Unpredictable Attributes in Market Comment Generation

Unpredictable Attributes in Market Comment Generation Yumi Hamazono1,2 Tatsuya Ishigaki2 Yusuke Miyao3,2 Hiroya Takamura2 Ichiro Kobayashi1,2 1Ochanomizu University 2National Institute of Advanced Industrial Science and Technology 3The University of Tokyo {hamazono.yumi, koba}@is.ocha.ac.jp {ishigaki.tatsuya, takamura.hiroya}@aist.go.jp yusuke@is.s.u-tokyo.ac.jp Abstract There are two types of datasets for data-to- text: one uses raw data obtained in the real world, and the other is constructed artificially for a controlled task. A straightforwardly out- put text is generated from its paired input data for a manually constructed dataset because the dataset is well constructed without any excess or deficiencies. However, it may not be possi- ble to generate a correct output text from the input data for a dataset constructed with real- world data and text. In such cases, we have to provide additional data, for example, data or text attribute labels, in order to generate the ex- pected output text from the paired input. This paper discusses the importance of additional input labels in data-to-text for real-world data. The content and style of a market comment change depending on its medium, the market situation, and the time of day. However, as the stock price, which is the input data, does not contain any such aforementioned informa- tion, it cannot generate comments appropri- ately from the data alone. Therefore, we anal- yse the dataset and provide additional labels which are unpredictable with input data for the appropriate parts in the model. Thus, the accu- racy of sentence generation is greatly improved compared to the case without the labels.The re- sult suggests unpredictable attributes should be given as a part of the input in the training of the text generating model. 1 Introduction Data-to-text generation is the task of generating tex- tual descriptions from numerical time series or struc- tured data, including sports data (Wiseman et al., 2017) and market data (Murakami et al., 2017). The majority of the recently proposed models for this task are neural network-based language generators conditioned to the structured or numerical data. We usually rely on large training datasets for training neural networks. However, many datasets used for data-to-text gen- eration are artificial because they are constructed as benchmarks for language generation research. Such artificial datasets are usually constructed by crowd workers working under constraints; for example, the target texts should only mention the information in the input data. For instance, the target text in the E2E dataset (Novikova et al., 2017) is constrained to mention all parts of input Meaning Representations (MR) and ToTTo (Parikh et al., 2020) is the dataset for the task of generating a one-sentence description that mentions the highlighted cells of a Wikipedia table. The highlighted cells were manually selected to provide sufficient and not excessive information to generate the sentence description. All the existing datasets above focus on texts that contain only the information in the inputs. However, some datasets are constructed by ex- tracting real-world data that do not impose these constraints. For example, RotoWire (Wiseman et al., 2017) is a dataset consisting of basketball game …

Content truncated to 500 words