![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
guslDoes anyone here work on Bayesian models of sentence interpretation, whether focused on semantics or pragmatics?
First, a Bayesian model of sentence parsing:
If an agent hears an utterance U, the process of parsing is a search through the space of parse-trees that tries to find the maximum-probability parse-tree.
P(U) = SUM_i P(U|tree_i) P(tree_i)
This last factor is a prior over all possible tree-structures. The set of plausible subtrees is determined by the grammar... but to go to a finer level, it's also determined by what is being said (since the parse tree includes the sentence).
To go one level deeper, we could break the prior on trees to incorporate meanings:
P(tree) = SUM_i P(tree|meaning_i) P(meaning_i)
(We could now search for the most probable meaning, i.e. the one that dominates this sum)
This last term is a prior over all possible meanings. This is determined by:
* the world in which the speaker lives
* his reasons for deciding to express this meaning
Speakers try to avoid ambiguity, while keeping the utterance as short as possible, by relying on context.
To go one level deeper, we could break up the prior over meanings expressed.
P(meaning) = P(meaning|goals) P(goals)
The probability of expressing a given meaning is dependent on the speaker's communicative goals. (Can we assume that people follow Grice's maxims? http://en.wikipedia.org/wiki/Gricean_maxim )
One of the most common communicative goals is to provide information that the listener desires.
In this case, our prior on goals should take into account the speaker's model of the listener (in particular, what the listener knows and wants to know).
Enough dabbling... has anyone seen this before?
-Gustavo
P.S. to go off on a tangent: I've seen papers about game theory applied to some pragmatic phenomena (e.g. politeness).
... but I've never seen a working system... probably for the same reason that all computational semantics is hard: ontology engineering is expensive.
(facebook post)
First, a Bayesian model of sentence parsing:
If an agent hears an utterance U, the process of parsing is a search through the space of parse-trees that tries to find the maximum-probability parse-tree.
P(U) = SUM_i P(U|tree_i) P(tree_i)
This last factor is a prior over all possible tree-structures. The set of plausible subtrees is determined by the grammar... but to go to a finer level, it's also determined by what is being said (since the parse tree includes the sentence).
To go one level deeper, we could break the prior on trees to incorporate meanings:
P(tree) = SUM_i P(tree|meaning_i) P(meaning_i)
(We could now search for the most probable meaning, i.e. the one that dominates this sum)
This last term is a prior over all possible meanings. This is determined by:
* the world in which the speaker lives
* his reasons for deciding to express this meaning
Speakers try to avoid ambiguity, while keeping the utterance as short as possible, by relying on context.
To go one level deeper, we could break up the prior over meanings expressed.
P(meaning) = P(meaning|goals) P(goals)
The probability of expressing a given meaning is dependent on the speaker's communicative goals. (Can we assume that people follow Grice's maxims? http://en.wikipedia.org/wiki/Gricean_maxim )
One of the most common communicative goals is to provide information that the listener desires.
In this case, our prior on goals should take into account the speaker's model of the listener (in particular, what the listener knows and wants to know).
Enough dabbling... has anyone seen this before?
-Gustavo
P.S. to go off on a tangent: I've seen papers about game theory applied to some pragmatic phenomena (e.g. politeness).
... but I've never seen a working system... probably for the same reason that all computational semantics is hard: ontology engineering is expensive.
(facebook post)