Interests

#NLG

My current work concerns with machine production of natural language (~English). We want to see what kind of information we need/should provide to a system (~neural network) for it to learn generating natural language texts. For instance:

CONTRAST (visits US Bill) (visits Canada Alice) ----|> Unlike Bill, who is visiting U.S., Alice is visiting Canada.

(visits US Bill) (visits Canada Alice) ----|> Unlike Bill, who is visiting U.S., Alice is visiting Canada.

SIMILARITY (visits US Bill) (visits US Alice) ----|> Like Bill, Alice is visiting U.S.

(visits US Bill) (visits US Alice) ----|> Like Bill, Alice is visiting U.S.

We want to find out whether having explicit markers of discourse relations (e.g. CONTRAST, SIMILARITY) are helpful to have in the input in order for a model to generate a fluent text with the intended meaning.

#DiscourseGrammars

Formal grammars such (as the following one and slightly more complicated ones) have been long argued to be able to encode much of natural language syntax (clause/sentence level descriptions of co-occurrence patterns in natural language).

S -> NP VP .

NP -> Alice | Bill | Canada | U.S.

VP -> V NP

V -> visits

While Chomsky's interests mainly concerned studying syntax with grammars, Montague translated sentences generated by grammars (like the one above) into formulas of a logic.

What about texts, discourses, can we think of a grammar describing them? E.g. words unlike, but, and, can we design a grammar that encodes them as discourse connectives? We want to write down a grammar rules like to build surface forms. We want to be able interpret those surface forms, i.e., to attribute meanings to them. For instance, we want to be able to model texts such as the following one:

Bill is visiting Canada. But he planned to visit U.S.

A grammar may have the following rule to encode the latter text:

x But y -> CONTRAST (x', y')

Where x and y are sentences; x' is the meaning of x and y' is the meaning of y.

With a little more effort, using grammars like CFGs, we may achieve to encode some simple cases like above.

What about a slightly more complicated cases, such as the following one:

Bill is visiting Canada. He however planned to visit U.S.

Here the word however acts as a VP modifier at the clause level, modifying planned to visit U.S.

At the discourse-level, the however signals CONTRAST between Bill is visiting Canada and he planned to visit U.S. How to model this behavior of however in a computationally affordable way is a question in a discourse grammar community, which I also addressed in my dissertation using Abstract Categorial Grammars (ACGs).

#ProbabilitySemanticsInference

Let us assume you are told the following statements hold:

All linguists are programmers. Bill is a linguist.

If you're asked if Bill is a programmer, you'd answer yes.

A vast majority of linguists are programmers. Bill is a linguist.

Is Bill a programmer?

If I were to answer it, I'd say yes is more probable, than no.

Our aim is to have a model that would enable one to make inferences in the cases like ones above.

To do that, we build a model where premises are true (we try to make them as true as possible).

In the model, we evaluate the conclusion---how likely it is to hold.

The model is probabilistic and a new piece of information updates the model by a Bayesian update.

For sake of illustration, imagine a model where linguists denote a set LNGVST. Programmers denote a set PRGRMMR. A premise is that a vast majority of linguists are programmers. The model is updated so that the intersection of the sets LNGVST and PRGRM is a vast majority of the set LNGVST.

Now take a random member, call it Bill, of the set LNGVST (now we're building a representation of the conclusion). What is the probability that Bill belongs to the intersection of the sets LNGVST and PRGRMMR? It should be rather high than low (because the intersection of the sets LNGVST and PRGRMMR makes a vast majority of LNGVST). Thus, the answer would be rather yes than no.

#SyntaxMorphology

Being a speaker of Georgian, which is considered to be one of agglutinative languages and the one to be highly irregular, I am interested in its linguistic properties. Georgian has polypersonal agreement (with both subject and objects in number and person). It also exhibits some traits of polysynthetic languages. Given all these facts, till today there is no clear understanding where Georgian belongs to: Is it polysynthetic? If yes, then is it agglutinative or fusional? If no, is it still agglutinative or it is fusional? Is it possible that Georgian could be something else? In these terms, I want to find out what are other languages that might have similar characterizations as Georgian.

Georgian, like a number of languages, is not a Nominative-Accusative language. That is, a noun phrase referring to an agent may have case markers that are not nominative case (NOM). Georgian is rather a Split Ergative language. In particular, the noun phrase corresponding to an agent/subject can have three different case markers. This nature of Georgian gives rise a number of interesting questions to study. One of them is coordination in Georgian, and in particular, the coordination and ellipses when the same noun phrase in different case markers is being used in two coordinated verb phrases. Let us consider the following example:

kata xeze azvra da [katam] chiti shechama.

cat_NOM tree_on climbed and [CAT_ERG] bird_NOM ate.

A cat climbed on a tree and ate a bird.

As you can see, kata (cat) is in nominative, but katam, which is an agentive form required by the verb shechama (ate) is in ergative. In addition, the experiencer chiti (bird) is in nominative, which is the same case marker as the one of the agent. So, we have the coordination and partial agreement (with only one of the verb phrases). Moreover, there is an experiencer that has a case marker in an agentive case (nominative).

#XbarGeorgian

In a non-standard student of mathematics, according to X-bar, student and mathematics are close (adjacent) because student is an argument of the predicate mathematics (let us call this principle the function-argument closeness principle).

In Georgian, we say:

matematikis arastandartuli studenti

mathematics_GEN non-standard_NOM student_NOM

A non-standard student of mathematics

Thus, the function-argument closeness is not that evident in the case of Georgian. We can still say:

arastandartuli matematikis studenti

non-standard_NOM mathematics_GEN student_NOM

A student of non-standard mathematics

Hence, when matematikis (mathematics_GEN) and studenti (student_NOM) are adjacent, we have that arastandartuli, which is on the right of matematikis studenti, modifies matematikis (mathematics_GEN) rather than studenti (student_NOM). That is, in the latter example, arastandartuli matematikis studenti means a student of non-standard mathematics. This is the salient, primary reading. Using intonations, one can argue that arastandartuli matematikis studenti still can have a meaning that a non-standard student of mathematics has but this is hardly the salient reading. It would be interesting to investigate such cases in more details. Also, it would be interesting to see whether other language also show similar grammatical traits when it comes to noun phrases such the one discussed here.