For FSAs and HMMs, strings were generated based on the notion of a transition from one latent state/category to another. Consider the following sentence.

Call me.

We could generate such a sentence using an FSA/HMM using a transition from the verb call to the noun me. There is a fundamental assymetry, however, between the two words in this sentence which is not captured in the way we have conceptualized FSAs/HMMs so far in the course. In this sentence, call is the head of the sentence while me is a dependent of this head. In general, the transitive form of the verb call must always co-occur with a direct object—which is a phrase headed by a noun. In general, heads of phrases are those words which select other arguments as dependents.

Such argument structure is typical of verbs. The verb call selects a dependent noun phrase argument as direct object. Other verbs, such as arrive do not take a noun argument as direct object. Other categories also have such argument structure requirements. For example, prepositions such as on cannot occur with a nominal object: on shore.

The asymmetry between heads and arguments captures many linguistic facts. Phrases are composed by satisfying the argument requirements of head words and they properties are dependent on the head word. For example, in the sentence

I thought I would sail about a little and see the watery part of the world.

The phrse watery part of the world is headed by the noun part and is thus a noun phrase. It’s other properties, such as the existence of an adjective modifying the head are also predicted by the category of the head.

How can we model such head-dependent relations using transitions in an HMM? One basic issue with using an HMM to model argument structure is that the transitions in an HMM don’t make reference to the preceding word. Thus, for example, when deciding whether the word following a verb like call or arrive should be a noun, the HMM cannot make reference to the identity of the verb itself to distinguish between these cases. We might fix this problem by making the next category distributions dependents on the proceeding word in addition to the proceeding category. However, there is a bigger problem.

Consider the following sentence:

You call me.

The verb call takes not only a direct object argument on the right as a dependent, but also a subject argument on the left! Furthermore, consider what happens when you embed the sentence above within another.

I think you call me.

Here, the word you is selected by the head call and not by the word to it’s immediate left, the word think. Our HMM/FSA formalism only had a mechanism to select arguments to the immediately right of each head, but not to both directions!

In the next several chapters, we will study systems that generalize HMMs/FSAs to allow heads to select to both the left and the right. These systems will open up new possibilities for modeling linguistic structure.


39 Markov Chain Monte Carlo 41 Context-Free Grammars