Background. Nowadays, Requirements Engineering(RE) focuses not only the pri- mary stakeholders but also large scale data, which comes from an amount of cus- tomers’ operations and feedback. For instance, the user reviews in the mobile app platform become an important analysis target, since they contain a lot of informa- tion. However, data-driven requirements technology is a new domain. There might be many challenges hindering elicitation.

Objectives. This research aims to investigate the state of the art in data-driven requirements elicitation. Conclude challenges in using data-driven requirements elic- itation techniques. And provide a solution to overcome a challenge for data-driven requirements elicitation technique in practical.

Methods. In this project, we use two methods for our topic. We choose to use the systematic literature review and controlled experiment as our methods. According to the systematic literature review, we can get to know the state-of-art of data-driven requirements engineering and the challenges in using these techniques. And we will try to raise a solution for one of the challenges which we listed through literature. For the feasibility of this solution, we will design an experiment to test it. The controlled experiment can help us to know whether the solution can work or not.

Results. In the systematic literature review part, we finally get 44 papers in data- driven requirements elicitation area. We make a list for techniques in these papers, and divide them into three categories. These techniques are getting requirements from natural language, usage data and management data. We also list challenges from these papers. In the experiment part, we try to modify an irrelevant classifier in a technique called ri-analytics-classification-twitter. And finally, we get a new model for the classification in these techniques.

Conclusions. In systematic literature review, we find only several techniques have implemented in industry and had cases. These are still having lots of challenges in technology itself and human factors. In the experiment, we develop a model for classifying irrelevant tweets. And this model still has a lot of room to improve.