In this paper,we present a neural model to map structured table into document-scale descriptive texts.Most existing neural net-work based approaches encode a table record-by-record and generate long summaries by attentional encoder-decoder model,which leads to two problems.(1)portions of the generated texts are incoherent due to the mismatch between the row and corresponding records.(2)a lot of irrelevant information is described in the generated texts due to the in-correct selection of the redundant records.Our approach addresses both problems by modeling the row representation as an intermediate struc-ture of the table.In the encoding phase,we first learn record-level rep-resentation via transformer encoder.Afterwards,we obtain each rows representation according to their corresponding records representation and model row-level dependency via another transformer encoder.In the decoding phase , we first attend to row-level representation to find important rows.Then,we attend to specific records to generate texts.Experiments were conducted on ROTOWIRE,a dataset which aims at producing a document-scale NBA game summary given structured ta-ble of game statistics.Our approach improves a strong baselines BLEU score from 14.19 to 15.65(+10.29%).Furthermore,three extractive eval-uation metrics and human evaluation also show that our model has the ability to select salient records and the generated game summary is more accurate.