Neither - they are usually effectively equivalent.
This is because:
password crackers are very aware of non-random, human-understandable passphrases - and can (and do) emulate both subtypes at high speed, and
injecting memorable additional sequences like "!" and ":)" is just as well known and "simulatable", and
the effective entropy of either kind of passphrase-based construction technique once the methodology is known to the attacker are virtually identical.
It's #3 that is the most important lesson here, and why raw bruteforce is irrelevant: because the effective entropy of this method - how much work a real-world attacker has to perform - is dramatically less than the bruteforce entropy. This is because the psychology of naive password-memorization strategies is well-known to the attacker.
In your specific methodology, you only have to remember about eight things to remember the password:
- the fact that you're using a passphrase at all (this is a freebie for the attacker, because they will be attacking passphrases as a class)
- the rough meaning components of the original phrase (liking, food, and which two foods, perhaps roughly four pieces of information) (since the phrase makes grammatical sense, much of this is also easy for the attacker)
- the fact that you did or did not use whole words vs the first letters (since there are only a few options here, this isn't hard for the attacker to exhaust both variants)
- the placement and location of the two additional items (attackers know that people tend to put the special-character "extra" stuff at the end)
If the attacker had no idea at all in the history of password psychology that anyone would ever do this, you'd be in pretty good shape. But attackers know very well how people think about passwords, and what they do to "chunk" them into "memorizable" components. (In other words, your methodology does not adhere to Kerkhoffs' Principle).
So an attacker only has to "bruteforce" the psychologically likely effective entropy of your password. The attacker doesn't have to bruteforce all possible characters of length X; instead, they can dramatically reduce the things they have to guess by only trying the same possibilities that make them easier for you to remember.
By contrast, the most effective passphrases are random passphrases: five or six words, drawn purely randomly from a wordlist of, say, 20,000 words or more (20,000^5, or 3.2x10^21). Even if the attacker knows everything that you do about the construction methodology - exactly how many words are in the list, and exactly how many words you selected, and how you separated them, etc. - the sheer volume of possibilities is resistant to bruteforce. You only have to memorize five words - but they are so many possibilities that trying them all would take a very, very long time. If you do the math of this, you'll quickly start to understand.
(And the concern about "using words in the dictionary" is a leftover from old password complexity requirements that were trying to prevent people from using single words from a dictionary. The underlying principle - avoiding "guessability" - is a good one, but it's totally OK to use a large number of random words that happen to be in the dictionary)
Any password system whose strength is based solely on a large amount of randomness will always be stronger than one that isn't.