Reducing the Vision and Language Bias for Temporal Sentence Grounding