We consider semantic analysis to consist of a) basic semantic dependency structure (SDS) building, interleaved with b) expert augmentation of that semantic dependency structure with further constraints or information by a variety of microtheories. The SDS process, in our view, necessarily relies on the syntax-semantics interface, imbedded in the lexical syntactic and lexical semantic specifications.
The bulk of the work described here falls in the category of SDS-building; a brief description of microtheory-based augmentation follows. Each microtheory treats a particular language phenomenon, whether language-specific or general; a working machine translation system in this paradigm would need to have a battery of such microtheories to handle such phenomena as definite reference and aspect. <<<reference to microtheory papers>>> In the process of adding information to (thus further constraining) the semantic analysis produced by the SDS-building process, the microtheories also assist in traversing the search space (described below) by selecting from among candidate analyses generated by the SDS process or other microtheories, by pruning out certain readings entirely, or by adjusting the preference (defined below) of a particular reading or set of readings.
The readings or semantic interpretations generated by the SDS-building process (both as intermediate and as final results) are expressed in TMR. The particular language that is used is not of critical importance, as long as it meets a number of criteria regarding its expressiveness and deterministic properties. Regardless of the language, the meaning is represented as an augmented network of instantiations of concepts from the ontology.
The notion of a preference is utilized to reflect the absolute and relative likelihood of a particular reading or interpretation (i.e., a particular state in the search space, discussed below). The goal of the SDS-building process is to find the most probable semantic interpretation of the input text; the preference of each interpretation considered by the SDS-building process should reflect that interpretation's "goodness". If the assignment of preferences by the search process (and the microtheories) is appropriate, then the interpretation with the highest preference value at the end of processing should be the correct or preferred one (as judged by human translators). The preference is used by the search heuristic both for a) pruning paths with preferences on the interpretation that fall below a minimum threshhold, and b) guiding the search heuristic by contributing towards selecting the most promising path (when a best-first or similar heuristic is used). The preference is maintained currently as a value in the interval [0.0, 1.0], with adjustments to the preference typically made by a multiplier.
Both incrementing and decrementing adjustments are possible, reflecting an increased likelihood on that reading (for example, if the reading reflects the use of a typical collocation or idiom) or a decreased likelihood on that reading (as is the case when any constraint violation occurs). Determining the values by which the preference should be adjusted in a particular case is an issue of critical importance to the success of this approach, and a variety of factors influence those decisions.
The process of building semantic dependency structure is considered here from the point of view of traversing a search space of all possible semantic constructions (both well-formed and incomplete) in order to find the semantic construction that best represents the meaning of the input text. Each state in the search space is referred to as a reading, and has an associated preference reflecting the likelihood of that reading. A particular state may be final (i.e., well-formed and complete), or incomplete (where portions of the meaning of the text have not been incorporated into the reading yet). The two operators for expanding or traversing nodes in the search space (i.e., the processes which actually build the SDS) are instantiation and combination.
The instantiation process instantiates each syntactically appropriate word sense in the input syntactic structure according to the lexical semantic specification of that word sense. It is significant to point out that the final TMR does not represent syntactic information about the input string. However, the approach taken here is consistent with the mainstream semantic analysis approaches in that it does involve syntactic analysis. The syntactic parse produced in the early stage of analysis is considered to be one type of knowledge produced by a dynamic knowledge source; this knowledge is used as input to the heuristics which help guide the two phases of the semantic analysis process. The syntactic parse identifies the lexemes corresponding to words, idioms, or morphemes in the input string, and eliminates those lexemes which do not meet basic syntactic constraints. The syntactic parse structure is used by a heuristic to help guide the application of the combination operator, which actually builds the semantic dependency structure that forms the bulk of the TMR.
Given the instantiated word senses, the combination operator attempts to combine pairs of partial semantic structures according to the dereferenced variables of the syntax-semantics interface. For example, the syntactic specification for a sense of eat may subcategorize for an object, and the syntax-semantics interface (i.e., the $vars) indicates that the meaning associated with the object serves as the THEME of the meaning of eat. Thus the combination operator builds a semantic structure by attempting to insert the meaning representation produced by instantiating the syntactic object into that role in the meaning representation of eat.
This combination process essentially involves creating a relation between two concepts; in the representation, this is done by having a slot on one concept have the name of another concept as its value. A number of constraints guides this linking --- a) the constraints on the range (and domain) of the relation in the ontology, b) possible further constraining of the range in the head concept's local specification of that relation, and c) the lexical semantics of the head may further constrain the possible fillers of that slot. For example, the case role agent has a constrained range (animals or other animate entities may be agents); the concept ingest constrains the agent to be animal; the German word freßen maps to the concept ingest and further constrains the agent to be non-human.
The determination of which slot that one meaning representation element is to be inserted into (i.e., the specific relation that holds between the two instantiated meanings) is part of the SDS-building process. In fact, determination of which element is to be the head and which is to be the filler is also part of the SDS-building process. Three cases occur:
The constraints on slot fills are defined in terms of concepts from the ontology. Thus, since the candidate filler is a constrained concept from the ontology, and the constraint is a concept from the ontology, the constraint satisfaction process is (in the base case) a matter of verifying that the filler head is in the subtree of the ontology headed by the constraint concept. In other cases, the constraint satisfaction process involves exploring other (non-taxonomic) paths between the candidate filler and the constraint. These other paths may define a metaphorical or metonymic relationship between the candidate filler and the constraining concept.
Given this view of semantic composition as a constraint satisfaction problem, and that the candidate filler and the constraints are both elements in an interconnected graph, this process is treated as a low-cost path problem over a graph <<<give references to appropriate graph lit.>>>. The determination of the arc costs provides the control over the graph traversal. Conceptually, this process determines the distance over the ontology between the two concepts. The more ontologically related two concepts are (in a particular definition of related), the shorter the path. The relation between the two concepts can be vertically taxonomic, or any of a variety of other relations that reflect conceptual relatedness between two concepts (such as composer and his work, sword and scabbard, part and whole, to taxi and airplane, landing strip and airplane).
Each arc in the graph (i.e., all relations) are directional; inverse links are present for each link, typically with a different cost, however. Since the graph has some tree-like structure, the graph traversal is typically computed as originating at the candidate filler, and the destination is the constraint; in discussing the paths, the reverse direction is often more intuitive. In the trivial case of verifying that the candidate filler is in a subtree headed by the constraint, the graph is treated as a tree (i.e., ignoring non-taxonomic links); the cost of an IS-A arc is set to be zero or very low, and the cost of a SUBLASSES arc (the inverse of the IS-A arc), as well as all other links, is set to infinite. Thus the constraint satisfaction test is trivially treated.
In many cases, however, the simple IS-A test will fail, because the base constraints are established for literal meaning. Thus cases of metonymic or metaphoric text the IS-A constraints fail; the graph traversal is expanded to include other arcs (relations) in the ontology, where the other arcs, appropriately weighted, reflect the variety of relations that exist between concepts. The sorts of relations that are exhibited in cases of metonymy and (some) metaphor are among those additional relations in the ontology.
The example in Figure 5A. and Figure 5B. illustrates how two paths over the ontology can result in different weights: .81 and .85, respectively. [Figures 5A and 5B are not available.] This path mught be appropriate in an example such as Fred drove his dual-cam V8 down Main street, where the engine is used metonymically for the vehicle; in the semantic representation for drive, the constraint on what could be driven might specify the VEHICLE concept from the ontology. If the arc weights are set appropriately, the shortest path from the constraint to the filler will be reflect the metonymy by traversing the arc capturing the relation embodied in the metonymic.
5.2 Generation