Introduction
Automatically extracting useful and appealing insights, i.e. the data mining results, from a multidimensional table is a challenging yet important task in the areas of Business Intelligence (BI), Data Mining, TabletoText Generation, etc. For example, we can derive the insight ”Sales of Brand A is increasing year over year while sales of Brand B is decreasing from 2015 to 2017 in China” from a multidimensional car sales table. In this work, insight is defined as a data structure that includes subspace, type, significance value, and description. It can be described in any forms for different applications. In the whole process of automatic business data analysis, generating abundant insights from multidimensional structured data can be accomplished with elaborate predefined rules, while modeling their usefulness or interestingness and ranking the top ones are much more difficult. Handcrafted ranking rules are less efficient and cannot cover every possible situation, and therefore a learning method for insight ranking is worth studying.
Previously effort has been made to explore how to extract insights according to its statistical significance score [Tang et al.2017]. However, statistical significance has some limitation in insight importance ranking. First, as its scoring method suggests, it neglects the semantics of the data (such as the horizontal and vertical headers in the bidimensional table), which is proved to greatly contribute to the importance of data in our later experiments. As a result, insights that have a higher preference in realworld data are possible to get less attention. For example, in financial reports a statistically significant increase of “Operating Income” usually enjoys less popularity than that of a more common item “Total Revenue”, as shown in Figure 1
, but is possible to get a higher statistical significance score. Besides, the significance values of insights in different types are incomparable. It is inaccurate to rank an insight of trend (shape insight) and an insight of outliers (point insight) according to their significant values since the two significant values have their own statistical meanings under different statistical hypothesis and measurements. Also, the statistical analysis method is unsuited for small tables since it requires a minimum number of data points to calculate the statistical significance.
The main reason that the previous ranking methods are usually rulebased is that there is a lack of available training data for the insight ranking problem, as explicitly scoring an insight or giving it a rank usually requires domain knowledge and a thorough understanding of the table and context which is difficult and timeconsuming. To address this problem, we take advantage of the human written table descriptions and analytical text, and use the text as ”weak supervision” signals to learn an insight ranking model. Such texts involve latent prior common knowledge and domain knowledge and provide valuable information on what insights are more important and are more likely to be mentioned. To our best knowledge, this is the first work to explore the ranking problem of the insights with the assistance of text.
The importance of an insight can be measured in many dimensions. We find that the semantics information of insights contributes to its importance measuring. The advantage of introducing it into the ranking model is that it provides the meaning to a cell of number, and the context of the data application. Moreover, it breaks the limitation to table structure, as tables in any form and of any size can be universally represented as a list of labels and values. Inspired by this, in this paper we focus on ranking the insights by capturing both the semantic features and statistical characteristics of the data. In addition, the global table context, such as table structure and the relationship among all the insights, should also be taken into consideration. For example, a yearoveryear decreasing insight is more valuable than an increasing insight when all the other data are of increasing trends.
The challenges are threefold. First, despite its prospect, there is no existing available dataset and no annotated insight importance labels for ranking models. Second, it is hard to model the interestingness of insights as it can be measured in many dimensions. Both the content relevance and the statistical significance of insight need exploration. Third, the comparison or ranking process among insights should be done in groups. For a fair comparison, insights within one table should be compared in one group since they are closely related inherently. Therefore, the table context needs to be introduced as external information to enable the comparison of relative interestingness values in a ranking model.
To overcome the above limitations, we present a textassisted ranking model with header semantics and a global contextaware memory component. We estimate the importance of an insight according to its probability of being interpreted in the description text and feed the score into the ranking model. The ranking model consists of two parts. The neural ranking model explores the data characteristics, such as its semantics and statistics information simultaneously. The keyvalue memory network model introduces table structure information into the ranking process. The experiment results on two datasets demonstrate that our model achieves significant progress compared with baselines.
In summary, our contributions are as follows:

We formally formulate the problem of text assisted insight ranking, which has not been fully investigated yet.

We construct a new financial dataset, in which we labeled the insight importance with text assistance.

We propose a contextaware memory network to model the importance of insights. The experimental results on two datasets show that our approach significantly outperforms the baseline methods.
Related Work
Insight Ranking
Earlier works have explored the insight importance evaluation problem. Notice that the insight has different names in different studies. A broader definition of the interestingness of insights, or data mining results, is conciseness, coverage, reliability, peculiarity, diversity, novelty, surprisingness, utility, and actionability [Geng and Hamilton2006].
tang2017extracting ̵̃tang2017extracting proposes that the insight score should be applicable to and fair across different types of insight. The insight score function in their paper measures the market share and the pvalue based uncommonness significance score. In their work, different insights follow different distribution and have different null hypothesis. We argue that such statistical methods do not satisfy the comparability requirement of insight importance score.
demiralp2017foresight ̵̃demiralp2017foresight also uses predefined strength metrics for each kind of insights, such as the Pearson correlation coefficient for linear relationship insight, the number of outliers for outliers insight, and standardized skewness coefficient for skew insight. More previous works in data exploration and data mining areas measure the insight importance by how surprising that value is different from the expectation
[Wu, Sismanis, and Reinwald2007, Sarawagi, Agrawal, and Megiddo1998]. User preference is also taken into account in the area of interactive data exploration [Wasay, Athanassoulis, and Idreos2015, Dimitriadou, Papaemmanouil, and Diao2016, Çetintemel et al.2013]. Different from their work, we introduce header semantics and table context into the insight ranking process.In the task of tabletotext generation in Natural Language Processing (NLP), the generation process is divided into three modules, content planning, sentence planning, and surface realization
[Sha et al.2018, Lebret, Grangier, and Auli2016, Mei, Bansal, and Walter2016, Liu et al.2018]. Similar to our insight ranking problem, the content planning module is required to decide which parts of the input table should be paid attention to. The difference is that the selection process is not explicitly formulated as a ranking problem that assigns each candidate a significance score.Learning to Rank
The aforementioned insight importance ranking methods are mostly based on handcrafted rules, different from which our approach applies the “learning to rank” method in machine learning.
The ranking methods are usually classified into 3 categories, pointwise ranking, pairwise ranking, and listwise ranking. The pointwise approach considers the ranking problem as multiclass classification problem
[Li, Burges, and Wu2007] or regression problem [Cossock and Zhang2006]. It considers the ranked candidates as independent, and is regardless of the final ranked result. The pairwise approach considers the ranking problem as binary classification problem and classifies the candidate pairs into two categories, correctly or incorrectly ranked pairs [Burges et al.2005, Freund et al.2003, Burges, Ragno, and Le2006, Tsai et al.2007]. There is a gap between its loss function and the evaluation metrics of the ranking results. The listwise method scores the candidates within a list together and directly optimizes the evaluation metrics
[Pareek and Ravikumar2014, Cao et al.2007, Burges, Ragno, and Le2006, Xu and Li2007, Taylor et al.2008, Burges2010].Problem Formulation
Insight
Definition 1 (MultiDimensional Table). A multi dimensional table is defined as the set of data cells, i.e. . Each data cell is represented as , where is one dimension in a table, is the total number of dimensions in a table, and is the value.
For example, table 1 is a bidimensional table with dimension Brand and Year. For the cell in the up left corner, .
Brand, Year  2015  2016  2017 

A  13  14  20 
B  51  49  60 
C  13  20  23 
Definition 2 (Subspace). A subspace is defined as a set of cells that , in which at least one dimension of the cells in the subset is the same:
(1) 
where is the number of cells in the subspace, and is the th dimension in each cell .
In table 1, a subspace consists of:
(2) 
where the cells share the same dimension . The subspace is usually formed when we fixed some dimensions of the table and enumerate the combination of other chosen dimensions. The subspace usually has a particular meaning when selected. In the example, the subspace represents the sales of Brand A over years.
For each subspace, we can perform statistical test with specific hypothesis. The hypothesis is defined by insight type which includes summary statistics, correlations, outliers, empirical distributions, density functions, clusters, and so on [Demiralp et al.2017]. Under the statistical hypothesis of insight type , we can calculate the statistical significance value . If the significance value exceeds a predefined threshold, it is considered as an informative observation from the table, and we can generate a description from the header semantics for each dimension using some predefined templates, such as “Sales of A is increasing from 2015 to 2017.” in the example subspace we give in table 1.
Formally, we define the above elements as the insight:
Definition 3 (Insight). An insight is defined as four parts , where is the subspace, is the insight type, is the significance value, and is the corresponding description.
Subspace  A,2015, A,2016, A,2017 

Insight Type  Tread Increasing 
Significance Value  0.5 
Description  Sales of A is increasing year over year. 
TextAssisted Insight Ranking
From a multidimensional table, we can derive a great many insights as informative observations especially when the table has many dimensions. However, people will only pay attention to several important insights, which requires insight ranking. As introduced, it is difficult to explicitly calculate the importance of an insight directly. And human written table description and analysis text provide valuable information about what insights are more likely to be worth analyzing and which are not.
In this study, we use the assistance of the description text corresponding to a table. Suppose the description is a set of sentences . For each insight , we can calculate the similarity between the description from the insight and each sentence in the text by a similarity function , and find the most similar sentence . When the similarity score is higher than a certain threshold, we can assume that human writer does mention that insight in the text, and correspondingly, the similar human written sentence is an expression of the insight. As a result, the semantic similarity score represents the possibility of an insight’s being mentioned in the text, which further represents its importance or interestingness. The similarity can be seen as a ”week supervision” of how likely the insight will be interpreted by people in the corresponding text. Therefore, given an insight set , we can get the rank of those insights by the similarity score as .
Definition 4 (Insight Ranking). Given a set of insights from a table, we learn a ranking function
(3) 
where is the rank of insight from the ranking function, and is a listwise loss function.
In order to compare all the insights from the same table together, we build our ranking function according to the listwise ranking method in [Cao et al.2007].
Model
As shown in Figure 2, our proposed model consists of two parts. The neural ranking model explores the data characteristics, including its semantic information, insight type, statistic information and subspace, and assigns importance scores to each insight. Additionally, the keyvalue memory network model introduces other insights within one group, namely the table context, into the ranking process.
Insight Representation
We represent the insight with a vector of fixed size
. Four kinds of insight features are encoded in different ways into vectors with the same vector length .Significance value : The significant score is embedded into a vector with a fullyconnected layer.
Insight Type : We treat each insight type as a special word token and encode them using the same embedding matrix used in header semantics representation.
Subspace : The cells in a subspace is considered as a sequence of continuous cell values along with their shared dimension. The sequence is then processed by a singlelayer CNN to form the subspace representation. The CNN regards as an input channel, and alternates convolution operation.
Suppose that denotes the output of feature maps of channel. On the convolution layer, we employ a 1D convolution operation with a window size , and define as:
(4) 
where is a tanh, and are parameters. The output of the feature maps are projected to a vector of dimension
with a linear transformation.
Semantics : The semantics of an insight is expressed as the concatenation of all headers of the cells in a insight subspace, which produces a sequence of word tokens
(5) 
where is the length of the headers. Then the distributed semantics representation is defined as a bagofwords using the embedding matrix :
(6) 
where maps the tokens to a bag of dimension (the vocabulary size), and is a matrix.
Finally, the feature of an insight is represented by summing up the four features:
(7) 
KeyValue Memory Network
Our intuition is to introduce the table context such as table structure and the relations between insights in the same table into the ranking model. We represent the table as a set of insights extracted from the table.
Since the insights are not naturally expressed as sorted sequence, a memorylike framework is more appropriate than structuresensitive models such as RNN and CNN. Assuming that relation between insights can be revealed by their header semantics, we apply a keyvalue memory network (KVMemNN) [Miller et al.2016] to search semantically similar insights for each insight candidate.
We define the memory slots as a vector of pairs
(8) 
where there are related insights, is the th insight and is the semantic vector of insight . We denote the semantic of current insight as query . The key addressing and reading of the memory involves the following two steps.
Key Addressing: During addressing, we perform a selfattention operation by computing the inner product between and the memory keys followed by a softmax:
(9) 
which yields a vector of attention weights over the semantics of related insights.
Value Reading: In the reading step, the values of the memories (insight representations) are read by taking their weighted sum using the addressing attentions, and the memory output vector is returned as:
(10) 
The final insight representation will be an input of the ranking model described in the next section.
Since the representation of the insight itself is also contained in the memory, it will definitely produce very high attention to address the insight self. We do not concatenate the output of the memory with other feature vectors as the other memory network often does.
Ranking Model
The model is implemented as a multilayer perceptron (MLP) which receives insight representations and outputs the ranking scores of the insights.
The model is trained by minimizing the L2 loss of the output scores and the similarity scores of the insights:
(11)  
where and are the model outputs and groundtruth scores, respectively. We apply the listwise approach and sums up the total losses of the insights in the same table, as the total loss relies on the table context. For the baseline models without the memory network, we apply the pointwise approach and calculate the L2 loss for each insight as a training sample.
Dataset
Financial Report Dataset
The financial report dataset is built upon the public annual and quarterly reports from United States Securities and Exchange Commission^{1}^{1}1https://www.sec.gov/edgar/searchedgar/companysearch.html. The dataset contains in total 5,670 reports and 49,129 tables from 2,762 companies. Table 3 summarizes the data statistics. In the experiment, we randomly split the dataset into training, validation, and test sets consisting of 60%, 20%, and 20% summaries, respectively.
We filtered the sentences out that are less than 50 characters or 10 words, and those do not include any numbers or keywords. Year information is substituted with “this year”, “last year”, and so on. More detailed date information is deleted as we only consider annual report. Special tokens are also processed to avoid noise.
Mean  Percentile  

5%  95%  
# tokens per cell  5.29  1  12 
# tokens per sentence  32.36  15  64 
# sentences per report  774.98  282  1434 
SBNation Dataset
To validate the generality of our model, we also evaluate the effectiveness of our model on SBNation Dataset from [Wiseman, Shieber, and Rush2017]. This dataset consists of 10,903 humanwritten NBA basketball game summaries aligned with their corresponding boxscores and linescores. We randomly split the dataset into training, validation, and test sets consisting of 60%, 20%, and 20% summaries.
Insight Extraction
We defined two representative types of insight in this work:

Point insight: we measure how outstanding the data point is among all the data points in the subspace.

Shape insight (trend): we detect the rising or falling trend among a series of data points.
In the financial dataset, the “point” is defined as the change ratio of one item from the current year to last year in the point insight. The “trend” is defined as the increasing or decreasing trend yearoveryear in the shape insight.
In the SBNation dataset, we only extract the point insight. The “point” is defined as the one of the game statistic such as scores, rebounds and assistants of a player.
The significance score of each insight type is calculated with the same approach described in [Tang et al.2017].
Similarity Function
We propose two similarity functions here, and , and select the better one by human evaluation.
First, we count same words in the insight description and the human written sentence :
(12) 
where is the count of same words, represents the length of .
To assign more weights to the words in the headers, we calculate the similarity of a header and a sentence :
(13) 
where represents the percentage of the number of words matched in the header, and is the normalization factor.
We add the similarity of headers to the similarity of sentences to construct the second similarity function:
(14) 
where and are the weights. In this paper, we set them both to .
Finally, the maximum similarity score among the insight description and all the candidate sentences in the text represents the probability of the insight’s being interpreted in the report, which is further used as the guideline for ranking.
Text Assistance
To test the effectiveness of the text assistance method, we randomly sample 4,000 pairs of insights and their most similar sentences in the reports, and ask 10 human annotators who are familiar with financial reports to label whether the pairs are of the same meaning. The evaluation data is equally split into three groups according to their similarity scores.
As shown in figure 3, for both similarity functions, the higher similarity is the higher accuracy of their having the same meaning is. Therefore we can use the similarity score as the ground truth of the insight importance.
In addition, we find that performs better than . It obtains nearly accuracy for the high similarity group. The reason is that emphasizes the headers explicitly compared to the . For the low similarity group, the accuracy of the similarity function with headers is lower than that with sentences, which indicates that the similarity function with headers is more distinguishable.
According to the human evaluation, in the later experiments we will use as the similarity function.
Experiment
Experiment Settings
Based on the performance on the validation set, we set the embedding size to for the baseline methods and the proposed model. The vocabulary sizes in the financial report dataset and the SBNation dataset are 8,409 and 900, respectively.
The parameters are updated by Adam algorithm [Kingma and Ba2014]
on a single 1080 Ti GPU and initialized by sampling from the uniform distribution (
). The initial learning rate is . The model is trained in minibatches with a batch size of .Evaluation Metrics
We report the ranking accuracy in three evaluation metrics: Mean Average Precision(mAP)@k, Normalized Discounted Cumulative Gain(NDCG)@k, and Precision@k.
Precision@1  Precision@3  Precision@5  mAP@3  mAP@5  NDCG@3  NDCG@5  

0.098  0.246  0.399  0.474  0.624  0.646  0.688  
0.107  0.249  0.408  0.473  0.621  0.649  0.692  
0.110  0.261  0.416  0.481  0.632  0.658  0.703  
0.118  0.278  0.444  0.525  0.686  0.738  0.757  
0.162  0.411  0.605  0.668  0.756  0.799  0.815  
0.170  0.425  0.626  0.684  0.772  0.812  0.829 
Comparing Methods
We first compare three significant score calculation methods. The detailed calculation methods follow the definition of point insight and shape insight in [Tang et al.2017].

calculates the significance from the data distributions in one table. It represents the insight importance when the insights are compared to other insights in the same table.

calculates the significance from the data distributions in all tables. We assume that all tables are inherently related to each other.

first clusters the subspaces of all the insights in the dataset using the word embedding of the headers, then calculate the significance score of the data distributions in one cluster. We employ the KMeans method for clustering, and
is set to for the best performance.
We also implement the Text Assisted Ranking (TAR) model with different components.

adds the CNN to capture more statistical features in addition to the basic insight significance and insight type features.

adds the table header as semantics information to the input in addition to the .

adds the memory component to the to introduce the table context and relations among the insights.
Experiment Results and Analysis
Financial Report Dataset
Evaluation results on financial report dataset are shown in Table 4. In general, our proposed method achieves the best overall performance, which demonstrates its ability to fully explore the insight characteristics and modeling the insight importance.
We first compare the three baseline methods, , and , which calculate the significance scores in different ways. The performance of is slightly better than that of , as the former method calculate the significance with a much larger space of data points. The comparison result also supports our assumption that the statistical significance score method does not suit for small tables, as the significance score is unreliable while there are only very few insights from a table. The cluster method achieves the best result, which demonstrates the importance of the header semantics since it is the clustering rule. According to the result, we use the as significance score in the models.
A series of incremental experiments are conducted to evaluate the contributions of the key components in our proposed model. Three versions of model in incremental sequence, , and , are provided. is a basic version that explores the insight type, the significance score and the subspace of insights. By introducing the subspace information, the model is exposed to more available information on the statistical data distribution instead of a single significance score, and slightly improves the ranking performance.
Comparing to the gap between and , the improvement between and is much more obvious. The result suggests that the semantics is an important factor when we determine the importance value of an insight. Explicitly introducing the semantics largely enriches the insight representation space and improve the ranking performance significantly.
The model, the complete version of our proposed model, achieves the best performance in all evaluation metrics. Compared with , the introduces the related insight information within one group for comparison. The result supports our assumption that global table context and grouped insight relationship make a contribution to the process of insight ranking.
Human Evaluation
Precision@1  Precision@3  Precision@5  

0.727  0.629  0.540  
0.886  0.813  0.745 
We randomly sample 400 tables and ask the human experts to determine if the topk insights and their most similar descriptions in the report are of the same meanings. The result in Table 5 implies that the recommendations of the insights according to our ranking model are of high accuracy and reliability.
SBNation Dataset
Precision@1  Precision@3  

0.503  0.513  
0.788  0.754  
0.797  0.759 
The experimental result on SBNation dataset is shown in Table 6. Different from the annual financial report dataset, the description in SBNation is much more rigid and lacks variation. Therefore we consider label matched sentences as the target, and mark insight importance as 0or1, either relevant or irrelevant, rather than continuous 0to1 values. NDCG and mAP cannot adapt to such labels in ranking problems. The value in Precision@k is set to 1 and 3, as the tables in SBNation are relatively smaller and most of them contain only 3 to 4 insights. Similar to the results in financial report dataset, the achieves the best performance.
Case Study
Insight Descriptions  Gold  TAR 

Collaboration and license revenue was 71.7  1  2 
million for the year ended, an increase of  
58.7 million compared to the year ended.  
General and administrative expenses were  2  4 
27.8 million for the year ended, an increase  
of 18.8 million compared to the year ended.  
Research and development expenses were  3  1 
58.6 million for the year ended, an increase  
of 35.1 million compared to the year ended.  
We had 111 fulltime employees including  4  9 
82 employees engaged in development.  
The net valuation allowance increased by  5  3 
4.9 million and 0.6 million respectively. 
We present a ranking result example in Table 7. It consists of the top 5 insights in 10 insight candidates from one table.
The Precision@5 is , a relatively high accuracy. The more detailed relative position of the top 5 insights is of less usefulness. Because the target ranking results only represent the probability of the insight’s being in the text, and the importance of the top 5 insights are of little distinction.
The reason why the fourth insight is wrongly labeled is that the similarity score is incorrectly calculated and the gold standard is in fact inaccurate. We analyzed the insight description and found that the sentence is coincidently matched with a wrong insight because it contains some keywords in the headers and similar numbers. This serves as an example of the optimization direction of the text assisted approach. We would like to solve this problem by introducing the position of the sentence to the text assistance to derive more accurate similarity function.
Conclusion
In this paper, we propose a contextaware memory network to rank the insight importance. The model explores the data characteristics and introduces table structure and semantics information into the ranking process. We construct a financial report dataset, in which the insight interestingness inferred from the human written description is used as annotated training data. Experimental results show that our approach largely improves the ranking precision.
In the future, we would like to investigate a more reliable similarity function to take the sentence position into account. Also, instead of text assistance, we can explore more methods such as figure assistance and metadata assistance to estimate the approximate score of the insight importance.
References
 [Burges et al.2005] Burges, C. J. C.; Shaked, T.; Renshaw, E.; Lazier, A.; Deeds, M.; Hamilton, N.; and Hullender, G. N. 2005. Learning to rank using gradient descent. In Machine Learning, Proceedings of the TwentySecond International Conference (ICML 2005), Bonn, Germany, August 711, 2005, 89–96.
 [Burges, Ragno, and Le2006] Burges, C. J. C.; Ragno, R.; and Le, Q. V. 2006. Learning to rank with nonsmooth cost functions. In Advances in Neural Information Processing Systems 19, Proceedings of the Twentieth Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 47, 2006, 193–200.
 [Burges2010] Burges, C. J. 2010. From ranknet to lambdarank to lambdamart: An overview. Learning 11(23581):81.
 [Cao et al.2007] Cao, Z.; Qin, T.; Liu, T.; Tsai, M.; and Li, H. 2007. Learning to rank: from pairwise approach to listwise approach. In Machine Learning, Proceedings of the TwentyFourth International Conference (ICML 2007), Corvallis, Oregon, USA, June 2024, 2007, 129–136.
 [Çetintemel et al.2013] Çetintemel, U.; Cherniack, M.; DeBrabant, J.; Diao, Y.; Dimitriadou, K.; Kalinin, A.; Papaemmanouil, O.; and Zdonik, S. B. 2013. Query steering for interactive data exploration. In CIDR 2013, Sixth Biennial Conference on Innovative Data Systems Research, Asilomar, CA, USA, January 69, 2013, Online Proceedings.
 [Cossock and Zhang2006] Cossock, D., and Zhang, T. 2006. Subset ranking using regression. In Learning Theory, 19th Annual Conference on Learning Theory, COLT 2006, Pittsburgh, PA, USA, June 2225, 2006, Proceedings, 605–619.
 [Demiralp et al.2017] Demiralp, Ç.; Haas, P. J.; Parthasarathy, S.; and Pedapati, T. 2017. Foresight: Rapid data exploration through guideposts. CoRR abs/1709.10513.

[Dimitriadou, Papaemmanouil, and
Diao2016]
Dimitriadou, K.; Papaemmanouil, O.; and Diao, Y.
2016.
AIDE: an active learningbased approach for interactive data exploration.
IEEE Trans. Knowl. Data Eng. 28(11):2842–2856.  [Freund et al.2003] Freund, Y.; Iyer, R. D.; Schapire, R. E.; and Singer, Y. 2003. An efficient boosting algorithm for combining preferences. Journal of Machine Learning Research 4:933–969.
 [Geng and Hamilton2006] Geng, L., and Hamilton, H. J. 2006. Interestingness measures for data mining: A survey. ACM Comput. Surv. 38(3):9.
 [Kingma and Ba2014] Kingma, D. P., and Ba, J. 2014. Adam: A method for stochastic optimization. CoRR abs/1412.6980.
 [Lebret, Grangier, and Auli2016] Lebret, R.; Grangier, D.; and Auli, M. 2016. Neural text generation from structured data with application to the biography domain. 1203–1213.

[Li, Burges, and
Wu2007]
Li, P.; Burges, C. J. C.; and Wu, Q.
2007.
Mcrank: Learning to rank using multiple classification and gradient boosting.
In Advances in Neural Information Processing Systems 20, Proceedings of the TwentyFirst Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 36, 2007, 897–904.  [Liu et al.2018] Liu, T.; Wang, K.; Sha, L.; Chang, B.; and Sui, Z. 2018. Tabletotext generation by structureaware seq2seq learning.
 [Mei, Bansal, and Walter2016] Mei, H.; Bansal, M.; and Walter, M. R. 2016. What to talk about and how? selective generation using lstms with coarsetofine alignment. 720–730.
 [Miller et al.2016] Miller, A. H.; Fisch, A.; Dodge, J.; Karimi, A.; Bordes, A.; and Weston, J. 2016. Keyvalue memory networks for directly reading documents. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 14, 2016, 1400–1409.
 [Pareek and Ravikumar2014] Pareek, H. H., and Ravikumar, P. 2014. A representation theory for ranking functions. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 813 2014, Montreal, Quebec, Canada, 361–369.
 [Sarawagi, Agrawal, and Megiddo1998] Sarawagi, S.; Agrawal, R.; and Megiddo, N. 1998. Discoverydriven exploration of OLAP data cubes. In Advances in Database Technology  EDBT’98, 6th International Conference on Extending Database Technology, Valencia, Spain, March 2327, 1998, Proceedings, 168–182.
 [Sha et al.2018] Sha, L.; Mou, L.; Liu, T.; Poupart, P.; Li, S.; Chang, B.; and Sui, Z. 2018. Orderplanning neural text generation from structured data.
 [Tang et al.2017] Tang, B.; Han, S.; Yiu, M. L.; Ding, R.; and Zhang, D. 2017. Extracting topk insights from multidimensional data. In Proceedings of the 2017 ACM International Conference on Management of Data, SIGMOD Conference 2017, Chicago, IL, USA, May 1419, 2017, 1509–1524.
 [Taylor et al.2008] Taylor, M. J.; Guiver, J.; Robertson, S.; and Minka, T. 2008. Softrank: optimizing nonsmooth rank metrics. In Proceedings of the International Conference on Web Search and Web Data Mining, WSDM 2008, Palo Alto, California, USA, February 1112, 2008, 77–86.
 [Tsai et al.2007] Tsai, M.; Liu, T.; Qin, T.; Chen, H.; and Ma, W. 2007. Frank: a ranking method with fidelity loss. In SIGIR 2007: Proceedings of the 30th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Amsterdam, The Netherlands, July 2327, 2007, 383–390.
 [Wasay, Athanassoulis, and Idreos2015] Wasay, A.; Athanassoulis, M.; and Idreos, S. 2015. Queriosity: Automated data exploration. In 2015 IEEE International Congress on Big Data, New York City, NY, USA, June 27  July 2, 2015, 716–719.
 [Wiseman, Shieber, and Rush2017] Wiseman, S.; Shieber, S. M.; and Rush, A. M. 2017. Challenges in datatodocument generation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP 2017, Copenhagen, Denmark, September 911, 2017, 2253–2263.
 [Wu, Sismanis, and Reinwald2007] Wu, P.; Sismanis, Y.; and Reinwald, B. 2007. Towards keyworddriven analytical processing. In Proceedings of the ACM SIGMOD International Conference on Management of Data, Beijing, China, June 1214, 2007, 617–628.
 [Xu and Li2007] Xu, J., and Li, H. 2007. Adarank: a boosting algorithm for information retrieval. In SIGIR 2007: Proceedings of the 30th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Amsterdam, The Netherlands, July 2327, 2007, 391–398.
Comments
There are no comments yet.