Online news commenting allows Internet users to express their opinions, show their attitudes, and communicate with each other under the news articles they read. The commentaries extend the content of the articles and improve user engagement of the news websites by encouraging users to browse the comment stream, to share new information, and to debate with one another. With the prevalence of online news articles with comments, it is of great interest to build an automatic news commenting system with data-driven approaches. Such systems can enable commenting service for a news website from cold start, enhance the reading experience for less commented news articles, and enrich skill lists of other artificial intelligence applications, such as chatbotsShum et al. (2018).
|Title: FIFA rankings: France number one, Croatia and England soar, Germany and Argentina plummet|
|Body (truncated): World Cup glory has propelled France to the top of FIFA’s latest world rankings, with the impact of Russia 2018 felt for better or worse among a number of football’s heavyweight nations.|
|These are the first set of rankings released under FIFA’s new formula that ”relies on adding/subtracting points won or lost for a game to/from the previous point totals rather than averaging game points over a given time period”.|
|FIFA world rankings: 1.France 2.Belgium 3. Brazil 4. Croatia 5. Uruguay 6. England 7. Portugal 8. Switzerland 9. Spain 10. Denmark|
|Comment A: If it’s heavily based on the 2018 WC, hence England leaping up the rankings, how are Brazil at 3?|
|Comment B: England above Spain, Portugal and Germany. Interesting.|
In this work, we study the problem of automatic news comment generation, which is a less explored task in the literature of natural language generationGatt and Krahmer (2018). Existing work does some preliminary studies, where a comment is generated either from the title of a news article only Zheng et al. (2018); Qin et al. (2018) or by feeding the entire article (title plus body) to a basic sequence-to-sequence (s2s) model with an attention mechanism Qin et al. (2018). News titles are short and succinct, and thus only using news titles may lose quite a lot of useful information in comment generation. On the other hand, a news article and a comment is not a pair of parallel text. The news article is much longer than the comment and contains much information that is irrelevant to the comment. Thus, directly applying the s2s model, which has proven effective in machine translation Bahdanau et al. (2015), to the task of news comment generation is unsuitable, and may bring a lot of noise to generation. Both approaches oversimplify the problem of news comment generation and are far from how people behave on news websites. In practice, people read a news article, draw attention to some points in the article, and then present their comments along with the points they are interested in. Table 1 illustrates news commenting with an example from Yahoo! News.111https://www.yahoo.com/news/fifa-rankings-france-number-one-112047790.html The article is about the new FIFA ranking, and we pick two comments among the many to explain how people behave in the comment section. First, both commenters have gone through the entire article, as their comments are built upon the details in the body. Second, the article gives many details about the new ranking, but both commenters only comment on a few points. Third, the two commenters pay their attention to different places of the article: the first one notices that the ranking is based on the result of the new world cup, and feels curious about the position of Brazil; while the second one just feels excited about the new position of England. The example indicates a “read-attend-comment” behavior of humans and sheds light on how to construct a model that is close to the real practice.
To mimic human behavior, we propose a reading network and a generation network that generate a comment from the entire news article. The reading network simulates how people digest a news article, and acts as an encoder of the article. The generation network then simulates how people comment the article after reading it, and acts as a decoder of the comment. Specifically, from the bottom to the top, the reading network consists of a representation layer, a fusion layer, and a prediction layer. The first layer represents the title of the news with a recurrent neural network with gated recurrent units (RNN-GRUs)Cho et al. (2014)
and represents the body of the news through self-attention which can model long-term dependency among words. The second layer forms a representation of the entire news article by fusing the information of the title into the representation of the body with an attention mechanism and a gate mechanism. The attention mechanism selects useful information in the title, and the gate mechanism further controls how much such information flows into the representation of the article. Finally, the third layer is built on top of the previous two layers and employs a multi-label classifier and a pointer networkVinyals et al. (2015) to predict a bunch of salient spans (e.g., words, phrases, and sentences, etc.) from the article. With the reading network, our model comprehends the news article and boils it down to some key points (i.e., the salient spans). The generation network is an RNN language model that generates a comment word by word through an attention mechanism Bahdanau et al. (2015) on the selected spans and the news title. In training, since salient spans are not explicitly available, we treat them as a latent variable, and jointly learn the two networks from article-comment pairs by optimizing a lower bound of the true objective through a Monte Carlo sampling method. Thus, training errors in comment prediction can be back-propagated to span selection and used to supervise news reading comprehension.
We conduct experiments on two large scale datasets. One is a Chinese dataset published recently in Qin et al. (2018), and the other one is an English dataset built by crawling news articles and comments from Yahoo! News. Evaluation results on the two datasets consistently indicate that our model can significantly outperform existing methods on both automatic metrics and human judgment. The Yahoo! News corpus and our source code are available online.222https://github.com/TobeyYang/DeepCom
Our contributions are four-folds: (1) proposal of “read-attend-comment” procedure for news comment generation with a reading network and a generation network; (2) joint optimization of the two networks with an end-to-end learning approach; (3) empirical verification of the effectiveness of the proposed model on two public datasets with both automatic metrics and human annotations; and (4) release of an English news commenting dataset to the research community.
2 Related Work
News comment generation is a sub-task of natural language generation (NLG). Among various NLG tasks, the task studied in this paper is most related to summarization Rush et al. (2015); Nallapati et al. (2016); See et al. (2017) and product review generation Tang et al. (2016); Dong et al. (2017). However, there is stark difference between news comment generation and the other two tasks: the input of our task is an unstructured document, while the input of product review generation is structured attributes of a product; and the output of our task is a comment which often extends the content of the input with additional information, while the output of summarization is a condensed version of the input that contains the main information from the original. Very recently, there emerge some studies on news comment generation. For example, zheng2018automatic propose a gated attention neural network model to generate news comments from news titles. The model is further improved by a generative adversarial net. qin2018automatic publish a dataset with results of some basic models. Different from all the existing methods, we attempt to comprehend the entire news articles before generation and perform end-to-end learning that can jointly optimize the comprehension model and the generation model.
Our model is partially inspired by the recent success of machine reading comprehension (MRC), whose prosperity can be attributed to an increase of publicly available large scale annotated datasets, such as SQuAD Rajpurkar et al. (2016, 2018) and MS Marco Nguyen et al. (2016) etc. A great number of models have been proposed to tackle the MRC challenges, including BiDAF Seo et al. (2016), r-net Wang et al. (2017), DCN Xiong et al. (2016), Document Reader Chen and Bordes (2017), QANet Yu et al. (2018), and s-net Tan et al. (2018) etc. Our work can be viewed as an application of MRC to a new NLG task. The task aims to generate a comment for a news article, which is different from existing MRC tasks whose goal is to answer a question. Our learning method is also different from those in the MRC works.
3.1 Problem Formalization
3.2 Model Overview
Figure 1 illustrates the architecture of our model. In a nutshell, the model consists of a reading network and a generation network. The reading network first represents a news title and a news body separately in a representation layer, then forms a representation of the entire article by fusing the title into the body through a fusion layer, and finally distills some salient spans from the article by a prediction layer. The salient spans and the news title are then fed to the generation network to synthesize a comment. With the two networks, we can factorize the generation probability as , where refers to a set of spans in , represents the reading network, and refers to the generation network.
3.3 Reading Network
In the representation layer, let be a news title with the -th word, and be the associated news body with the -th word, we first look up an embedding table and represent and as and respectively, where and are randomly initialized, and jointly learned with other parameters. Different from the title, the body is long and consists of multiple sentences. Hence, to emphasize positional information of words in the body, we further expand with and , where are positional embeddings with the former indexing the position of in its sentence and the latter indicating the position of the sentence in the entire body. The representation of is then given by , where ,
refers to a multi-layer perceptron with two layers, andmeans the concatenation of the three arguments.
Starting from and as initial representations of and respectively, the reading network then transforms
into a sequence of hidden vectorswith a recurrent neural network with gated recurrent units (RNN-GRUs) Cho et al. (2014). In the meanwhile, is transformed to with the -th entry defined as (two layers) and is an attention-pooling vector calculated by a scaled dot-product attention Vaswani et al. (2017) denoted as :
In Equation (1), a word is represented via all words in the body weighted by their similarity. By this means, we try to capture the dependency among words in long distance.
The fusion layer takes and as inputs, and produces as new representation of the entire news article by fusing into . Specifically, , we first let attend to and form a representation , where is parameterized as Equation (1). With this step, we aim to recognize useful information in the title. Then we combine and as with a gate which balances the impact of and and filters noise from . is defined by:
The top layer of the reading network extracts a bunch of salient spans based on . Let denote the salient spans, where and refer to the start position and the end position of the -th span respectively, we propose detecting the spans with a multi-label classifier and a pointer network. Specifically, we formalize the recognition of as a multi-label classification problem with as an input and as an output, where , means that the -th word is a start position of a span, otherwise . Here, we assume that are independent with each other, then the multi-label classifier can be defined as binary classifiers with the -th classifier given by
where refers to a two-layer MLP, and is a probability distribution with the first entry as and the second entry as . The advantage of the approach is that it allows us to efficiently and flexibly detect a variable number of spans from a variable-length news article, as there is no dependency among the classifiers, and they can be calculated in parallel.
Given , the end position is recognized via a probability distribution which is defined by a pointer network:
where is an attention-pooling vector based on parameter :
is defined in a similar way with replaced by .
Let us denote as and as , then can be formulated as
In practice, we recognize the -th word as a start position if , and determine the associated end position by . Note that we do not adopt a pointer network to detect the start positions, because in this case, either we have to set a threshold on the probability distribution, which is sensitive to the length of the news article and thus hard to tune, or we can only pick a fixed number of spans for any articles by ranking the probabilities. Neither of them is favorable.
3.4 Generation Network
With the salient spans, the representation of the news article, and the representation of the news title given by the three layers of the reading network respectively, we define a representation of as . The generation network takes and as inputs and decodes a comment word by word via attending to both and . At step , the hidden state is , . is the embedding of the word generated at step -, and are context vectors that represent attention on the title and the spans respectively. is defined as Equation (5).
With , we calculate and via and respectively and obtain a probability distribution over vocabulary by . Let be a comment where , is the index of the -th word of in vocabulary, then is defined as
where refers to the -th entry of . In decoding, we define the initial state as an attention-pooling vector over the concatenation of and given by . is a parameter learned from training data.
3.5 Learning Method
We aim to learn and from , but is not explicitly available, which is a common case in practice. To address the problem, we treat as a latent variable, and consider the following objective:
where refers to the space of sets of spans, and is a set of salient spans for . Objective is difficult to optimize, as logarithm is outside the summation. Hence, we turn to maximizing a lower bound of Objective which is defined as:
Let denote all parameters of our model and denote the gradient of on an example , then is given by
To calculate the gradient, we have to enumerate all possible s for , which is intractable. Thus, we employ a Monte Carlo sampling method to approximate . Suppose that there are samples, then the approximation of is given by
). Although the Monte Carlo estimator is unbiased, it typically suffers from high variance. To reduce variance, we subtract baselines from. Inspired by Mnih and Gregor (2014), we introduce an observation-dependent baseline to capture the systematic difference in news-comment pairs during training. Besides, we also exploit a global baseline to further control the variance of the estimator. The approximation of is then re-written as
To calculate , we first encode the word sequences of and with GRUs respectively, and then feed the last hidden states of the GRUs to a three-layer MLP. is calculated as an average of over the current mini-batch. The parameters of the GRUs and the MLP are estimated via .
The learning algorithm is summarized in Algorithm 1. To speed up convergence, we initialize our model through pre-training the reading network and the generation network. Specifically, , we construct an artificial span set , and learn the parameters of the two networks by maximizing the following objective:
is established in two steps: first, we collect all associated comments for
, extract n-grams () from the comments, and recognize an n-gram in as a salient span if it exactly matches with one of the n-grams of the comments. Second, we break as sentences and calculate a matching score for a sentence and an associated comment. Each sentence corresponds to a group of matching scores, and if any one of them exceeds , we recognize the sentence as a salient span. The matching model is pre-trained with with as a positive example and a randomly sampled comment from other news as a negative example. In the model, and are first processed by GRUs separately, and then the last hidden states of the GRUs are fed to a three-layer MLP to calculate a score.
We test our model on two large-scale news commenting datasets.
4.1 Experimental Setup
The first dataset is a Chinese dataset built from Tencent News (news.qq.com) and published recently in Qin et al. (2018). Each data point contains a news article which is made up of a title and a body, a group of comments, and some side information including upvotes and categories. Each test comment is labeled by two annotators according to a 5-scale labeling criteria presented in Table 3. All text in the data is tokenized by a Chinese word segmenter Jieba (https://github.com/fxsjy/jieba). The average lengths of news titles, news bodies, and comments are words, words and words respectively. In addition to the Chinese data, we also build another dataset by crawling news articles and the associated comments from Yahoo! News. Besides upvotes and categories, side information in Yahoo data also includes paragraph marks, WIKI-entities, downvotes, abusevotes, and sentiment tagged by Yahoo!. Text in the data is tokenized by Stanford CoreNLP pipline Manning et al. (2014). As pre-processing, we filter out news articles shorter than words in the body and comments shorter than words or longer than words. Then, we remove news articles with less than comments. If the number of comments of an article exceeds , we only keep top comments with the most upvotes. On average, news titles, news bodies, and comments contain words, words and words respectively. More information about Yahoo data can be found in Appendix A. After the pre-processing, we randomly sample a training set, a validation set, and a test set from the remaining data, and make sure that there is no overlap among the three sets. Table 2 summarizes the statistics of the two datasets. Note that we only utilize news titles, news bodies and comments to learn a generation model in this work, but both datasets allow modeling news comment generation with side information, which could be our future work.
|Avg. # Cmts per News||27||27||27|
|Avg. # Cmts per News||20.6||20.5||20.5|
Following Qin et al. (2018), we evaluate the performance of different models with both automatic metrics and human judgment. In terms of automatic evaluation, we employ BLEU Papineni et al. (2002), METEOR Banerjee and Lavie (2005), ROUGE Lin (2004), and CIDEr Vedantam et al. (2015) as metrics on both data. Besides these metrics, qin2018automatic propose human score weighted metrics including W-BLEU, W-METEOR, W-ROUGE and W-CIDEr. These metrics, however, requires human judgment on each comment in the test set. Thus, we only involve results w.r.t. these metrics in Tencent data. As qin2018automatic do not publish their code for metric calculation, we employ a popular NLG evaluation project available at https://github.com/Maluuba/nlg-eval, and modify the scripts with the scores provided in the data according to the formulas in Qin et al. (2018) to calculate all the metrics. In human evaluation, for each dataset, we randomly sample articles from the test data and recruit three native speakers to judge the quality of the comments given by different models. For every article, comments from all models are pooled, randomly shuffled, and presented to the annotators. Each comment is judged by the three annotators under the criteria in Table 3.
|5||Rich in content; attractive; deep insights; new yet relevant viewpoints|
|4||Highly relevant with meaningful ideas|
|3||Less relevant; applied to other articles|
|1||Hard to read; Broken language; Only emoji|
The following models are selected as baselines:
Basic models: the retrieval models and the generation models used in Qin et al. (2018)
including (1) IR-T and IR-TC: both models retrieve a set of candidate articles with associated comments by cosine of TF-IDF vectors. Then the comments are ranked by a convolutional neural network (CNN) and the top position is returned. The difference is that IR-T only utilizes titles, while IR-TC leverages both titles and news bodies; (2) Seq2seq: the basic sequence-to-sequence modelSutskever et al. (2014) that generates a comment from a title; and (3) Att and Att-TC: sequence-to-sequence with attention Bahdanau et al. (2015) in which the input is either a title (Att) or a concatenation of a title and a body (Att-TC). In Seq2seq, Att, and Att-TC, top 1 comment from beam search (beam size=5) is returned.
GANN: the gated attention neural network proposed in Zheng et al. (2018). The model is further improved by a generative adversarial net.
We denote our model as “DeepCom” standing for “deep commenter”, as it is featured by a deep reading-commenting architecture. All baselines are implemented according to the details in the related papers and tuned on the validation sets.
4.3 Implementation Details
For each dataset, we form a vocabulary with the top
k frequent words in the entire data. We pad or truncate news titles, news bodies, and comments to make them in lengths of, , and respectively. The dimension of word embedding and the size of hidden states of GRU in all models are set as . In our model, we set as and (i.e., dimension of the position embedding in the reading network) as . The size of hidden layers in all MLPs is . The number of samples in Monte Carlo sampling is
. In pre-training, we initialize our model with a Gaussian distributionand optimize Objective (12) using AdaGrad Duchi et al. (2011) with an initial learning rate and an initial accumulator value . Then, we optimize
using stochastic gradient descent with a learning rate. In decoding, top comment from beam search with a size of is selected for evaluation. In IR-T and IR-TC, we use three types of filters with window sizes , , and in the CNN based matching model. The number of each type of filters is .
4.4 Evaluation Results
|Dataset||Metrics||No Reading||No Prediction||No Sampling||Full Model|
reports evaluation results in terms of both automatic metrics and human annotations. On most automatic metrics, DeepCom outperforms baseline methods, and the improvement is statistically significant (t-test with-value ). The improvement on BLEU-1 and W-BLEU-1 is much bigger than that on other metrics. This is because BLEU-1 only measures the proportion of matched unigrams out of the total number of unigrams in the generated comments. In human evaluation, although the absolute numbers are different from those reported in Qin et al. (2018) due to the difference between human judgements, the overall trend is consistent. In human evaluation, the values of Fleiss’ kappa over all models are more than , indicating substantial agreement among the annotators. Although built in a complicated structure, GANN does not bring much improvement over other baseline methods, which demonstrates that only using news titles is not enough in comment generation. IR-TC and Att-TC represent the best retrieval model and the best generation model among the baselines on both datasets, implying that news bodies, even used in a simple way, can provide useful information to comment generation.
We compare the full model of DeepCom with the following variants: (1) No Reading: the entire reading network is replaced by a TF-IDF based keyword extractor, and top keywords (tuned on validation sets) are fed to the generation network; (2) No Prediction: the prediction layer of the reading network is removed, and thus the entire is used in the generation network; and (3) No Sampling: we directly use the model pre-trained by maximizing Objective (12). Table 5
reports the results on automatic metrics. We can see that all variants suffer from performance drop and No Reading is the worst among the three variants. Thus, we can conclude that (1) span prediction cannot be simply replaced by TF-IDF based keyword extraction, as the former is based on a deep comprehension of news articles and calibrated in the end-to-end learning process; (2) even with sophisticated representations, one cannot directly feed the entire article to the generation network, as comment generation is vulnerable to the noise in the article; and (3) pre-training is useful, but optimizing the lower bound of the true objective is still beneficial.
To further understand why DeepCom is superior to its variants, we calculate BLEU-1 (denoted as BLEU) with the predicted spans and the ground truth comments in the test sets of the two data, and compare it with a baseline BLEU-1 (denoted as BLEU) which is calculated with the entire news articles and the ground truth comments. On Tencent data, BLEU and BLEU are and respectively, and the two numbers on Yahoo data are and respectively. This means that by extracting salient spans from news articles, we can filter out redundant information while keeping the points that people like to comment on, which explains why DeepCom is better than No Prediction. When comparing DeepCome with No Sampling, we find that spans in DeepCom are longer than those in No Sampling. In the test set of Tencent data, the average lengths of salient spans with and without sampling are and respectively, and the two numbers in Yahoo data are and respectively. Thus, DeepCom can leverage discourse-level information rather than a few single words or bi-grams in comment generation, which demonstrates the advantage of our learning method.
Analysis of human annotations:
We check the distributions of human labels for DeepCom, Att-TC, and IR-TC to get insights on the problems these models suffer from. Table 6 shows the results. Most of the bad comments from IR-TC are labeled as “2”, meaning that although IR-TC can give attractive comments with rich and deep content, its comments are easy to diverge from the news articles and thus be judged as “irrelevant”. In terms of Att-TC, there are much more comments judged as “1” than the other two models, indicating that Att-TC often generates ill-formed sentences. This is because a news article and a comment are highly asymmetric in terms of both syntax and semantics, and thus the generation process cannot be simply modeled with an encoder-decoder structure. Bad cases from DeepCom concentrate on “3”, reminding us that we need to further enrich the content of comments and improve their relevance in the future.
Finally, to further understand our model, we visualize the predicted salient spans and the generated comment using a test example from Tencent dataset in Table 7. Due to space limitation, we truncate the body and only show three of the selected spans in the truncated body. The full article with the full set of spans and another test example from Yahoo! News dataset are shown in Appendix B. In spite of this, we can see that the model finds some interesting points after “reading” the article and synthesizes a comment along one of the spans (i.e. “Chinese Paladin 3”). More interestingly, the model extends the content of the article in the comment with “Luo Jin”, who is Tiffany Tang’s partner but is not talked about in the article. On the other hand, comments given by the baseline methods are either too general (Att-TC, the best generation baseline), or totally irrelevant with the article (IR-TC, the best retrieval baseline). The example demonstrates that our model can generate relevant and informative comments through analyzing and understanding news articles.
|Title: 唐嫣为什么不演清宫剧？ (Why Tiffany Tang never plays a role in a drama set in Qing dynasty?)|
|Body (truncated): 众多的戏剧影视作品让唐嫣获得不少人气，众多活动的多变造型树立其在大众心中百变小天后的形象。… 如果说唐嫣最美的造型是《仙剑奇侠传三》中的紫萱，|
|(… The numerous television series and movies Tiffany Tang acted in have made her very popular, and her varied modelling in many activities has set her an image of ”the queen of reinvention” in the hearts of the public. … If her most beautiful role is Zi Xuan in “Chinese Paladin 3”, then the most ugly one should be this! … )|
|DeepCom: 唐嫣罗晋的演技真的很好，特别喜欢她演[blue]《仙剑奇侠传》。我觉得这部剧挺好看。(Tiffany Tang and Luo Jin are really good actors. I especially like her role in [blue]“Chinese Paladin 3”. I think the TV drama is worth watching.)|
|Att-TC: 我也是醉了 (I have nothing to comment.)|
|IR-TC: 星爷和谁开撕过嘛, 都是别人去撕星爷！(Stephen Chow never fights with others. It is others that fight with Stephen Chow!)|
We propose automatic news comment generation with a reading network and a generation network. The reading network comprehends a news article and boils it down to some salient points. The generation network then leverages the salient points and the news title to synthesize a comment. Experimental results on two public data sets indicate that the proposed model can significantly outperform baseline methods in terms of both automatic evaluation and human evaluation.
We appreciate the valuable comments provided by the anonymous reviewers. This work is supported in part by the National Natural Science Foundation of China (Grand Nos. U1636211, 61672081, 61370126), and the National Key R&D Program of China (No. 2016QY04W0802).
- Bahdanau et al. (2015) Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations.
- Banerjee and Lavie (2005) Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65–72. Association for Computational Linguistics.
- Chen and Bordes (2017) Danqi Chen and Antoine Bordes. 2017. Reading wikipedia to answer open-domain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870–1879.
- Cho et al. (2014) Kyunghyun Cho, Bart van Merriënboer, Dzmitry Bahdanau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder–decoder approaches. Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 103–111.
- Dong et al. (2017) Li Dong, Shaohan Huang, Furu Wei, Mirella Lapata, Ming Zhou, and Ke Xu. 2017. Learning to generate product reviews from attributes. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 623–632.
- Duchi et al. (2011) John Duchi, Elad Hazan, and Yoram Singer. 2011. Adaptive subgradient methods for online learning and stochastic optimization. J. Mach. Learn. Res., 12:2121–2159.
- Gatt and Krahmer (2018) Albert Gatt and Emiel Krahmer. 2018. Survey of the state of the art in natural language generation: Core tasks, applications and evaluation. Journal of Artificial Intelligence Research, 61:65–170.
- Lin (2004) Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. Text Summarization Branches Out.
- Manning et al. (2014) Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In Association for Computational Linguistics (ACL) System Demonstrations, pages 55–60.
Mnih and Gregor (2014)
Andriy Mnih and Karol Gregor. 2014.
variational inference and learning in belief networks.
Proceedings of the 31st International Conference on International Conference on Machine Learning - Volume 32, ICML’14, pages II–1791–II–1799. JMLR.org.
- Nallapati et al. (2016) Ramesh Nallapati, Bowen Zhou, Cicero dos Santos, Ça glar Gulçehre, and Bing Xiang. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. pages 280–290.
- Nguyen et al. (2016) Tri Nguyen, Mir Rosenberg, Jianfeng Gao, and Li Deng. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268.
- Papineni et al. (2002) Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318.
- Qin et al. (2018) Lianhui Qin, Lemao Liu, Wei Bi, Yan Wang, Xiaojiang Liu, Zhiting Hu, Hai Zhao, and Shuming Shi. 2018. Automatic article commenting: the task and dataset. arXiv preprint arXiv:1805.03668.
- Rajpurkar et al. (2018) Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for squad. arXiv preprint arXiv:1806.03822.
Rajpurkar et al. (2016)
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016.
questions for machine comprehension of text.
Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392.
- Rush et al. (2015) Alexander M Rush, Chopra, and Jason Weston. 2015. A neural attention model for abstractive sentence summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 379–389.
- See et al. (2017) Abigail See, Peter J Liu, and Christopher D Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073–1083.
- Seo et al. (2016) Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. 2016. Bidirectional attention flow for machine comprehension. arXiv preprint arXiv:1611.01603.
- Shum et al. (2018) Heung-Yeung Shum, Xiaodong He, and Di Li. 2018. From eliza to xiaoice: Challenges and opportunities with social chatbots. Frontiers of Information Technology & Electronic Engineering, 19(1):10–26.
- Sutskever et al. (2014) Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, NIPS’14, pages 3104–3112, Cambridge, MA, USA. MIT Press.
- Tan et al. (2018) Chuanqi Tan, Furu Wei, Nan Yang, Bowen Du, Weifeng Lv, and Ming Zhou. 2018. S-net: From answer extraction to answer generation for machine reading comprehension. In AAAI’18.
- Tang et al. (2016) Jian Tang, Yifan Yang, Sam Carton, Ming Zhang, and Qiaozhu Mei. 2016. Context-aware natural language generation with recurrent neural networks. arXiv preprint arXiv:1611.09900.
- Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 5998–6008. Curran Associates, Inc.
- Vedantam et al. (2015) Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In
- Vinyals et al. (2015) Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 2692–2700. Curran Associates, Inc.
- Wang et al. (2017) Wenhui Wang, Nan Yang, Furu Wei, Baobao Chang, and Ming Zhou. 2017. Gated self-matching networks for reading comprehension and question answering. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 189–198.
- Xiong et al. (2016) Caiming Xiong, Victor Zhong, and Richard Socher. 2016. Dynamic coattention networks for question answering. arXiv preprint arXiv:1611.01604.
- Yu et al. (2018) Adams Wei Yu, Rui Dohan, Kai Chen, Mohammad Norouzi, and Quoc V Le. 2018. Qanet: Combining local convolution with global self-attention for reading comprehension. In International Conference on Learning Representations.
- Zheng et al. (2018) Hai-Tao Zheng, Wei Wang, Wang Chen, and Arun Kumar Sangaiah. 2018. Automatic generation of news comments based on gated attention neural networks. IEEE Access, 6:702–710.
Appendix A Yahoo! News Dataset
More information about the Yahoo dataset is shown in Table 8. The side information associated with news including:
Paragraph. After pre-processing, we retain the paragraph structure of news articles.
Category. There are 31 news categories and the distribution is shown in Figure 2.
Wiki-Entities. The Wikipedia entities mentioned in the news articles are extracted.
Vote. Each comment has upvote, downvote and abusevote information from news readers.
Sentiment. Each comment is annotated with POSITIVE, NEGATIVE or NEUTRAL by Yahoo!.
|Avg. # Comments per News||27||27||27|
|Avg. #Upvotes per Comment||5.9||4.9||3.4|
|Avg. # Comments per News||20.6||20.5||20.5|
|Avg. #Upvotes per Comment||31.4||30.2||32.0|
|Avg. #DownVotes per Comment||4.8||4.8||4.9|
|Avg. #AbuseVotes per Comment||0.05||0.05||0.05|
Appendix B Case Study
We demonstrate the advantage of DeepCom over IR-TC and Att-TC with examples from the test sets of the two datasets. Table 9 and Table 10 show the comments given by the three models. The content in the red box represents salient spans predicted by the reading network and that in the blue box is generated entities included in the salient spans. We can see that compared to the two baseline methods, comments from DeepCom are more relevant to the content of the news articles. Comments from IR-TC are rich with content, but are irrelevant to the news articles, while Att-TC is prone to generate generic comments.
|Title: 唐嫣为什么不演清宫剧？ (Why Tiffany Tang never plays a role in a drama set in Qing dynasty?)|
(The numerous television series and movies Tiffany Tang acted in have made her very popular, and her varied modelling in many activities has set her an image of ”the queen of reinvention” in the hearts of the public. Her elegant and unique nature showed in commercial events and media events brings her a lot of praise and spotlight. Tiffany Tang is tall, slim, has sweet smile and a beautiful face, and is always dressed in fashion. Thus, she is called “9-heads tall” girl by her fans. If her most beautiful role is Zi Xuan in “Chinese Paladin 3”, then the most ugly one should be this! Both hair-style and makeup look terrible. She is most alluring at the time of “Chinese Paladin 3”. When she dressed up with a costume in Qing Dynasty, however, she
looks really bad, like an idiot. This is how Tiffany Tang looks like in a TV series “A Beauty in Troubled Times”, but this might be her forgettable moment. Her costume in Republic of China looks terrible too. I finally know whyTiffany Tang never plays a role in a drama set in Qing dynasty. If she does not use fake bangs, even her face is small, her forehead looks too big! What do you think?)
|DeepCom: 唐嫣罗晋的演技真的很好，特别喜欢她演[blue]《仙剑奇侠传》。我觉得这部剧挺好看。(Tiffany Tang and Luo Jin are really good actors. I especially like her role in [blue]Chinese Paladin 3. I think the TV drama is worth watching.)|
|Att-TC: 我也是醉了 (I have nothing to comment.)|
|IR-TC: 星爷和谁开撕过嘛, 都是别人去撕星爷！（Stephen Chow never fights with others. It is others that fight with Stephen Chow!）|
|Title: NBA notebook : Rockets targeting Anthony after losing Mbah a Moute|
|Body: The Houston Rockets are now determined to sign forward Carmelo Anthony after forward Luc Mbah a Moute joined the Los Angeles Clippers on a one-year, $4.3 million deal on Monday, according to an ESPN report. Anthony is currently a member of the Oklahoma City Thunder, but the two sides are reportedly working on parting ways, whether through a trade, a buyout or waiving via the stretch provision. Anthony is likely to become a free agent even if he is traded, as his new team would likely waive him. Multiple reports on Sunday said rockets guard Chris Paul wants Anthony, a good friend of his, to join the rockets, while Anthony is also believed to have interest in joining Lebron James with the Los Angeles Lakers. The Miami Heat are also reportedly interested in adding Anthony. Mbah a Moute spent the 2015-16 and 2016-17 seasons with the Clippers before joining the Rockets last season. The 31-year-old averaged 7.5 points , 3.0 rebounds and 1.2 steals in 25.6 minutes per game across 61 games (15 starts) in Houston. – The Cleveland cavaliers are looking to deal 37-year-old guard Kyle Korver and transition to a younger lineup, according to Terry Pluto of the Cleveland Plain Dealer. …Korver’s contract has $ 15.1 million remaining over the next two seasons, although only $ 3.4 million of his 2019-20 salary is guaranteed. He could prove to be a good option for a team looking for better perimeter shooting to compete for a league title. …– It ’s unclear whether Joakim Noah will remain with the New York Knicks moving forward, but the center said he hopes to say in the big apple, in a video published by TMZ. “I love New York,” Noah said, “I don’t know what’s going to happen , but Coach Fiz is cool, man.” …|
|DeepCom: the [blue]rockets are going to have a lot of fun in this series .|
|Att-TC: i think he is going to be a contender|
|IR-TC: the kings have a great future if cousins is n’t the team ’s leader . but if continues to be propped up as the guy , then the kings will only have a good future along with the bottom half of the west . cousins just is n’t enough of a leadership guy who will go through 3 more coaches before he finds a decent match not just for himself , but the team .|