{ "paper_id": "2020", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:27:47.746877Z" }, "title": "Market Comment Generation from Data with Noisy Alignments", "authors": [ { "first": "Yumi", "middle": [], "last": "Hamazono", "suffix": "", "affiliation": { "laboratory": "", "institution": "Ochanomizu University", "location": {} }, "email": "hamazono.yumi@is.ocha.ac.jp" }, { "first": "Yui", "middle": [], "last": "Uehara", "suffix": "", "affiliation": {}, "email": "yui.uehara@aist.go.jp" }, { "first": "Hiroshi", "middle": [], "last": "Noji", "suffix": "", "affiliation": {}, "email": "noji@aist.go.jp" }, { "first": "Yusuke", "middle": [], "last": "Miyao", "suffix": "", "affiliation": {}, "email": "yusuke@is.s.u-tokyo.ac.jp" }, { "first": "Hiroya", "middle": [], "last": "Takamura", "suffix": "", "affiliation": {}, "email": "takamura.hiroya@aist.go.jp" }, { "first": "Ichiro", "middle": [], "last": "Kobayashi", "suffix": "", "affiliation": { "laboratory": "", "institution": "Ochanomizu University", "location": {} }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "End-to-end models on data-to-text learn the mapping of data and text from the aligned pairs in the dataset. However, these alignments are not always obtained reliably, especially for the time-series data, for which real time comments are given to some situation and there might be a delay in the comment delivery time compared to the actual event time. To handle this issue of possible noisy alignments in the dataset, we propose a neural network model with multitimestep data and a copy mechanism, which allows the models to learn the correspondences between data and text from the dataset with noisier alignments. We focus on generating market comments in Japanese that are delivered each time an event occurs in the market. The core idea of our approach is to utilize multitimestep data, which is not only the latest market price data when the comment is delivered, but also the data obtained at several timesteps earlier. On top of this, we employ a copy mechanism that is suitable for referring to the content of data records in the market price data. We confirm the superiority of our proposal by two evaluation metrics and show the accuracy improvement of the sentence generation using the time series data by our proposed method.", "pdf_parse": { "paper_id": "2020", "_pdf_hash": "", "abstract": [ { "text": "End-to-end models on data-to-text learn the mapping of data and text from the aligned pairs in the dataset. However, these alignments are not always obtained reliably, especially for the time-series data, for which real time comments are given to some situation and there might be a delay in the comment delivery time compared to the actual event time. To handle this issue of possible noisy alignments in the dataset, we propose a neural network model with multitimestep data and a copy mechanism, which allows the models to learn the correspondences between data and text from the dataset with noisier alignments. We focus on generating market comments in Japanese that are delivered each time an event occurs in the market. The core idea of our approach is to utilize multitimestep data, which is not only the latest market price data when the comment is delivered, but also the data obtained at several timesteps earlier. On top of this, we employ a copy mechanism that is suitable for referring to the content of data records in the market price data. We confirm the superiority of our proposal by two evaluation metrics and show the accuracy improvement of the sentence generation using the time series data by our proposed method.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "In recent time, various industries such as finance, pharmaceuticals and telecommunications have increased opportunities to treat large-scale data. Hence, there is an increasing demand to automatically generate a text from large and complex data. In recent studies, neural network-based models have achieved significant progress on the data-to-text which is a text generation task from input data (Puzikov and Gurevych, 2018; Iso et al., 2019) .", "cite_spans": [ { "start": 396, "end": 424, "text": "(Puzikov and Gurevych, 2018;", "ref_id": "BIBREF26" }, { "start": 425, "end": 442, "text": "Iso et al., 2019)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "One important issue in constructing a dataset for data-to-text is to obtain the correct alignment between data and text. It is not very problematic when there is a clear correspondence between data and text, for example, when the text is manually provided by an annotator for each input, including E2E NLG Challenge dataset (Puzikov and Gurevych, 2018) . However, this is often not a trivial problem in the wild, in particular for the application of real-time text generation, such as sequential comment generation on sports games (Taniguchi et al., 2019) and stock markets (Murakami et al., 2017) , for which we can only obtain a loose alignment of data and text. This problem has been taken into account in a classical task (Chen and Mooney, 2008) , but has been overlooked in the recent neural-based models cited above.", "cite_spans": [ { "start": 324, "end": 352, "text": "(Puzikov and Gurevych, 2018)", "ref_id": "BIBREF26" }, { "start": 531, "end": 555, "text": "(Taniguchi et al., 2019)", "ref_id": "BIBREF32" }, { "start": 574, "end": 597, "text": "(Murakami et al., 2017)", "ref_id": "BIBREF20" }, { "start": 726, "end": 749, "text": "(Chen and Mooney, 2008)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In Murakami et al. (2017) , for example, they constructed a dataset for market comment generation in Japanese from a chart of stock price trends and its market reports, wherein the alignments between data and texts sometimes deviate. Figure 1 presents examples of data which are a five-minute chart of Nikkei 225 (Nikkei Stock Average) and comments describing the chart trends. All of the (I) to (III) comments are about an event at \"9:00 am, 29th of January\". However, since there is often a delay in comment delivery time for an event (e.g., (II) and (I II)), if a different movement occurs during this time period, the expressions in comment may not exactly reflect the movement at the delivery time. The word \"rebound\" indicates a downward and then upward movement on three points (the closing prices of the last two days and the latest price). This is valid for the prices at (I) and (II), but does not hold at (III) because the latest price (17039.22 yen) is lower than the last closing price (17041.45 yen). In addition, the expression \"gains 88 yen\" is only valid at the opening time (I) and is not valid at (II). To deal with these inconsistencies, the models have to be aware of these possible mismatch of data and text due to the delay, but a simple encoder-decoder-based 2017, which does not tell the difference between the event and delivery times, may perform an undesirable generalization between data and text. In this study, we extend Murakami et al. (2017) 's model with a new architecture to solve the problem due to the noisy alignments. The presented architecture is a multi-timestep architecture, which employs multiple input vectors to compensate the lack of information about the actual event time, treating it as a hidden variable and learning the correspondences from the ambiguous data. Our model employs a copy mechanism to generate a price value in a text, in which an attention weight to the time can be regarded as an induced alignment between the text and actual event time. The experimental results reveal that our proposed model outperforms the existing model in terms of the correctness of market price movement expressions, in addition to the BLEU scores.", "cite_spans": [ { "start": 3, "end": 25, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" }, { "start": 544, "end": 548, "text": "(II)", "ref_id": null }, { "start": 1452, "end": 1474, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" } ], "ref_spans": [ { "start": 234, "end": 242, "text": "Figure 1", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this section, we explain the model proposed by Murakami et al. (2017) , which be used as a base model. They proposed a model for generating a market comment, which is a news headline about the movement of the Nikkei 225, from the timeseries of the stock price of the Nikkei 225. Their model is based on the encoder-decoder (Sutskever et al., 2014) .", "cite_spans": [ { "start": 50, "end": 72, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" }, { "start": 326, "end": 350, "text": "(Sutskever et al., 2014)", "ref_id": "BIBREF30" } ], "ref_spans": [], "eq_spans": [], "section": "Generating Market Comments", "sec_num": "2" }, { "text": "Murakami et al. (2017) use both a long-term vector x long and a short-term vector x short . To capture the long-term price movement, they use a vector consisting of the closing prices of the preceding trading days represented as x long = ( long, 1 , long, 2 , . . . , long, ). Similarly, to capture the short-term price movement, they use a vector consisting of the prices of previous timesteps on the five-minute chart, starting from the comment delivery time (e.g., 10:00 am), represented as x short = ( short, 1 , short, 2 , . . . , short, ). Each of these vectors undergoes the following two preprocessing steps :", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Base Model", "sec_num": "2.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "std = \u2212 ,", "eq_num": "(1)" } ], "section": "Base Model", "sec_num": "2.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "norm = 2 \u00d7 move \u2212 (\u00af max +\u00af min ) max \u2212\u00af min ,", "eq_num": "(2)" } ], "section": "Base Model", "sec_num": "2.1" }, { "text": "where is the -th element of x. and are the mean and the standard deviation of the values in the training data, respectively. move is defined to be \u2212 , where is the closing price of the previous trading day.\u00af max and\u00af min are the maximum and the minimum of move , respectively. Equation 1is a standardization method and Equation (2) is a normalization method for moving reference. By applying these preprocessings to x long and x short , we obtain x std long ,x norm long , x std short , and x norm short . In the encoding step, these vectors are passed to multi-layer perceptrons (MLPs) to obtain the vectors h std long , h norm long , h std short , and h norm short . These are then concatenated as h long = [h std long ; h norm long ] and h short = [h std short ; h norm short ]. These vectors are combined to obtain the hidden state m of the encoder:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Base Model", "sec_num": "2.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "m = W [h long ; h short ] + b .", "eq_num": "(3)" } ], "section": "Base Model", "sec_num": "2.1" }, { "text": "In the decoding step, Murakami et al. (2017) set the initial hidden state s 0 of the decoder as m above, and use LSTM cells (Hochreiter and Schmidhuber, 1997) . They further use additional vectors called time embeddings t. To obtain t, time is discretized into intervals of one hour (e.g., 9:00 am to 10:00 am), and an embedding is obtained for each interval; 9:10 am and 9:30 am are associated with the same embedding. The time embedding of the interval, wherein the delivery time of the comment falls into is used as an additional input in each step of LSTM:", "cite_spans": [ { "start": 22, "end": 44, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" }, { "start": 124, "end": 158, "text": "(Hochreiter and Schmidhuber, 1997)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Base Model", "sec_num": "2.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "s = LSTM( [t ; w \u22121 ], s \u22121 ).", "eq_num": "(4)" } ], "section": "Base Model", "sec_num": "2.1" }, { "text": "As in the standard LSTM decoder, this output is fed into a linear layer, followed by a softmax layer to calculate the next word probability.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Base Model", "sec_num": "2.1" }, { "text": "Market comments often mention numerical values including the market prices themselves and the values obtained through arithmetic operations such as difference and rounding. To allow the model to generate numerical values with such computation during decoding, Murakami et al. (2017) introduce generalization tags (Table 1) , which specify which operation should be performed to obtain a value.", "cite_spans": [ { "start": 260, "end": 282, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" } ], "ref_spans": [ { "start": 313, "end": 322, "text": "(Table 1)", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Generalization Tags for Estimation of Arithmetic Operations", "sec_num": "2.2" }, { "text": "In Murakami et al. (2017) , for simplicity, the input value to these operations is fixed as the short,1 , the first (latest) price at the delivery time, which is converted to and in Table 1 (see caption) . However, this simplification ignores the possible mismatch of the delivery and event times (Section 1). We extend this method with a variant of a copy mechanism (Section 3.2), in which the event time is Murakami et al. (2017) used MLPs, convolutional neural networks (CNNs) and long short-term memory networks (LSTMs). We use MLPs in the encoder for our baseline model because the differences of the performance were small in their experiment.", "cite_spans": [ { "start": 3, "end": 25, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" }, { "start": 409, "end": 431, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" } ], "ref_spans": [ { "start": 182, "end": 203, "text": "Table 1 (see caption)", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Generalization Tags for Estimation of Arithmetic Operations", "sec_num": "2.2" }, { "text": "Arithmetic operation Return Round down to the nearest 10 Round down to the nearest 100 Round up to the nearest 10 Round up to the nearest 100 Return as it is Round down to the nearest 100 Round down to the nearest 1,000 Round down to the nearest 10,000 Round up to the nearest 100 Round up to the nearest 1,000 Round up to the nearest 10,000 softly predicted with attention and each numerical value is generated on demand during decoding with an operation in Table 1 .", "cite_spans": [], "ref_spans": [ { "start": 618, "end": 625, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Tag", "sec_num": null }, { "text": "Murakami et al. 2017's model put an assumption that the event and delivery times are identical. This assumption simplifies the task and thus Murakami et al. 2017propose a model with a basic encoderdecoder architecture. However, due to the time gap between the actual event time and delivery time, this assumption is not realistic. We extend the Murakami et al. (2017) 's model with a multi-timestep architecture, aiming at solving that problem occurred by the noisy alignments. Figure 2 presents our model architecture. To compensating the lack of information caused by the time gap mentioned above, we extend the encoder with additional input vectors. Each additional vector corresponds to x short starting from preceding timesteps instead of the delivery time. This allows us to treat the actual event time as a latent variable. On top of this, we introduce a copy mechanism with attention in the decoder, which facilitates learning correspondences between data and text from the noisy training data.", "cite_spans": [ { "start": 345, "end": 367, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" } ], "ref_spans": [ { "start": 478, "end": 486, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Multi-timesteps for Time-series Data", "sec_num": "3" }, { "text": "For long-term stock prices, we use a long-term vector x long following Murakami et al. (2017) . For short-term stock prices, instead of x short alone, we use x short-0step ,. . . , x short-step which we abbreviate as x 0 ,. . . , x for brevity. Each x is a short-term vector consisting of the stock prices of timesteps, Figure 2 : Overview of Multi-timesteps for Time-series Data which, instead of starting from the comment delivery time, starts from the time steps prior to the delivery time. When the delivery time is 10:00 am, for example, since we use the five-minute chart, our short-term vectors comprise of + 1 vectors corresponding to the vectors starting from 10:00 am, 9:55 am, 9:50 am, and so on.", "cite_spans": [ { "start": 71, "end": 93, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" } ], "ref_spans": [ { "start": 320, "end": 328, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "We note that each interval of five minutes is not simply associated with one value, but rather with four different values: open (the price at the very beginning of the interval), low (the lowest price in the interval), high (the highest price in the interval), and close (the price at the last of the interval). We encode these as four different vectors x open , x low , x high , and x close . Thus, there are in total 4( + 1) \u2212 1 vectors for short-term prices. In the following, we use\u02dc \u2208 [0, 4( + 1)) as an index for these vectors.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "Each of the input vectors undergoes the preprocessing methods as Murakami et al. (2017) . For long-term vectors, we obtain x std long and x norm long in the same way. Similarity, for each\u02dc , we obtain x std and x norm from x\u02dc . Given these, each MLP emits the corresponding hidden states h long and h\u02dc . Murakami et al. (2017) use only the close prices. We use four values instead since prices could largely move even in a single interval.", "cite_spans": [ { "start": 65, "end": 87, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" }, { "start": 304, "end": 326, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "Following Equation 3, we also obtain the encoder hidden state m with h long and h short . The vector h short depends on a sequence of h\u02dc to which an encoder maps the input data. It is computed as a weighted sum of h\u02dc as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "h short = 4( +1)\u22121 \u02dc =0 \u02dc h\u02dc .", "eq_num": "(5)" } ], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "The weight \u02dc of each h\u02dc is computed by:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\u02dc = exp( \u02dc )/ 4( +1)\u22121 =0 exp( ).", "eq_num": "(6)" } ], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "e = ( 0 , \u2022 \u2022 \u2022 , 4( +1)\u22121 ) is obtained by MLP [t ; a ; f 0 ; h 0 ; . . . ; f 4( +1)\u22121 ; h 4( +1)\u22121 ] .", "eq_num": "(7)" } ], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "\u02dc scores the importance of x\u02dc , based on the hidden states h\u02dc , time embedding vector t (Section 2.1), and additional two kinds of vectors, five-minute time embedding vectors f\u02dc and article-type embedding vector a.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "f\u02dc is an embedding to identify\u02dc , which maps from the starting time and the kind of price (e.g., open) to a vector. a is a vector obtained whether the comment is regular or irregular. The motivation to use these vectors is that an important x\u02dc is primarily determined by either the price history (encoded by h\u02dc ), or the delivery time (encoded by f\u02dc ), depending on the article type. For instance, when the delivery time is the same as the market opening time, that is, 9:00 am, and the comment is a regular comment, then the comment usually mentions the price at 9:00 am (e.g., \"The opening price is 15,430 yen.\"). In this case, the important x\u02dc is primarily determined by the time encoded by f\u02dc ; that is, for regular comments, the event times are rather fixed regardless of the variations in delivery times. On the other hand, even if the delivery time is the same as the market opening time, when the comment is irregular, the comment mentions some distinguished price movement (e.g. \"The price is over 100 yen higher than the last closing price.\") rather than the mere price at 9:00 am. In this case, the important x\u02dc would be determined by the price history itself. We expect these additional vectors to provide a useful inductive bias for a model to learn those distinctions.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Encoder with Multi-timesteps", "sec_num": "3.1" }, { "text": "We adapt a copy mechanism (Gu et al., 2016) in our decoder to generate numerical values by attending to the input. Recall that in the current task, the values in the output text usually do not appear in the input data (Section 2.2); rather, they can be obtained by applying an arithmetic operation to the certain value in the data.", "cite_spans": [ { "start": 26, "end": 43, "text": "(Gu et al., 2016)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "We generate a numerical value by an extension of a copy mechanism, wherein a value is generated by applying one of the operations in Table 1 to a data point \u02dc ,1 , which is the first value (latest price) of x\u02dc . Denoting an arithmetic operation as \u2208 { 1 , \u2022 \u2022 \u2022 , 12 }, the value is identified by a pair ( ,\u02dc ). We reduce the generation of numerical values to identification of these pairs, followed by the execution of an operation.", "cite_spans": [], "ref_spans": [ { "start": 133, "end": 140, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "We wish to obtain a probability distribution on numerical values that are determined by ( ,\u02dc ). One consideration for obtaining this is that there can be multiple pairs of ( ,\u02dc ) that result in the same value. For example, for an input \u02dc ,1 = = 3200, 6 7 , and 10 all result in the same value (3200). In practice, we obtain a probability to generate a numerical value by a weighted sum of scores for ( ,\u02dc ), for which ( \u02dc ,1 ) = holds, according to the obtained weights \u02dc in the encoder, which we regard as attention. (\u2022) denotes execution of operation to the input. We note that unlike the standard copy mechanism, we use the fixed attention weights \u02dc throughout the decoder. This is because, in a headline comment, which is our target, the event to be mentioned would not change throughout a single piece of text. In other words, the important x\u02dc would not change throughout a comment.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "Our model generates all numerical values in the text with this copy mechanism. To do this, we exclude numerical values from the vocabulary of the model. To switch the copy mode and noncopy mode, we add a special token \"\", which is inserted before every numerical value in the training data and indicates that the next token is a price value. Utilizing \"\", we define each conditional probability of generating target word at time as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "( | < , m) = ( | < , m), ( \u22121 = \"\") ( | < , m), (otherwise) (8)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "where", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "( |\u2022) and ( |\u2022)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "are obtained by the generation mode and the copy mode, respectively. This method is inspired by Pointer-generator network introduced by See et al. (2017). These two probabilities are defined as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( |\u2022) = [softmax(W v + b )]", "eq_num": "(9)" } ], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( |\u2022) = \u02dc , : ( \u02dc ,1 )= (\u02dc , )", "eq_num": "(10)" } ], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "(\u02dc , ) = \u02dc \u2022 [softmax(W c + b )] .", "eq_num": "(11)" } ], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "v and c are both obtained from the output of decoder LSTM, s , at each step : ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "s = LSTM( [t ; w \u22121 ; q \u22121 ], s \u22121 ) (12) v = W \u210e s + b \u210e (13) c = MLP(s ).", "eq_num": "(14)" } ], "section": "Decoder with Copy Mechanism", "sec_num": "3.2" }, { "text": "All MLPs in the model are three layers with hidden dimension of 256. The decoder LSTM is a single layer with hidden dimensions of 256. For the length of short-and long-term vectors, we set = 7 for x long and = 62 for x\u02dc , following Murakami et al. (2017) , changing the range of by setting \u2208 [0, 6].", "cite_spans": [ { "start": 232, "end": 254, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "4.2" }, { "text": "The embedding sizes of a word, five-minute tag f\u02dc , article-tag a, and time tag t are 128, 80, 64, and 64, respectively. We trained the models for 150 epochs with the mini-batch size of 100, using Adam (Kingma and Ba, 2015) optimizer with the initial learning rate 1 \u00d7 10 \u22124 , and saved the parameters every epoch, selecting the model with the highest BLEU score on the validation dataset.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "4.2" }, { "text": "We conduct two types of evaluation: one is BLEU (Papineni et al., 2002) to measure the matching degree between the market comments written by humans as references and the output comments generated by the models, and the other is a new evaluation metric that we created. The new metric uses the matching between the market price movement in the data and the movement expressions in the comments. Using (x , w , w ), which are the -th sample of the input data, the market comment written by humans, and the output comment generated by the models, we define the following https://hosted.datascope.reuters.com/ DataScope/ variables:", "cite_spans": [ { "start": 48, "end": 71, "text": "(Papineni et al., 2002)", "ref_id": "BIBREF23" } ], "ref_spans": [], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "4.3" }, { "text": "(w) = \uf8f1 \uf8f4 \uf8f4 \uf8f2 \uf8f4 \uf8f4 \uf8f3 (w \u2208 rise ) (w \u2208 fall ) ( \u210e ) (x) = \uf8f1 \uf8f4 \uf8f4 \uf8f2 \uf8f4 \uf8f4 \uf8f3 (Latest move > 0) (Latest move < 0) ( \u210e ) C = { | (w ) = (x )} C = { | (w ) = (x )} D = { | (w ) \u2260 (x )} D = { | (w ) \u2260 (x )}", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "4.3" }, { "text": "where rise is {\u7d9a\u4f38 (continuous rise), \u53cd\u767a (rebound), \u4e0a\u3052 (up / rise)}, fall is {\u7d9a\u843d (continuous fall), \u53cd\u843d (fall back), \u4e0b\u3052 (down/fall)}, move is the same as move shown in Section 2.1, defined as \u2212 , where is the closing price of the previous trading day. Using the above variables, we obtain the following metrics:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "4.3" }, { "text": "Concord = |C \u2229 C | |C | Concord = |C \u2229 C | |C | Diff = |D \u2229 D | |D | Diff = |D \u2229 D | |D | .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "4.3" }, { "text": "These metrics can be seen as a proxy for the model's ability to attend to an intermediate (not latest) step according to the movement in the data. To know the frequency of these concordances and differences in the data, we count them in the training and valid dataset, which we summarize in Table 2 .", "cite_spans": [], "ref_spans": [ { "start": 291, "end": 298, "text": "Table 2", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "4.3" }, { "text": "We evaluate the BLEU scores for both validation and test sets, while we performed the correspondence evaluation, which will be described below, only for the validation set. We train models with six different seeds for each setting and report the macro averages on them. According to the experimental results provided in Table 3 , our models mostly outperformed the baseline, especially when used with the copy mechanism. For the test set, our models always outperformed the baseline. In particular, = 2 with the copy mechanism achieved the highest score, 7.68 points improvement on the BLEU score compared to the baseline (see the bold font in Table 3) . Conversely, the result shows that just adding attention to the encoder (+Attention), keeping the decoder the same as the baseline, is not helpful. It shows that simply applying the attention mechanism does not enable the model to obtain the correspondence between data and text correctly, while the copy mechanism certainly helps to obtain the correspondence. In a comparison among the models using = 6, increasing the number of steps does not necessarily contribute to improving the BLEU score.", "cite_spans": [], "ref_spans": [ { "start": 320, "end": 327, "text": "Table 3", "ref_id": "TABREF3" }, { "start": 644, "end": 652, "text": "Table 3)", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "4.3" }, { "text": "Furthermore, Table 4 indicates that most of our models outperformed the baseline. However, in the same tendency as the BLEU score evaluation, this result further reveals that increasing the number of steps is not necessarily an important factor in improving the score. Figure 3 depicts distributions regarding the time gap between the comment delivery time and the event occurrence time. When we use = 5, we can cover 89.05% in training data and 94.07% in validation set of their time gap data (i.e., sum of an Irregular value and 0 to 5 of Regular values in Figure 3 ). Alternatively, using more x\u02dc would add more noise, therefore, they would be related to the transactions. According to Table 3, Table 4 and Figure 3 , using = 3 or = 4, covered around 80% are considered as the best choice in this dataset. Table 5 provides examples of the generated comments where our model generated the correct movement expression while the baseline generated an incorrect expression (see, the bold font in Table 5 ). Moreover, the method with copy mentioned the cor- Figure 3 : Data distribution regarding the time gap between the comment delivery time and the event occurrence time. First, we classified a human-written market comment in the dataset into two types, whether it was a regular comment (Regular) or not (Irregular) using the expressions specific to regular comments. In the case of a regular comment, it is classified by the time gap between that event occurrence time and the comment delivery time in five-minute increments, for example, 0 = from 0 min to 5 min gap, 1 = from 5 min to 10 min gap.", "cite_spans": [], "ref_spans": [ { "start": 13, "end": 20, "text": "Table 4", "ref_id": "TABREF6" }, { "start": 269, "end": 277, "text": "Figure 3", "ref_id": null }, { "start": 559, "end": 567, "text": "Figure 3", "ref_id": null }, { "start": 689, "end": 705, "text": "Table 3, Table 4", "ref_id": "TABREF3" }, { "start": 710, "end": 718, "text": "Figure 3", "ref_id": null }, { "start": 809, "end": 816, "text": "Table 5", "ref_id": null }, { "start": 995, "end": 1002, "text": "Table 5", "ref_id": null }, { "start": 1056, "end": 1064, "text": "Figure 3", "ref_id": null } ], "eq_spans": [], "section": "Results", "sec_num": "4.4" }, { "text": "rect numerical value, while the baseline generated an incorrect numerical value. However, there is another problem with the fluency. For instance, at the expression of \"in the higher yen range\", the numerical value should be a round number, e.g., round to the nearest 100, 1,000, or 10,000. However, the model with the copy mechanism generates a specific number (see, the underline font in Table 5 )", "cite_spans": [], "ref_spans": [ { "start": 390, "end": 397, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "Results", "sec_num": "4.4" }, { "text": "The task of generating text describing input data, which is called data-to-text, has been worked on various domains, for instance weather forecasts (Belz, 2007; Angeli et al., 2010) , healthcare (Portet et al., 2009; Banaee et al., 2013) , and sports (Liang et al., 2009) . Traditionally, data-to-text is divided into two sub-problems (Kukich, 1983; Goldberg et al., 1994) : content selection, which is about \"what to say\", and surface realization, which is about \"how to say\". Moreover, Reiter and Dale (1997) divides three modules, adding micro planning between the above sub-problems. In the early stage of this task, surface realization is often realized using templates (van Deemter et al., 2005) or statistically learned models with hand-crafted features (Belz, 2008; Konstas and Lapata, 2012) .", "cite_spans": [ { "start": 148, "end": 160, "text": "(Belz, 2007;", "ref_id": "BIBREF2" }, { "start": 161, "end": 181, "text": "Angeli et al., 2010)", "ref_id": "BIBREF0" }, { "start": 195, "end": 216, "text": "(Portet et al., 2009;", "ref_id": "BIBREF24" }, { "start": 217, "end": 237, "text": "Banaee et al., 2013)", "ref_id": "BIBREF1" }, { "start": 251, "end": 271, "text": "(Liang et al., 2009)", "ref_id": "BIBREF17" }, { "start": 335, "end": 349, "text": "(Kukich, 1983;", "ref_id": "BIBREF14" }, { "start": 350, "end": 372, "text": "Goldberg et al., 1994)", "ref_id": "BIBREF7" }, { "start": 488, "end": 510, "text": "Reiter and Dale (1997)", "ref_id": "BIBREF27" }, { "start": 675, "end": 701, "text": "(van Deemter et al., 2005)", "ref_id": "BIBREF6" }, { "start": 761, "end": 773, "text": "(Belz, 2008;", "ref_id": "BIBREF3" }, { "start": 774, "end": 799, "text": "Konstas and Lapata, 2012)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "In recent times, various industries such as finance, pharmaceuticals, and telecommunications Nikkei opens with a continual rise slightly. \u00d7 The price is in the higher 17024 yen range. \u65e5\u7d4c\u5e73\u5747\u3001\u5c0f\u5e45\u7d9a\u4f38\u3067\u59cb\u307e\u308b 17024\u5186\u53f0\u5f8c\u534a Table 5 : Generated comment. The latest stock price movement of comment delivery time is -48.91 yen, so that (x) is . Conc (concordance) row shows whether the movement expression in the comment matches the movement with the latest stock price ( ) or not (\u00d7).", "cite_spans": [], "ref_spans": [ { "start": 208, "end": 215, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "have been increased providing opportunities to treat various types of large-scale data, so that they are interested in automatically learning a correspondence relationship from data to text and generating a description of this relationship. Therefore, recent works have focused on generating text from data with neural networks, that can solve the above sub-tasks in one through. Especially the models, which utilize an encoder-decoder model (Sutskever et al., 2014 ) have proven to be useful (Mei et al., 2016; Lebret et al., 2016) . While text generation by neural network can describe the text fluently, they do not describe the exact entities or numbers. Therefore, a copy mechanism (Vinyals et al., 2015; Gu et al., 2016) , which provides a way to directly copy words from the input, has been utilized. By these neural networks, the works such as conditional language generation based on tables (Yang et al., 2017) , short biographies generation from Wikipedia tables (Lebret et al., 2016; Chisholm et al., 2017; and sports scoreboards (Wiseman et al., 2017; Li and Wan, 2018; Puduppully et al., 2019) are well performed. Contrastingly, they can only generate the superficial contents that appear in their input table, and cannot generate contents that require arithmetic operations.", "cite_spans": [ { "start": 442, "end": 465, "text": "(Sutskever et al., 2014", "ref_id": "BIBREF30" }, { "start": 493, "end": 511, "text": "(Mei et al., 2016;", "ref_id": "BIBREF19" }, { "start": 512, "end": 532, "text": "Lebret et al., 2016)", "ref_id": "BIBREF15" }, { "start": 687, "end": 709, "text": "(Vinyals et al., 2015;", "ref_id": "BIBREF33" }, { "start": 710, "end": 726, "text": "Gu et al., 2016)", "ref_id": "BIBREF8" }, { "start": 900, "end": 919, "text": "(Yang et al., 2017)", "ref_id": "BIBREF36" }, { "start": 973, "end": 994, "text": "(Lebret et al., 2016;", "ref_id": "BIBREF15" }, { "start": 995, "end": 1017, "text": "Chisholm et al., 2017;", "ref_id": "BIBREF5" }, { "start": 1041, "end": 1063, "text": "(Wiseman et al., 2017;", "ref_id": "BIBREF35" }, { "start": 1064, "end": 1081, "text": "Li and Wan, 2018;", "ref_id": "BIBREF16" }, { "start": 1082, "end": 1106, "text": "Puduppully et al., 2019)", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "However, Joulin and Mikolov (2015) and Neelakantan et al. (2016) indicate that current neural models have difficulties in learning arithmetic operations such as addition and comparisons by neural program inductions. Thus, there have been some methods to prepare the numerical values with arithmetic operations in advance. Murakami et al. (2017) post-process the price by extending the copy mechanism and replacing numerical values with defined arithmetic operations after generation. Nie et al. (2018) utilizes information from pre-computed operations on raw data to consider incorporating the facts that can be inferred from the input data to guide the generation process. Our model prepares numerical values with defined arithmetic operations as Murakami et al. (2017) for copy and that copy target is guided by encoded input.", "cite_spans": [ { "start": 9, "end": 34, "text": "Joulin and Mikolov (2015)", "ref_id": "BIBREF11" }, { "start": 39, "end": 64, "text": "Neelakantan et al. (2016)", "ref_id": "BIBREF21" }, { "start": 322, "end": 344, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" }, { "start": 484, "end": 501, "text": "Nie et al. (2018)", "ref_id": "BIBREF22" }, { "start": 748, "end": 770, "text": "Murakami et al. (2017)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "In this paper, we have proposed an encoder-decoder model with multi-timestep data and a copy mechanism for generating the market comment from data with the noisy alignments. Both BLEU scores and our proposal evaluation showed the accuracy of sentence generation with time-series data has been improved by our proposed method, especially utilizing a copy mechanism.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "6" }, { "text": "t is the time embedding defined in Section 2.1. Comparing to Equation (4), we add q to the input to the LSTM. Each element of q is (\u02dc , ). We add this vector to properly propagate the information about an applied arithmetic operation, which may not be kept directly in s .4 Experiments4.1 DatasetsWe used a five-minute chart of the Nikkei 225 from March 2013 to October 2016 as numerical", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This paper is based on results obtained from projects commissioned by the New Energy and Industrial Technology Development Organization (NEDO) JPNP20006 and JPNP15009, and JSPS KAKENHI Grant Number JP20H04217.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "A simple domain-independent probabilistic approach to generation", "authors": [ { "first": "Gabor", "middle": [], "last": "Angeli", "suffix": "" }, { "first": "Percy", "middle": [], "last": "Liang", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Klein", "suffix": "" } ], "year": 2010, "venue": "Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "502--512", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gabor Angeli, Percy Liang, and Dan Klein. 2010. A simple domain-independent probabilistic approach to generation. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Process- ing, pages 502-512, Cambridge, MA. Association for Computational Linguistics.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Towards NLG for physiological data monitoring with body area networks", "authors": [ { "first": "", "middle": [], "last": "Hadi Banaee", "suffix": "" }, { "first": "Amy", "middle": [], "last": "Mobyen Uddin Ahmed", "suffix": "" }, { "first": "", "middle": [], "last": "Loutfi", "suffix": "" } ], "year": 2013, "venue": "Proceedings of the 14th European Workshop on Natural Language Generation", "volume": "", "issue": "", "pages": "193--197", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hadi Banaee, Mobyen Uddin Ahmed, and Amy Loutfi. 2013. Towards NLG for physiological data monitor- ing with body area networks. In Proceedings of the 14th European Workshop on Natural Language Gen- eration, pages 193-197, Sofia, Bulgaria. Association for Computational Linguistics.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Probabilistic generation of weather forecast texts", "authors": [ { "first": "Anja", "middle": [], "last": "Belz", "suffix": "" } ], "year": 2007, "venue": "Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference", "volume": "", "issue": "", "pages": "164--171", "other_ids": {}, "num": null, "urls": [], "raw_text": "Anja Belz. 2007. Probabilistic generation of weather forecast texts. In Human Language Technologies 2007: The Conference of the North American Chap- ter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 164-171, Rochester, New York. Association for Computational Linguistics.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Automatic generation of weather forecast texts using comprehensive probabilistic generation-space models", "authors": [ { "first": "Anja", "middle": [], "last": "Belz", "suffix": "" } ], "year": 2008, "venue": "Natural Language Engineering", "volume": "14", "issue": "4", "pages": "431--455", "other_ids": { "DOI": [ "10.1017/S1351324907004664" ] }, "num": null, "urls": [], "raw_text": "Anja Belz. 2008. Automatic generation of weather forecast texts using comprehensive probabilistic generation-space models. Natural Language Engi- neering, 14(4):431-455.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Learning to sportscast: A test of grounded language acquisition", "authors": [ { "first": "L", "middle": [], "last": "David", "suffix": "" }, { "first": "Raymond", "middle": [ "J" ], "last": "Chen", "suffix": "" }, { "first": "", "middle": [], "last": "Mooney", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 25th International Conference on Machine Learning, ICML '08", "volume": "", "issue": "", "pages": "128--135", "other_ids": { "DOI": [ "10.1145/1390156.1390173" ] }, "num": null, "urls": [], "raw_text": "David L. Chen and Raymond J. Mooney. 2008. Learn- ing to sportscast: A test of grounded language ac- quisition. In Proceedings of the 25th International Conference on Machine Learning, ICML '08, page 128-135, New York, NY, USA. Association for Com- puting Machinery.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Learning to generate one-sentence biographies from Wikidata", "authors": [ { "first": "Andrew", "middle": [], "last": "Chisholm", "suffix": "" }, { "first": "Will", "middle": [], "last": "Radford", "suffix": "" }, { "first": "Ben", "middle": [], "last": "Hachey", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "633--642", "other_ids": {}, "num": null, "urls": [], "raw_text": "Andrew Chisholm, Will Radford, and Ben Hachey. 2017. Learning to generate one-sentence biogra- phies from Wikidata. In Proceedings of the 15th Conference of the European Chapter of the Associa- tion for Computational Linguistics: Volume 1, Long Papers, pages 633-642, Valencia, Spain. Association for Computational Linguistics.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Real versus template-based natural language generation: A false opposition?", "authors": [ { "first": "Mariet", "middle": [], "last": "Kees Van Deemter", "suffix": "" }, { "first": "Emiel", "middle": [], "last": "Theune", "suffix": "" }, { "first": "", "middle": [], "last": "Krahmer", "suffix": "" } ], "year": 2005, "venue": "Computational Linguistics", "volume": "31", "issue": "", "pages": "15--24", "other_ids": { "DOI": [ "10.1162/0891201053630291" ] }, "num": null, "urls": [], "raw_text": "Kees van Deemter, Mariet Theune, and Emiel Krahmer. 2005. Real versus template-based natural language generation: A false opposition? Computational Lin- guistics, 31:15-24.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Using natural-language processing to produce weather forecasts", "authors": [ { "first": "Eli", "middle": [], "last": "Goldberg", "suffix": "" }, { "first": "Norbert", "middle": [], "last": "Driedger", "suffix": "" }, { "first": "Richard", "middle": [ "I" ], "last": "Kittredge", "suffix": "" } ], "year": 1994, "venue": "IEEE Expert", "volume": "9", "issue": "2", "pages": "45--53", "other_ids": {}, "num": null, "urls": [], "raw_text": "Eli Goldberg, Norbert Driedger, and Richard I Kittredge. 1994. Using natural-language processing to produce weather forecasts. IEEE Expert, 9(2):45-53.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Incorporating copying mechanism in sequence-to-sequence learning", "authors": [ { "first": "Jiatao", "middle": [], "last": "Gu", "suffix": "" }, { "first": "Zhengdong", "middle": [], "last": "Lu", "suffix": "" }, { "first": "Hang", "middle": [], "last": "Li", "suffix": "" }, { "first": "O", "middle": [ "K" ], "last": "Victor", "suffix": "" }, { "first": "", "middle": [], "last": "Li", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "1631--1640", "other_ids": { "DOI": [ "10.18653/v1/P16-1154" ] }, "num": null, "urls": [], "raw_text": "Jiatao Gu, Zhengdong Lu, Hang Li, and Victor O.K. Li. 2016. Incorporating copying mechanism in sequence-to-sequence learning. In Proceedings of the 54th Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), pages 1631-1640, Berlin, Germany. Association for Computational Linguistics.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Long short-term memory", "authors": [ { "first": "Sepp", "middle": [], "last": "Hochreiter", "suffix": "" }, { "first": "J\u00fcrgen", "middle": [], "last": "Schmidhuber", "suffix": "" } ], "year": 1997, "venue": "Neural Computation", "volume": "9", "issue": "8", "pages": "1735--1780", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9(8):1735-1780.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Learning to select, track, and generate for data-to-text", "authors": [ { "first": "Hayate", "middle": [], "last": "Iso", "suffix": "" }, { "first": "Yui", "middle": [], "last": "Uehara", "suffix": "" }, { "first": "Tatsuya", "middle": [], "last": "Ishigaki", "suffix": "" }, { "first": "Hiroshi", "middle": [], "last": "Noji", "suffix": "" }, { "first": "Eiji", "middle": [], "last": "Aramaki", "suffix": "" }, { "first": "Ichiro", "middle": [], "last": "Kobayashi", "suffix": "" }, { "first": "Yusuke", "middle": [], "last": "Miyao", "suffix": "" }, { "first": "Naoaki", "middle": [], "last": "Okazaki", "suffix": "" }, { "first": "Hiroya", "middle": [], "last": "Takamura", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "2102--2113", "other_ids": { "DOI": [ "10.18653/v1/P19-1202" ] }, "num": null, "urls": [], "raw_text": "Hayate Iso, Yui Uehara, Tatsuya Ishigaki, Hiroshi Noji, Eiji Aramaki, Ichiro Kobayashi, Yusuke Miyao, Naoaki Okazaki, and Hiroya Takamura. 2019. Learn- ing to select, track, and generate for data-to-text. In Proceedings of the 57th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 2102- 2113, Florence, Italy. Association for Computational Linguistics.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Inferring algorithmic patterns with stack-augmented recurrent nets", "authors": [ { "first": "Armand", "middle": [], "last": "Joulin", "suffix": "" }, { "first": "Tomas", "middle": [], "last": "Mikolov", "suffix": "" } ], "year": 2015, "venue": "Advances in Neural Information Processing Systems", "volume": "28", "issue": "", "pages": "190--198", "other_ids": {}, "num": null, "urls": [], "raw_text": "Armand Joulin and Tomas Mikolov. 2015. Inferring al- gorithmic patterns with stack-augmented recurrent nets. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 190-198. Curran Associates, Inc.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Adam: A method for stochastic optimization", "authors": [ { "first": "P", "middle": [], "last": "Diederik", "suffix": "" }, { "first": "Jimmy", "middle": [], "last": "Kingma", "suffix": "" }, { "first": "", "middle": [], "last": "Ba", "suffix": "" } ], "year": 2015, "venue": "3rd International Conference on Learning Representations", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd Inter- national Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Unsupervised concept-to-text generation with hypergraphs", "authors": [ { "first": "Ioannis", "middle": [], "last": "Konstas", "suffix": "" }, { "first": "Mirella", "middle": [], "last": "Lapata", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "752--761", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ioannis Konstas and Mirella Lapata. 2012. Unsuper- vised concept-to-text generation with hypergraphs. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, pages 752-761, Montr\u00e9al, Canada. Association for Computational Linguistics.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Design of a knowledge-based report generator", "authors": [ { "first": "Karen", "middle": [], "last": "Kukich", "suffix": "" } ], "year": 1983, "venue": "21st Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "145--150", "other_ids": { "DOI": [ "10.3115/981311.981340" ] }, "num": null, "urls": [], "raw_text": "Karen Kukich. 1983. Design of a knowledge-based report generator. In 21st Annual Meeting of the As- sociation for Computational Linguistics, pages 145- 150, Cambridge, Massachusetts, USA. Association for Computational Linguistics.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Neural text generation from structured data with application to the biography domain", "authors": [ { "first": "R\u00e9mi", "middle": [], "last": "Lebret", "suffix": "" }, { "first": "David", "middle": [], "last": "Grangier", "suffix": "" }, { "first": "Michael", "middle": [], "last": "Auli", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "1203--1213", "other_ids": { "DOI": [ "10.18653/v1/D16-1128" ] }, "num": null, "urls": [], "raw_text": "R\u00e9mi Lebret, David Grangier, and Michael Auli. 2016. Neural text generation from structured data with ap- plication to the biography domain. In Proceedings of the 2016 Conference on Empirical Methods in Natu- ral Language Processing, pages 1203-1213, Austin, Texas. Association for Computational Linguistics.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Point precisely: Towards ensuring the precision of data in generated texts using delayed copy mechanism", "authors": [ { "first": "Liunian", "middle": [], "last": "Li", "suffix": "" }, { "first": "Xiaojun", "middle": [], "last": "Wan", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 27th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "1044--1055", "other_ids": {}, "num": null, "urls": [], "raw_text": "Liunian Li and Xiaojun Wan. 2018. Point precisely: Towards ensuring the precision of data in generated texts using delayed copy mechanism. In Proceed- ings of the 27th International Conference on Com- putational Linguistics, pages 1044-1055, Santa Fe, New Mexico, USA. Association for Computational Linguistics.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Learning semantic correspondences with less supervision", "authors": [ { "first": "Percy", "middle": [], "last": "Liang", "suffix": "" }, { "first": "Michael", "middle": [], "last": "Jordan", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Klein", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP", "volume": "", "issue": "", "pages": "91--99", "other_ids": {}, "num": null, "urls": [], "raw_text": "Percy Liang, Michael Jordan, and Dan Klein. 2009. Learning semantic correspondences with less super- vision. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th In- ternational Joint Conference on Natural Language Processing of the AFNLP, pages 91-99, Suntec, Sin- gapore. Association for Computational Linguistics.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Table-to-text generation by structure-aware seq2seq learning", "authors": [ { "first": "Tianyu", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Kexiang", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Lei", "middle": [], "last": "Sha", "suffix": "" }, { "first": "Baobao", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Zhifang", "middle": [], "last": "Sui", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tianyu Liu, Kexiang Wang, Lei Sha, Baobao Chang, and Zhifang Sui. 2018. Table-to-text generation by structure-aware seq2seq learning. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "What to talk about and how? selective generation using LSTMs with coarse-to-fine alignment", "authors": [ { "first": "Hongyuan", "middle": [], "last": "Mei", "suffix": "" }, { "first": "Mohit", "middle": [], "last": "Bansal", "suffix": "" }, { "first": "Matthew", "middle": [ "R" ], "last": "Walter", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "720--730", "other_ids": { "DOI": [ "10.18653/v1/N16-1086" ] }, "num": null, "urls": [], "raw_text": "Hongyuan Mei, Mohit Bansal, and Matthew R. Walter. 2016. What to talk about and how? selective gen- eration using LSTMs with coarse-to-fine alignment. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, pages 720-730, San Diego, California. Association for Computational Linguistics.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Learning to generate market comments from stock prices", "authors": [ { "first": "Soichiro", "middle": [], "last": "Murakami", "suffix": "" }, { "first": "Akihiko", "middle": [], "last": "Watanabe", "suffix": "" }, { "first": "Akira", "middle": [], "last": "Miyazawa", "suffix": "" }, { "first": "Keiichi", "middle": [], "last": "Goshima", "suffix": "" }, { "first": "Toshihiko", "middle": [], "last": "Yanase", "suffix": "" }, { "first": "Hiroya", "middle": [], "last": "Takamura", "suffix": "" }, { "first": "Yusuke", "middle": [], "last": "Miyao", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "1374--1384", "other_ids": { "DOI": [ "10.18653/v1/P17-1126" ] }, "num": null, "urls": [], "raw_text": "Soichiro Murakami, Akihiko Watanabe, Akira Miyazawa, Keiichi Goshima, Toshihiko Yanase, Hi- roya Takamura, and Yusuke Miyao. 2017. Learning to generate market comments from stock prices. In Proceedings of the 55th Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 1374-1384, Vancouver, Canada. Association for Computational Linguistics.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Neural programmer: Inducing latent programs with gradient descent", "authors": [ { "first": "Arvind", "middle": [], "last": "Neelakantan", "suffix": "" }, { "first": "V", "middle": [], "last": "Quoc", "suffix": "" }, { "first": "Ilya", "middle": [], "last": "Le", "suffix": "" }, { "first": "", "middle": [], "last": "Sutskever", "suffix": "" } ], "year": 2016, "venue": "4th International Conference on Learning Representations", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Arvind Neelakantan, Quoc V. Le, and Ilya Sutskever. 2016. Neural programmer: Inducing latent pro- grams with gradient descent. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Con- ference Track Proceedings.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Operation-guided neural networks for high fidelity data-to-text generation", "authors": [ { "first": "Feng", "middle": [], "last": "Nie", "suffix": "" }, { "first": "Jinpeng", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Jin-Ge", "middle": [], "last": "Yao", "suffix": "" }, { "first": "Rong", "middle": [], "last": "Pan", "suffix": "" }, { "first": "Chin-Yew", "middle": [], "last": "Lin", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "3879--3889", "other_ids": { "DOI": [ "10.18653/v1/D18-1422" ] }, "num": null, "urls": [], "raw_text": "Feng Nie, Jinpeng Wang, Jin-Ge Yao, Rong Pan, and Chin-Yew Lin. 2018. Operation-guided neu- ral networks for high fidelity data-to-text genera- tion. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3879-3889, Brussels, Belgium. Association for Computational Linguistics.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Bleu: a method for automatic evaluation of machine translation", "authors": [ { "first": "Kishore", "middle": [], "last": "Papineni", "suffix": "" }, { "first": "Salim", "middle": [], "last": "Roukos", "suffix": "" }, { "first": "Todd", "middle": [], "last": "Ward", "suffix": "" }, { "first": "Wei-Jing", "middle": [], "last": "Zhu", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "311--318", "other_ids": { "DOI": [ "10.3115/1073083.1073135" ] }, "num": null, "urls": [], "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic eval- uation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Com- putational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Automatic generation of textual summaries from neonatal intensive care data", "authors": [ { "first": "Fran\u00e7ois", "middle": [], "last": "Portet", "suffix": "" }, { "first": "Ehud", "middle": [], "last": "Reiter", "suffix": "" }, { "first": "Albert", "middle": [], "last": "Gatt", "suffix": "" }, { "first": "Jim", "middle": [], "last": "Hunter", "suffix": "" }, { "first": "Somayajulu", "middle": [], "last": "Sripada", "suffix": "" }, { "first": "Yvonne", "middle": [], "last": "Freer", "suffix": "" }, { "first": "Cindy", "middle": [], "last": "Sykes", "suffix": "" } ], "year": 2009, "venue": "Artificial Intelligence", "volume": "173", "issue": "7", "pages": "789--816", "other_ids": { "DOI": [ "10.1016/j.artint.2008.12.002" ] }, "num": null, "urls": [], "raw_text": "Fran\u00e7ois Portet, Ehud Reiter, Albert Gatt, Jim Hunter, Somayajulu Sripada, Yvonne Freer, and Cindy Sykes. 2009. Automatic generation of textual summaries from neonatal intensive care data. Artificial Intelli- gence, 173(7):789 -816.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Data-to-text generation with entity modeling", "authors": [ { "first": "Ratish", "middle": [], "last": "Puduppully", "suffix": "" }, { "first": "Li", "middle": [], "last": "Dong", "suffix": "" }, { "first": "Mirella", "middle": [], "last": "Lapata", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "2023--2035", "other_ids": { "DOI": [ "10.18653/v1/P19-1195" ] }, "num": null, "urls": [], "raw_text": "Ratish Puduppully, Li Dong, and Mirella Lapata. 2019. Data-to-text generation with entity modeling. In Pro- ceedings of the 57th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 2023- 2035, Florence, Italy. Association for Computational Linguistics.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "E2E NLG challenge: Neural models vs. templates", "authors": [ { "first": "Yevgeniy", "middle": [], "last": "Puzikov", "suffix": "" }, { "first": "Iryna", "middle": [], "last": "Gurevych", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 11th International Conference on Natural Language Generation", "volume": "", "issue": "", "pages": "463--471", "other_ids": { "DOI": [ "10.18653/v1/W18-6557" ] }, "num": null, "urls": [], "raw_text": "Yevgeniy Puzikov and Iryna Gurevych. 2018. E2E NLG challenge: Neural models vs. templates. In Proceed- ings of the 11th International Conference on Natural Language Generation, pages 463-471, Tilburg Uni- versity, The Netherlands. Association for Computa- tional Linguistics.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Building applied natural language generation systems", "authors": [ { "first": "Ehud", "middle": [], "last": "Reiter", "suffix": "" }, { "first": "Robert", "middle": [], "last": "Dale", "suffix": "" } ], "year": 1997, "venue": "Natural Language Engineering", "volume": "3", "issue": "1", "pages": "57--87", "other_ids": { "DOI": [ "10.1017/S1351324997001502" ] }, "num": null, "urls": [], "raw_text": "Ehud Reiter and Robert Dale. 1997. Building applied natural language generation systems. Natural Lan- guage Engineering, 3(1):57-87.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "Get to the point: Summarization with pointergenerator networks", "authors": [ { "first": "Abigail", "middle": [], "last": "See", "suffix": "" }, { "first": "J", "middle": [], "last": "Peter", "suffix": "" }, { "first": "Christopher", "middle": [ "D" ], "last": "Liu", "suffix": "" }, { "first": "", "middle": [], "last": "Manning", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "1073--1083", "other_ids": { "DOI": [ "10.18653/v1/P17-1099" ] }, "num": null, "urls": [], "raw_text": "Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer- generator networks. In Proceedings of the 55th An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073- 1083, Vancouver, Canada. Association for Computa- tional Linguistics.", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "Orderplanning neural text generation from structured data", "authors": [ { "first": "Lei", "middle": [], "last": "Sha", "suffix": "" }, { "first": "Lili", "middle": [], "last": "Mou", "suffix": "" }, { "first": "Tianyu", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Pascal", "middle": [], "last": "Poupart", "suffix": "" }, { "first": "Sujian", "middle": [], "last": "Li", "suffix": "" }, { "first": "Baobao", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Zhifang", "middle": [], "last": "Sui", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lei Sha, Lili Mou, Tianyu Liu, Pascal Poupart, Sujian Li, Baobao Chang, and Zhifang Sui. 2018. Order- planning neural text generation from structured data. In Proceedings of the Thirty-Second AAAI Confer- ence on Artificial Intelligence.", "links": null }, "BIBREF30": { "ref_id": "b30", "title": "Sequence to sequence learning with neural networks", "authors": [ { "first": "Ilya", "middle": [], "last": "Sutskever", "suffix": "" }, { "first": "Oriol", "middle": [], "last": "Vinyals", "suffix": "" }, { "first": "Quoc V", "middle": [], "last": "Le", "suffix": "" } ], "year": 2014, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Z. Ghahramani, M. Welling, C. Cortes, N. D.", "links": null }, "BIBREF31": { "ref_id": "b31", "title": "Advances in Neural Information Processing Systems", "authors": [ { "first": "K", "middle": [ "Q" ], "last": "Lawrence", "suffix": "" }, { "first": "", "middle": [], "last": "Weinberger", "suffix": "" } ], "year": null, "venue": "", "volume": "27", "issue": "", "pages": "3104--3112", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 3104-3112. Curran Associates, Inc.", "links": null }, "BIBREF32": { "ref_id": "b32", "title": "Generating live soccermatch commentary from play data", "authors": [ { "first": "Yasufumi", "middle": [], "last": "Taniguchi", "suffix": "" }, { "first": "Yukun", "middle": [], "last": "Feng", "suffix": "" }, { "first": "Hiroya", "middle": [], "last": "Takamura", "suffix": "" }, { "first": "Manabu", "middle": [], "last": "Okumura", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the AAAI Conference on Artificial Intelligence", "volume": "33", "issue": "", "pages": "7096--7103", "other_ids": { "DOI": [ "10.1609/aaai.v33i01.33017096" ] }, "num": null, "urls": [], "raw_text": "Yasufumi Taniguchi, Yukun Feng, Hiroya Takamura, and Manabu Okumura. 2019. Generating live soccer- match commentary from play data. In Proceedings of the AAAI Conference on Artificial Intelligence, vol- ume 33, pages 7096-7103.", "links": null }, "BIBREF33": { "ref_id": "b33", "title": "Pointer networks", "authors": [ { "first": "Oriol", "middle": [], "last": "Vinyals", "suffix": "" }, { "first": "Meire", "middle": [], "last": "Fortunato", "suffix": "" }, { "first": "Navdeep", "middle": [], "last": "Jaitly", "suffix": "" } ], "year": 2015, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In C. Cortes, N. D.", "links": null }, "BIBREF34": { "ref_id": "b34", "title": "Advances in Neural Information Processing Systems", "authors": [ { "first": "D", "middle": [ "D" ], "last": "Lawrence", "suffix": "" }, { "first": "M", "middle": [], "last": "Lee", "suffix": "" }, { "first": "R", "middle": [], "last": "Sugiyama", "suffix": "" }, { "first": "", "middle": [], "last": "Garnett", "suffix": "" } ], "year": null, "venue": "", "volume": "28", "issue": "", "pages": "2692--2700", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 2692-2700. Curran Associates, Inc.", "links": null }, "BIBREF35": { "ref_id": "b35", "title": "Challenges in data-to-document generation", "authors": [ { "first": "Sam", "middle": [], "last": "Wiseman", "suffix": "" }, { "first": "Stuart", "middle": [], "last": "Shieber", "suffix": "" }, { "first": "Alexander", "middle": [], "last": "Rush", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "2253--2263", "other_ids": { "DOI": [ "10.18653/v1/D17-1239" ] }, "num": null, "urls": [], "raw_text": "Sam Wiseman, Stuart Shieber, and Alexander Rush. 2017. Challenges in data-to-document generation. In Proceedings of the 2017 Conference on Empiri- cal Methods in Natural Language Processing, pages 2253-2263, Copenhagen, Denmark. Association for Computational Linguistics.", "links": null }, "BIBREF36": { "ref_id": "b36", "title": "Reference-aware language models", "authors": [ { "first": "Zichao", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Phil", "middle": [], "last": "Blunsom", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Dyer", "suffix": "" }, { "first": "Wang", "middle": [], "last": "Ling", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "1850--1859", "other_ids": { "DOI": [ "10.18653/v1/D17-1197" ] }, "num": null, "urls": [], "raw_text": "Zichao Yang, Phil Blunsom, Chris Dyer, and Wang Ling. 2017. Reference-aware language models. In Pro- ceedings of the 2017 Conference on Empirical Meth- ods in Natural Language Processing, pages 1850- 1859, Copenhagen, Denmark. Association for Com- putational Linguistics.", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "num": null, "text": "Nikkei stock average starts with rebound, gains 86 yen, a positive response in high price of crude oil. The stock price of Toyota increased. (III) 09:19 continuous fall dropped 2 yen Tosho begins with rebound. A positive response in high price of crude oil, Fanuc share prices dropped significantly.", "uris": null }, "FIGREF1": { "type_str": "figure", "num": null, "text": "Nikkei 225 and market comments about an event at 9:00 am, 29th of January model ofMurakami et al.", "uris": null }, "TABREF0": { "num": null, "content": "", "html": null, "text": "Generalization tags and corresponding arithmetic operations. is defined as the latest price, and is defined as the difference between and the closing price of the previous trading day.", "type_str": "table" }, "TABREF2": { "num": null, "content": "
: Statistics of the data
time-series data, which were collected from Thom-
son Reuters DataScope Select . As market com-
ments, we used 18,489 headlines of Nikkei Quick
News (NQN) that describe the Nikkei 225 behavior.
They were provided by Nikkei, Inc. and written
in Japanese. We divided them into three parts on
the basis of the period of publication: 15,035 for
training (December 2010-October 2015), 1,759 for
validation (October 2015-April 2016) and 1,695
for testing (April-October 2016).
", "html": null, "text": "", "type_str": "table" }, "TABREF3": { "num": null, "content": "
Valid Test Valid Test Valid Test
Baseline 21.37 21.30----
Multi-timesteps+Attention+Copy
= 021.63 22.66 21.33 22.38 28.16 28.68
121.59 22.74 21.43 22.25 27.90 28.54
221.64 23.03 21.20 22.47 28.31 28.98
321.82 23.11 21.02 21.94 27.29 27.75
421.68 22.92 20.59 21.71 28.13 28.93
521.73 22.89 20.71 21.74 27.29 27.78
621.73 22.66 20.51 21.62 26.68 27.25
", "html": null, "text": "summarizes the BLEU scores on the validation and test sets, andTable 4presents the correspondence evaluation on the validation set.", "type_str": "table" }, "TABREF4": { "num": null, "content": "
: BLEU (%)
", "html": null, "text": "", "type_str": "table" }, "TABREF5": { "num": null, "content": "
ConcordDiffConcordDiffConcordDiff
PRPRPRPRPRPR
Baseline 98.68 98.78 27.57 24.72--------
Multi-timesteps+Attention+Copy
= 098.61
", "html": null, "text": "98.94 41.00 33.93 98.90 99.08 51.35 45.28 98.85 99.25 53.90 42.61 1 98.77 99.02 46.03 40.52 98.76 99.28 56.83 41.91 98.81 99.02 45.86 39.80 2 98.79 98.47 40.66 46.13 98.99 98.78 45.00 49.74 98.71 99.06 47.08 38.56 3 98.82 98.61 43.12 46.20 98.93 98.87 44.65 46.25 98.73 98.95 45.38 38.79 4 98.88 98.86 46.58 47.05 98.80 98.44 30.60 36.29 98.68 98.68 36.46 36.06 5 98.88 98.91 44.13 43.47 98.69 98.14 27.88 34.92 98.57 98.57 34.04 31.54 6 98.80 98.70 41.88 43.43 98.70 97.11 25.54 38.74 98.46 98.81 32.35 26.17", "type_str": "table" }, "TABREF6": { "num": null, "content": "
ModelGenerated CommentConc
GoldNikkei opens with a continual rise.
The price is 17024 yen, 9 yen higher.\u00d7
\u65e5\u7d4c\u5e73\u5747\u3001\u7d9a\u4f38\u3067\u59cb\u307e\u308b
9\u5186\u9ad8\u306e17024\u5186
BaselineNikkei opens with a fall back.
The price is 16900 yen level.
\u65e5\u7d4c\u5e73\u5747\u3001\u53cd\u843d\u3067\u59cb\u307e\u308b 16900\u5186\u53f0
Multi-timesteps
= 3Nikkei opens with a continual rise slightly.\u00d7
\u65e5\u7d4c\u5e73\u5747\u3001\u5c0f\u5e45\u7d9a\u4f38\u3067\u59cb\u307e\u308b
+Attention Nikkei opens with a continual rise slightly.\u00d7
\u65e5\u7d4c\u5e73\u5747\u3001\u5c0f\u5e45\u7d9a\u4f38\u3067\u59cb\u307e\u308b
+Copy
", "html": null, "text": "The correspondence evaluation on validation set (%, P=precision, R=recall)", "type_str": "table" } } } }