Documentation for the Grammar Matrix Customization Coordination Library

Introduction

While the coordination library is optional (you can create a grammar without using it), if your language has AND-style coordination of any constituent type it is a good idea to have it in the grammar from the beginning. Coordination interacts with many different phenomena, and so building your analyses of the various phenomena with coordination in the mix from the start is a good idea.

Generally, the Grammar Matrix Coordination customization page is divided into two pieces. The coordination library enables users to add coordination strategies to their grammars ("Add a Coordination Strategy"). Additionally, users may describe how agreement features (like person, number, and gender) work in coordinated phrases in the language they are describing.

Descriptive grammars frequently describe coordination strategies, and somewhat less frequently describe how agreement features work within those coordination strategies. It may make sense to add a coordination strategy without also describing how agreement works in that language's coordinated phrases.

Coordination Strategy Options

Coordination Strategy Analyses

The coordination library allows the user to specify different coorindation strategies for different types of coordinands (e.g., NP and S). This results in specialized grammar rules for each type of coordinand. This may seem surprising in light of the very general coordination schemas sometimes encountered in the HPSG literatures, but in fact it is well-supported from both typological and theoretical angles. From a typological point of view, many languages use different coordination strategies for the coordination of different constituent types. From a theoretical point of view, the work that the coordination rule needs to do, both syntactically and semantically, differs according to the type of constituent being coordinated.

The coordination library does not presently constrain the value of any HEAD or INDEX features on the mother. Such constraints can be added to the starter grammar by hand. Note that in some cases, you'll need to create further subtypes of each rule type (e.g., to handle the resolution of such features as PERSON and NUMBER in NP coordination, based on the value of these features on the coordinands).

The analyses implemented in this library are partially described in Drellishak and Bender 2005.

Agreement Features in Coordinated Phrases

The following sections describe the additions to the Matrix that account for how agreement features work in coordinated phrases.

Agreement Pattern Options

Generally, the agreement features of a coordinated phrase are decided one of two ways: feature resolution or distinguished conjunct (which includes closest conjunct and first/last conjunct). The term "agreement pattern" is used throughout these MatrixDoc pages, as well as on the Coordination page, as an overarching term for the feature resolution and distinguished conjunct patterns defined on the page.

Feature Resolution

In feature resolution, the grammar follows fairly predictable logic and/or sets of rules (e.g. "whenever a noun phrase includes a masculine conjunct, the gender of the entire phrase will be masculine", "all coordinated phrases are plural", "any noun phrase including a 1st person element is also 1st person") to derive the entire coordinated phrase's agreement feature value.

If a language uses any sort of hierarchy in a coordinated phrase, this section is where to define that hierarchy. In addition, you can define other patterns that depend on sets of rules: requiring conjuncts to have the same feature values for particular features, or defining "default" values for mismatched values of conjunct.

Distinguished Conjunct

In distinguished conjunct patterns, a single conjunct in a predictable position (first, last, or closest to the verb) decides the features used by a word agreeing with the coordinated phrase. No matter how long or what other elements the phrase contains, the agreement features are always decided by the first conjunct, last conjunct, or the closest conjunct to the verb.

Agreement Pattern Analyses

This section describes the various options and typical paths for defining how agreement features work in coordinated phrases.

Once you have defined at least one agreement pattern, you may associate it with any coordination strategy. In fact, you must link it with a coordination strategy for it to work.

Feature Resolution

The feature resolution section allows you to add full feature resolution-style rule sets to a coordination strategy.

Optionally, you may define a name for each feature resolution pattern. The "name" is a label you write yourself (e.g. "case_only", "full", "resolution", "experimental") that will make the pattern easy to find when attaching it to a coordination strategy.

Adding a Feature Resolution Rule Set

In feature resolution, some or all coordinated phrases use sets of rules to decide the agreement value of the coordinated phrase as a whole. You can add a feature resolution pattern, which will allow you to define these sets of rules.

Feature resolution rules can be complex. Defining feature resolution rules for person values, for example, can easily result in a dozen rules. Unfortunately, you must list these out yourself, making sure not to duplicate or leave out any combinations.

Some things to remember:

First example: all coordinated phrases are plural

Suppose, in your language, all coordinated noun phrases are plural.

Define a feature resolution pattern. Then, add a feature. Select "number" from the feature list. Select "any" for both children, and "pl" for the parent. Now, all coordinated phrases that use this feature resolution pattern will be plural.

"any" & "any" = "pl"

Second example: two grammatical genders

In some languages, there are two grammatical genders, e.g. "m" and "f", and any coordinated phrase including one value will always be that value. You can define this pattern using three rules:

"any" & "f" = "f"

"f" & "m" = "f"

"m" & "m" = "m"

This covers all the possible combinations of the two values, without duplicating or skipping any.

Third example: Person

Values for the person feature can require many rules.

If you want any phrase that includes a 1st person conjunct to resolve to 1st person, you could follow the general pattern for defining a grammatical gender hierarchy:

"any" & "1st" = "1st"

"1st" & "2nd" = "1st"

"1st" & "3rd" = "1st"

For subsequent person values, if the language used 1st, 2nd, and 3rd person, with no inclusive or exclusive values, you could continue like this:

"2nd" & "3rd" = "2nd"

"3rd" & "2nd" = "2nd"

"3rd" & "3rd" = "3rd"

Special Feature Resolution Values

"The same": Identify two or more values

"The same" is used to identify two or more values.

For example, if your language has three genders, "m" , "n" , and "f", and a coordinated phrase with conjuncts of the same gender always has that gender value, you could use this option. By choosing "the same" for Child 1, "the same" for Child 2, and "the same" for Parent, these values will be identified.

"Any": An underspecified value

Selecting a value of "any" will leave the feature value for that child underspecified.

For example, to write a feature resolution rule that any coordinated phrase including a 1st person conjunct is 1st person, you could choose "1st" for Child 1, "any" for Child 2, and "1st" for Parent.

Note that in this case, you would also want to fill in the remaining two feature resolution rules, covering all possible combinations of person that include 1st person: "2nd" & "1st" = "1st"; "3rd" & "1st" = "1st".

"Any non-matching value in that list"

Some languages have feature resolution patterns that define a "default" value for non-matching conjuncts.

For example, if a language has 5 noun classes, and any combination of two different noun classes resulted in a value of noun class 3, you would use this option.

In this case, define a list of feature values under Child 1. The Child 1 list allows you to select multiple options, so select all possible feature values (e.g. noun classes 1, 2, 3, 4, and 5). Then, select "Any non-matching value in that list" for Child 2. Define the "default" value by selecting a feature value under Parent (e.g. noun class 3).

Feature Resolution Tips

Some descriptive grammars do not include very much information about how agreement features work in coordinated phrases. Even without detailed information about agreement patterns in a language, it can be useful to add some constraints to coordinated phrases to cut down on overgeneration and spurious ambiguity in the final grammar.

Out of the box, coordination strategies are underspecified for HEAD and INDEX features. If you do not plan to add additional information about agreement patterns, the Feature Resolution section is a place where you can add some constraints on coordinated phrases.

For example, even if you aren't specifically modeling how case works in coordinated phrases in the language we're modeling, it's possible to identify case values to reduce spurious ambiguity in sentences with coordinated phrases. (See the section on the "the same" value for information on how to identify feature values.)

This will force both coordinands to have the same case value, and the top-level node of the coordinated phrase will also have the matching case value.

Distinguished Conjunct

Compared to a feature resolution pattern, distinguished conjunct patterns are relatively simple to define.

Click the "Add a distinguished conjunct pattern" button, then select one of the radio button options: "First," "Last," or "Closest Conjunct."

Optionally, you may define a name for the pattern. This is simply a label that will make the pattern easy to find when attaching it to a coordination strategy.

Attaching an Agreement Pattern to a Coordination Strategy

As mentioned above, an agreement pattern must be "attached" to a coordination strategy. Coordination strategies are underspecified for agreement features by default. A coordination strategy with an associated agreement pattern will, when customized, allow only the feature combinations described in the agreement pattern.

Note that it is possible to attach multiple agreement patterns to a coordination strategy.

Typical Path

Typically, "All arguments" is the correct option for attaching an agreement pattern to a coordination strategy.

For example, if the language you're modeling uses feature resolution for all of its coordinated phrases, you would first define a feature resolution pattern. After defining as many coordination strategies as you would like, you would attach the feature resolution pattern to all coordination strategies, marking "all arguments" on each.

"All Arguments" vs. "Subject/Object Only"

Even when modeling languages whose verbs only agree with the subject, you should specify "all arguments" when attaching an agreement pattern to a coordination strategy, because feature resolution will typically be used for all coordinated phrases, whether they occur in the subject or object.

The "subject only" and "object only" options support languages that allow particular agreement patterns in the subject or object position only, and exclude it in the others.

For example, if a language allowed feature resolution for coordinated phrases in any argument, but additionally allowed closest conjunct for subjects, you could attach a feature resolution pattern for "all arguments" and the closest conjunct pattern for "subject only."

Similarly, some languages use e.g. feature resolution in subject position, but closest conjunct in object position. In this case, you would attach a feature resolution pattern for "subject only" and a closest conjunct pattern for "object only."

Upcoming Work

References

Greville G Corbett. 2006. Agreement. Cambridge University Press, Cambridge, UK.

Drellishak, Scott and Emily M. Bender. 2005. A Coordination Module for a Crosslinguistic Grammar Resource. Stephan Müller, ed. Proceedings of the 12th International Conference on Head-Driven Phrase Structure Grammar. Stanford: CSLI.

MatrixDoc/Coordination (last edited 2018-08-22 02:10:31 by LaurieDermer)

(The DELPH-IN infrastructure is hosted at the University of Oslo)