What's the fastest way to perform large inserts with foreign key relationships and preprocessing?

I need to regularly import large (hundreds of thousands of lines) tsv files into multiple related SQL Server 2008 R2 tables.

The input file looks something like this (it's actually even more complex and the data is of a different nature, but what I have here is analogous):

January_1_Lunch.tsv
+-------+----------+-------------+---------+
| Diner | Beverage | Food        | Dessert |
+-------+----------+-------------+---------+
| Nancy | coffee   | salad_steak | pie     |
| Joe   | milk     | soup_steak  | cake    |
| Pat   | coffee   | soup_tofu   | pie     |
+-------+----------+-------------+---------+

Notice that one column contains a character-delimited list that needs preprocessing to split it up.

The schema is highly normalized -- each record has multiple many-to-many foreign key relationships. Nothing too unusual here...

Meals
+----+-----------------+
| id |       name      |
+----+-----------------+
|  1 | January_1_Lunch |
+----+-----------------+

Beverages
+----+--------+
| id |  name  |
+----+--------+
|  1 | coffee |
|  2 | milk   |
+----+--------+

Food
+----+-------+
| id | name  |
+----+-------+
|  1 | salad |
|  2 | soup  |
|  3 | steak |
|  4 | tofu  |
+----+-------+

Desserts
+----+------+
| id | name |
+----+------+
|  1 | pie  |
|  2 | cake |
+----+------+

Each input column is ultimately destined for a separate table.

This might seem an unnecessarily complex schema -- why not just have a single table that matches the input? But consider that a diner may come into the restaurant and order only a drink or a dessert, in which case there would be many null rows. Considering that this DB will ultimately store hundreds of millions of records, that seems like a poor use of storage. I also want to be able to generate reports for just beverages, just desserts, etc., and I figure those will perform much better with separate tables.

The orders are tracked in relationship tables like this:

BeverageOrders
+--------+---------+------------+
| mealId | dinerId | beverageId |
+--------+---------+------------+
|      1 |       1 |          1 |
|      1 |       2 |          2 |
|      1 |       3 |          1 |
+--------+---------+------------+

FoodOrders
+--------+---------+--------+
| mealId | dinerId | foodId |
+--------+---------+--------+
|      1 |       1 |      1 |
|      1 |       1 |      3 |
|      1 |       2 |      2 |
|      1 |       2 |      3 |
|      1 |       3 |      2 |
|      1 |       3 |      4 |
+--------+---------+--------+

DessertOrders
+--------+---------+-----------+
| mealId | dinerId | dessertId |
+--------+---------+-----------+
|      1 |       1 |         1 |
|      1 |       2 |         2 |
|      1 |       3 |         1 |
+--------+---------+-----------+

Note that there are more records for Food because the input contained those nasty little lists that were split into multiple records. This is another reason it helps to have separate tables.

So the question is, what's the most efficient way to get the data from the file into the schema you see above?

Approaches I've considered:

1. Parse the tsv file line-by-line, performing the inserts as I go. Whether using an ORM or not, this seems like a lot of trips to the database and would be very slow.
2. Parse the tsv file to data structures in memory, or multiple files on disk, that correspond to the schema. Then use SqlBulkCopy to import each one. While it's fewer transactions, it seems more expensive than simply performing lots of inserts, due to having to either cache a lot of data or perform many writes to disk.
3. Per How do I bulk insert two datatables that have an Identity relationship and Best practices for inserting/updating large amount of data in SQL Server 2008, import the tsv file into a staging table, then merge into the schema, using DB functions to do the preprocessing. This seems like the best option, but I'd think the validation and preprocessing could be done more efficiently in C# or really anything else.

Are there any other possibilities out there?

The schema is still under development so I can revise it if that ends up being the sticking point.