和弦是旋律伴奏的基础，为旋律搭配合适的和弦是音乐制作中不可缺少的步骤。目前深度学习已经被广泛用于和弦生成任务，但是较少研究关注不同的和弦表征方式对于和弦生成结果的影响。本文以和弦的音阶组成对和弦进行表征，将传统的和弦生成任务从单标签分类转换为多标签分类，使得和弦生成模型可以充分利用和弦的音阶信息。通过搭建LSTM网络对传统标签法与音阶法和弦生成任务进行对比，实验结果表明以音阶表征和弦的方式在和弦类别预测上优于传统的标签法，且音阶法可以提升和弦生成模型对少数类和弦的预测能力。

Chord is the basis for accompaniment to the melody, and matching the right chord to melody is an indispensable part of music making. Nowadays, deep learning has been widely used in chord generation tasks, however, few studies have studied the impact of different chord representations on the final result. In this paper, chord composition is characterized by the scale composition of the chord, and the traditional chord generation task is replaced from single label classification to multi-label classification, so that the model can make full use of the scale information of the chord. By building an LSTM network to compare the two methods, representing chords by scale is better than the traditional label method in the prediction of chord categories, and the scale method can improve the prediction ability of the model for minority chords.