In many previous deep IR models, only the global matching score is employed to perform ranking. However, user’s queries may be of various nature which might require either low-level matching score, high-level matching score or a combination of theses scores. In order to train a deep IR system without click-through data and investigate the in influence of employing multi-level matching scores on the model, we proposed a multi-level abstraction convolutional model with weak supervision (MACM) to perform IR tasks. We thoroughly investigated the effectiveness of employing multi-level matching scores on IR models. Experimental results on classic dataset have demonstrated the effectiveness of our proposed MACM model. We hope that this study could offer researchers some insights into deep IR models.