNNs are known to have problems dealing with such long-range dependencies. There are some tricks to tackle with this problem, such as replacing vanilla RNNs with Long Short Term Memories (LSTMs) or feeding an input sequence twice. However, the long-range dependencies are still problematic. Attention mechanism is introduced to address the problem

[Bahdanau, Cho, and Bengio2015, Luong, Pham, and Manning2015]. With the attention mechanism, we no longer try to encode the full source sequence into a fixed-length vector. Rather, the decoder is allowed to “attend” to different parts of the source sequence at each step of the output generation. The conditional probability in Equation 2 can be rewritten as

(4)

where is a non-linear function, is the hidden state corresponding to the -th target word computed by

(5)

and is a context vector for generating the -th target word,

(6)

where is referred as an attention matrix (i.e., alignment matrix in machine translation). reflects the contribution of the -th input word to generating the -th word .

(7)

where measures how well and are aligned, noted as in short,

(8)

where , , and are learnable parameters of the model.

Bearing the attentive seq2seq model in mind, we proceed to our multi-task framework containing two attentive seq2seq tasks for product title compression. The structure of our multi-task framework is shown in Figure 3.

The main task is fed with product titles and generates corresponding short titles, named as title compression. Recall that the compressed titles are generated in an extractive manner. We implement the title compression task with pointer networks [Vinyals, Fortunato, and Jaitly2015], since the standard seq2seq model cannot be used for our compression problem where the output dictionary is composed of the words from the input sequence. Pointer networks do not blend the encode state to propagate extra information to the decoder, but instead, use as pointers to the input sequence directly.

The auxiliary task is fed with product titles and generates user search queries, named as query generation. We adopt the attentive encoder-decoder implementation for neural machine translation (NMT) [Bahdanau, Cho, and Bengio2015]. The motivation lies that query generation can be deemed as a machine translation task seamlessly. Product titles are written by product-sellers (written in merchants’ language), while search queries are issued by users (written in customers’ language). We believe the query generation task can contribute high-quality attention matrices on the original titles.

One intuitive approach for the multi-task setting is directly combine the losses of the two separate tasks, which cannot maximize the power of multi-task learning in our opinions. Instead, we tie the two tasks in an elegant manner with agreement-based learning [Liang, Klein, and Jordan2008, Cheng et al.2016]. Since our two tasks share an identical encoder which receives the original product title, we would obtain two attention distributions over the product title sequence. We constrain these two distributions agree with each other, which means the important words recognized by two separate decoders are accordant. With respect to implementation, we introduce an attention distribution agreement-based loss, as depicted in Equation 9.

(9)

is the attention matrix of Title compression, where and are the lengths of the generated short title and the original title respectively. is the attention matrix of Query generation, where

is the length of the user search query. Due to the different sizes of these two matrices, we perform max-pooling on them to get two vectors (

) before calculating the agreement-based loss, as shown in Equation 10.

(10)

To evaluate the agreement between and , we adopt their KL-divergence as our agreement-based loss.

(11)

Finally, the loss function of our multi-task framework becomes

(12)

where represents Title compression task, and Q represents Query generation task. and are hyper-parameters to tune the impacts of the two tasks.

5 Experiments

5.1 Comparison Methods

To evaluate the summarization performance of our approach, we implement rich extractive sentence summarization methods. Among them, the first two are common baselines.

  • Truncation-based Method (Trunc.). In most E-commerce scenarios where title compression is not available, long product titles are simply truncated to adapt to the limitation. Thus Truncation-based method is a naïve baseline for product title compression. Given a product title, we keep the words in their original order until the limit is reached.

  • ILP-based Method (ILP). We also include the traditional ILP-based method [Clarke and Lapata2008]

    as another baseline, which is an unsupervised method that relies on preprocessing (i.e., word segmentation, NER and term weighting) results of the input titles. With respect to preprocessing, we adopt our internal Chinese processing toolkit. The term weighting algorithm is based on some heuristic rules. For example, the

    Product words possess higher weights than Brand words, and Brand words possess higher weights than Modifier words. Figure 4 presents an example of the preprocessing results given a product. Please note that the ILP-based approach is a rather strong baseline, which has been deployed in real scenarios of the mentioned E-commerce website.

    Figure 4: The Chinese Word segmentation, NER and Term Weighting preprocessing results for a given product title.
  • Pointer Network (Ptr-Net). An attentive seq2seq model “translating” redundant long product titles to short ones. As we introduced before, to achieve the extractive summarization, we implement the Pointer Networks [Vinyals, Fortunato, and Jaitly2015].

  • Vanilla MTL. The proposed multi-task learning method, where the final loss is the linear combination of the two separate seq2seq tasks, i.e., , the hyper-parameter is tuned with the development data set.

  • Agreement-based MTL (Agree-MTL

    ). The proposed attention distribution agreement-based multi-task learning approach. In implementation, the attention distribution agreement-based loss is interpolated into the loss function during model training, i.e.,

    . The hyper-parameters are tuned with the development data set.

5.2 Seq2seq Model Settings and Parameter Tuning

For the implementation of our seq2seq models (i.e., Ptr-Net, Vanilla-MTL and Agree-MTL), we adopt two 128-dimensional LSTMs for the bidirectional encoder and one 256-dimensional LSTMs for the decoder. To avoid the effect of Chinese word segmentation error, both our encoder and decoder inputs are Chinese characters instead of words. We initialize a 128-dimensional word embedding following normal distribution

. All the models are trained on a single Tesla M40 GPU, and optimized with Adagrad [Duchi, Hazan, and Singer2011] (learning rate=, and batch size=

). We use gradient clipping with a maximum gradient norm of

, but do not use any form of regularization. At test time, our short titles are produced with a decoder whose beam search size is 10 and maximum decoding step size is 12. We randomly select 80% of the triplet data as training data, and remain the remainder for development and test (10% for each). We first tune of Vanilla MTL using grid search on [0.1, 0.9]. The model performance cannot be improved further when . Therefore of Vanilla MTL is set as . Then we conduct the parameter tuning for Agree-MTL. We fix as and further tune . The model achieves best over the development data set when reaches .

5.3 Automatic Evaluation

Summarization systems are usually evaluated using several variants of the recall-oriented ROUGE metric [Lin2004]

. ROUGE measures the summary quality by counting the overlapping units such as n-grams between the generated summary and reference summaries. We consider ROUGE-1 (uni-grams), ROUGE-2 (bi-grams) and ROUGE-L (longest-common subsequence) as our automatic evaluation metrics. The results are shown in Table

2. As expected, the truncation-based method achieves the worst ROUGE scores because of its ignorance of semantics. The main reason is that core product words are usually appear in the latter part of the titles, and truncation-based methods tend to miss some of them in the compressed titles by taking only the lead.

Although a reversed variant of the truncation-based method may mitigate the problem, it risks missing some important brand names that normally appear early in the titles. Compared with truncation-based method, ILP-method can improve the ROUGE-1 by 18 percents and the ROUGE-2 by 10 percents. This gap may be resulted from Chinese word segmentation mistakes. All the three seq2seq variants perform better than ILP-based method obviously, revealing that seq2seq models are more capable of imitating edited short titles than unsupervised methods. The Vanilla-MTL without any constraint on two separate seq2seq tasks cannot beat the independent seq2seq task (i.e., Ptr-Net). Nevertheless, the introduction of attention distribution agreement-based constraint can enhance the performance obviously, and perform best on all variants of ROUGE metrics.

Method ROUGE-1 ROUGE-2 ROUGE-L
Trunc. 30.43 19.13 29.00
ILP 48.28 29.84 43.65
Ptr-Net 69.03 55.30 67.98
Vanilla-MTL 65.92 52.94 65.20
Agree-MTL 70.89 56.80 69.61
Table 2: ROUGE performance of various methods on the test set.

5.4 Manual Evaluation

Like related summarization work [Filippova and Altun2013, Tan, Wan, and Xiao2017], we also conduct manual evaluation on the generated short titles. We randomly sampled 300 products in the Women’s Clothes category, and asked three participants to annotate the quality of generated short titles. Three perspectives are considered during manual evaluation process: (1) Core Product Recognition. Is the core product word detected correctly in the compressed title? (2) Readability. How fluent, grammatical the short title is? (3) Informativeness. How informative the short title is?

Compared with other sentence summarization work, core product recognition is a particular requirement for product title compression. Consider a product with a Chinese title “2017 new-style Women’s Dress Slim Bowknot One-Piece Skirt”, there exists three product words in the original title, including “Women’s Dress”, “bowknot” and “one-piece dress”. Obviously, “Women’s Dress” is too general to be a good core product term, and “bowknot” is a noun-modifier of “one-piece dress”. A good compression method should reserve the core product term and drop the other redundant ones.

The core product recognition property is assessed with a binary score (i.e, for correct and for incorrect), and the other two properties are assessed with a score from (worst) to (best). The results are presented in Table 3

. To make annotation results more consistent and reliable, we exclude instances with divergent ratings (i.e., their variance is larger than a threshold). We conducted paired t-tests on the manual evaluation results. With respect to readability, our Agree-MTL is significantly (p

0.05) better than all other methods. With respect to informativeness, our Agree-MTL is significantly (p 0.05) better than all other methods except Ptr-Net (it would be significantly better than Ptr-Net when p 0.08)

The truncation-based method only achieves a accuracy of 8.3% on core product recognition. As we analyzed in last section, this is caused by the phenomenon that core product words usually appear in the latter part of product titles.

The results indicate that our MTL methods outperform the unsupervised methods (i.e., Trunc. and ILP), getting particularly good marks for both readability and informativeness. Note that the unsupervised baseline is also capable of generating readable compressions but does a much poorer job in selecting most important information. Our MTL method successfully learned to optimize both scores.

Method Avg. Accu Avg. Read Avg. Info
Trunc. 8.33 % 1.93 1.96
ILP 93.33% 4.63 3.90
Ptr-Net 98.33 % 4.66 4.13
Vanilla-MTL 96.67% 4.63 3.90
Agree-MTL 98.33 % 4.80 4.66
Table 3: Manual evaluation results, including average core product recognition accuracy (Avg. Accu), average readability score (Avg. Read) and average informativeness score (Avg. Info).

5.5 Case Studies

In this section, we present two realistic cases in Figure 5.

Figure 5: Case studies.

Generally, truncation-based method achieves the worst performance because of missing core product words in the short titles. ILP-based method can recognize core product words correctly and produce reasonable short titles. However, readability is sometimes unsatisfactory. In the right example of Figure 5, the short title generate by ILP is not fluent due to word segmentation mistakes (surrounded by square brackets).

In comparison to both baselines, three seq2seq-based methods can produce fluent and grammatic short titles. With the presence of attention distribution agreement constraint, Agree-MTL perform better than Vanilla-MTL and Ptr-Net by exposing important words that match users’ information need (i.e., top search queries) on the corresponding product. Consider the left example of Figure 5, Agree-MTL method produces the short title with the English brand word (i.e., D’ZZIT), while Vanilla-MTL uses the Chinese brand word (i.e., DiSu) instead. From the top search queries that successfully leads final transactions, we can conclude that users prefer to search this product with the English brand word, and Agree-MTL captures this information and exposes it during compression.

6 Online Deployment

Previous experimental results have proven the advantages of our proposed Agree-MTL approach, so we deploy it in a real world online environment to test its practical performance.

We perform A/B testing in the search result page scenario of the mentioned E-commerce website (over 4 million daily active users in the Women’s Clothes category). Note that ILP-based method is already deployed on the website. The A/B testing system split online users equally into two groups and direct them into two separate buckets respectively. Then for users in the A bucket (i.e., the baseline group), the short titles are generated by ILP-based method. While for users in the B bucket (i.e., the experimental group), the short titles are generated by Agree-MTL method.

The online A/B testing lasted for one week. All the conditions of the two buckets are identical except the title compression methods. Two indicative measures are adopted to test the performance, including product Click Through Rate (CTR) and Click Conversion Rate (CVR) which can be calculated as

(13)

where is the clicking times of the product, is the page views of the product, and is the number of purchases of the product.

We calculated the overall CTR and CVR for all the products in the Women’s Clothes category in the two buckets. We found that the performance in the experimental bucket (i.e., Agree-MTL) is significantly better () than that in the baseline bucket (i.e., ILP) on both measures. Agree-MTL improved the CTR by 2.58% and the CVR by 1.32% over the baseline. More specifically, the improvement of CTR implies that users are more likely to browser and click the Agree-MTL generated short titles. The improvement of CVR means that after a user’s click, Agree-MTL has higher probability to convert the click into a purchase action. This is quite impressive, if we consider the fact that product titles on a search result page occupy a relatively small space and thus only partially affect the users’ decision.

7 Conclusion

Product titles on E-commerce platforms are frequently redundant and lengthy as a result of SEO. Meanwhile, people are getting used to browsing E-commerce Apps on their phones, where long titles cannot be displayed properly due to the limited screen size. Hence, product title compression is much desired. Taking it as an extractive sentence summarization task, traditional methods require lots of preprocessing cost without taking transaction conversion rate optimization into consideration. We address this problem with a pointer network-based sequence to sequence model in a multi-task setting. Sharing an identical encoder, a parallel seq2seq model is trained to match product titles and their transaction-leading search queries. We apply the attention mechanism in this framework, where for the two seq2seq models their individual attention distributions over tokens in product titles are jointly optimized to better focus on important tokens during compression. This framework not only improves user experience by compressing redundant titles into concise ones, but also guarantees query-initiated transaction conversion rate by prioritizing query-related keywords in the resultant short titles. We perform extensive experiments with realistic data from a dominant Chinese E-commerce website. The advantages of the proposed framework in terms of its compression quality and business value are demonstrated by these experiments and online deployment.

8 Acknowledgments

The authors would like to thank Dr. Nan Li and Dr. Ding Liu with Alibaba Group for their valuable discussions and the anonymous reviewers for their helpful comments. Junfeng Tian is supported by the Science and Technology Commission of Shanghai Municipality Grant (No. 15ZR1410700) and the Open Project of Shanghai Key Laboratory of Trustworthy Computing Grant (No. 07dz22304201604).

References

  • [Ba, Mnih, and Kavukcuoglu2015] Ba, J.; Mnih, V.; and Kavukcuoglu, K. 2015. Multiple object recognition with visual attention. In Proceedings of ICLR.
  • [Bahdanau, Cho, and Bengio2015] Bahdanau, D.; Cho, K.; and Bengio, Y. 2015. Neural machine translation by jointly learning to align and translate. In Proceedings of ICLR.
  • [Cao et al.2016] Cao, Z.; Li, W.; Li, S.; Wei, F.; and Li, Y. 2016. Attsum: Joint learning of focusing and summarization with neural attention. In Proceedings of COLING, 547–556.
  • [Carbonell and Goldstein1998] Carbonell, J., and Goldstein, J. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In Proceedings of SIGIR, 335–336. ACM.
  • [Cheng and Lapata2016] Cheng, J., and Lapata, M. 2016. Neural summarization by extracting sentences and words. In Proceedings of ACL, 484–494.
  • [Cheng et al.2016] Cheng, Y.; Shen, S.; He, Z.; He, W.; Wu, H.; Sun, M.; and Liu, Y. 2016. Agreement-based joint training for bidirectional attention-based neural machine translation. In Proceedings of IJCAI.
  • [Cho et al.2014] Cho, K.; van Merrienboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; and Bengio, Y. 2014. Learning phrase representations using rnn encoder–decoder for statistical machine translation. In Proceedings of EMNLP, 1724–1734.
  • [Chopra et al.2016] Chopra, S.; Auli, M.; Rush, A. M.; and Harvard, S. 2016. Abstractive sentence summarization with attentive recurrent neural networks. In Proceedings of HLT-NAACL, 93–98.
  • [Clarke and Lapata2008] Clarke, J., and Lapata, M. 2008. Global inference for sentence compression: An integer linear programming approach.

    Journal of Artificial Intelligence Research

    31:399–429.
  • [Duchi, Hazan, and Singer2011] Duchi, J.; Hazan, E.; and Singer, Y. 2011. Adaptive subgradient methods for online learning and stochastic optimization.

    Journal of Machine Learning Research

    12(Jul):2121–2159.
  • [Erkan and Radev2004] Erkan, G., and Radev, D. R. 2004. Lexrank: Graph-based lexical centrality as salience in text summarization. Journal of Artificial Intelligence Research 22:457–479.
  • [Filippova and Altun2013] Filippova, K., and Altun, Y. 2013. Overcoming the lack of parallel data in sentence compression. In Proceedings of EMNLP, 1481–1491.
  • [Filippova et al.2015] Filippova, K.; Alfonseca, E.; Colmenares, C. A.; Kaiser, L.; and Vinyals, O. 2015. Sentence compression by deletion with lstms. In Proceedings of EMNLP, 360–368.
  • [Kalchbrenner and Blunsom2013] Kalchbrenner, N., and Blunsom, P. 2013. Recurrent continuous translation models. In Proceedings of EMNLP.
  • [Klerke, Goldberg, and Søgaard2016] Klerke, S.; Goldberg, Y.; and Søgaard, A. 2016. Improving sentence compression by learning to predict gaze. In Proceedings of NAACL-HLT, 1528–1533.
  • [Knight and Marcu2000] Knight, K., and Marcu, D. 2000. Statistics-based summarization-step one: Sentence compression. In Proceedings of AAAI, volume 2000, 703–710.
  • [Liang, Klein, and Jordan2008] Liang, P. S.; Klein, D.; and Jordan, M. I. 2008. Agreement-based learning. In NIPS, 913–920.
  • [Lin2004] Lin, C.-Y. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out: Proceedings of the ACL-04 workshop, volume 8. Barcelona, Spain.
  • [Luong et al.2016] Luong, M.-T.; Le, Q. V.; Sutskever, I.; Vinyals, O.; and Kaiser, L. 2016. Multi-task sequence to sequence learning. In Proceedings of ICLR.
  • [Luong, Pham, and Manning2015] Luong, M.-T.; Pham, H.; and Manning, C. D. 2015. Effective approaches to attention-based neural machine translation. In Proceedings of EMNLP, 1412–1421.
  • [McDonald2006] McDonald, R. T. 2006. Discriminative sentence compression with soft syntactic evidence. In Proceedings of EACL.
  • [McDonald2007] McDonald, R. 2007. A study of global inference algorithms in multi-document summarization. Advances in Information Retrieval 557–564.
  • [Miao and Blunsom2016] Miao, Y., and Blunsom, P. 2016. Language as a latent variable: Discrete generative models for sentence compression. In Proceedings of EMNLP, 319–328.
  • [Nallapati, Zhai, and Zhou2017] Nallapati, R.; Zhai, F.; and Zhou, B. 2017. Summarunner: A recurrent neural network based sequence model for extractive summarization of documents. In Proceedings of AAAI.
  • [Rush, Chopra, and Weston2015] Rush, A. M.; Chopra, S.; and Weston, J. 2015.

    A neural attention model for abstractive sentence summarization.

    In Proceedings of EMNLP, 379–389.
  • [See, Manning, and Liu2017] See, A.; Manning, C.; and Liu, P. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of ACL.
  • [Tan, Wan, and Xiao2017] Tan, J.; Wan, X.; and Xiao, J. 2017. Abstractive document summarization with a graph-based attentional neural model. In Proceedings of ACL, 1171–1181.
  • [Thadani and McKeown2013] Thadani, K., and McKeown, K. 2013. Sentence compression with joint structural inference. In Proceedings of CoNLL, 65–74.
  • [Vinyals, Fortunato, and Jaitly2015] Vinyals, O.; Fortunato, M.; and Jaitly, N. 2015. Pointer networks. In NIPS, 2692–2700.
  • [Xu et al.2015] Xu, K.; Ba, J.; Kiros, R.; Cho, K.; Courville, A.; Salakhudinov, R.; Zemel, R.; and Bengio, Y. 2015. Show, attend and tell: Neural image caption generation with visual attention. In ICML, 2048–2057.
  • [Yao and Wan2017] Yao, J., and Wan, X. 2017. Greedy flipping for constrained word deletion. In Proceedings of AAAI.