Increasing The Performance of Cognitively Inspired Data-Efficient Language Models via Implicit Structure Building