This baseline roll seen in this spectrum is an example of acoustic ringing, and is quite normal for many low frequency NMR experiments that are recorded over wide sweep widths. It occurs as a consequence of slight breakthrough of the excitation pulse, as it rings down to zero. Ideally, you would like to start collecting signal at your receiver immediately after your excitation pulse, however it does take a small amount of time (microseconds) for the pulse to decay to zero. If your pre-scan delay (typically configured for your spectrometer during the configuration process) is too short, you can get this ringing artifact. It is more pronounced for weaker samples, but usually nothing to worry about.
To correct this as an experimental acquisition parameter, you would make your pre-scan delay longer; for Topspin, this is the parameter DE in your acquisition parameter list. It is normally defined during spectrometer configuration, and needs to be as short as practically possible to maximise your signal detection - typically less than 10 microseconds. For some nuclei, much of the signal has already decayed during this pre-scan period, and require some backward linear prediction to help. I'd speak to your local friendly NMR spectroscopist before mucking about with changing parameters like DE.
It is perfectly normal, and you should be able to get rid of it by applying a simple baseline correction algorithm - for Topspin you should use the commands abs or absd. This should work for your spectrum. As a word of caution when applying baseline corrections to your 1H spectrum, this will also apply an automatic integration for you, overwriting any integration you may have done previously. You can avoid this by using the optional n switch (abs n). For severe cases of baseline roll, apply a partial basline correction using absf between the limits of absf1 and absf2. Your Topspin manual can provide a fuller description, including other baseline correction modules.