Thanks for your answer and now I understand how OLE works. Below I summarized my reasoning for the rest of the procedure to compute the secret shares of a Beaver's triple. Please let me know should anything be wrong in here.
=====
Basically, for a Beaver triple, we want to find the triple {z, x, y} such that z = xy. After generating the Beaver triple, Alice should get only {z_A, x_A, y_A}, whereas Bob should get only {z_B, x_B, y_B}.
Notice that z = xy is equivalent to (z_A + z_B) = (x_A + x_B)⋅(y_A + y_B) = x_A⋅y_A + x_A⋅y_B + x_B⋅y_A + x_B⋅y_B.
We suppose that Alice randomly generates the secret shares x_A and y_A, whereas Bob randomly generates the secret shares x_B and y_B. Based on this setup, Alice should find z_A and Bob should find z_B, without knowing about each other's secret shares.
Alice knows what x_A⋅y_A is, and Bob knows what x_B⋅y_B is. Now, their next goals are to find out the secret shares for x_A⋅y_B and x_B⋅y_A, each.
Let the secret shares of x_A⋅y_B = c = c_A + c_B, and x_B⋅y_A = d = d_A + d_B.
Then, we can let z_A = x_A⋅y_A + c_A + d_A, and z_B = x_B⋅y_B + c_B + d_B.
First, to compute c_A and c_B, Bob randomly generates n_B. Then, Bob now has (y_B, n_B) and Alice has x_A. Based on these, Alice and Bob can do OLE (Oblivious Linear Evaluation), which gives Alice x_A⋅y_B + n_B. Then, we can let c_A = x_A⋅y_B + n_B and c_B = -n_B (note that Bob knows -n_B because he knows n_B). Doing this, c_A + c_B = x_A⋅y_B. Thus, Alice and Bob has successfully created valid secret shares {c_A, c_B} for x_A⋅y_B.
Similarly, Alice and Bob can create valid secret shares {d_A, d_B} for x_B⋅y_A.
At this point, Alice can compute z_A = x_A⋅y_A + c_A + d_A, and Bob can compute z_B = x_B⋅y_B + c_B + d_B. Therefore, Alice now gets {z_A, x_A, y_A}, whereas Bob gets {z_B, x_B, y_B}, which are secret shares of a Beaver's triple.
=====
In fact, the secret shares for multiplication of two secret values can be computed by using only OLE. Yet, we still prefer doing this by Beaver's triples, because OLE requires processing log(q) sets of secret shares (for each bit of the secrets), whereas the method based on Beaver's triple requires fewer secret shares to be processed.