Semantic NLP applications often rely on dependency trees to recognize major elements of the proposition structure of sentences. Yet, while much semantic structure is indeed expressed by syntax, many phenomena are not easily read out of dependency trees, often leading to further ad-hoc heuristic post-processing or to information loss. To directly address the needs of semantic applications, we present PropS - an output representation designed to explicitly and uniformly express much of the proposition structure which
is implied from syntax, and an associated tool for extracting it from dependency trees.
Props currently supports both English and German versions.
Dependency trees are great for verbal predications, as each verb directly heads its arguments. This doesn't hold for other types of predicates.
We rearrange the tree structure so that this is true for other types of predications.
tall (adjectival modifier) directly heads boy and marked as a predicate.
For conditionals, we consider the marker word (e.g, because, although, unless, etc.) as a predicate, taking as argument both clauses, using the condition and outcome labels, as it introduces the logical relation between them.
Dependency, on the other hand, marks come as root, on which build is dependent, which, finding the conditional only as a daughter node of build, which requires the application to perform a multiple passes over the tree to recover the logical relation between the two clauses.
Different syntactic realizations may evoke the same proposition structure (e.g., passive versus active voice). We aim to present a unified proposition structure when applicable.
For example, note that this and the following example receive identical PropS structure, while the dependency trees diverge significantly
Different syntactic realizations may evoke the same proposition structure (e.g., passive versus active voice). We aim to present a unified proposition structure when applicable.
For example, note that this and the previous example receive identical PropS structure, while the dependency trees diverge significantly
Similarly to the previous examples, cases of raising-to-subject (you looked impatient) which we represent similarly to the semantically similar adjectival predication (you are impatient). Demoting the verbs (look, seem, etc.) to modifier status.
Similarly to the previous examples, cases of raising-to-subject (you looked impatient) which we represent similarly to the semantically similar adjectival predication (you are impatient). Demoting the verbs (look, seem, etc.) to modifier status.
We take special care in minimizing arguments scope where possible. This creates more salient entities, which can then be more easily used in various semantic tasks (OIE, knowledge base population, entailment, etc.).
We take special care in minimizing arguments scope where possible. This creates more salient entities, which can then be more easily used in various semantic tasks (OIE, knowledge base population, entailment, etc.).
Dependency introduces a special appos node from Obama to comma, which require delicate handling - identifying the two entities which stand in the apposition relation, and propagating the relations.
In PropS this is trivially extractable from the structure (as can be seen in the OIE clauses below), without introducing new types of edges.