Ida.liu.se

Integrating linguistic and domain knowledge for spoken dialogue systems in
multiple domains
Myroslava O. Dzikovska, James F. Allen, Mary D. Swift
Abstract
comes with its own ontology, many aspects of spoken dia-logue interaction are common across domains. This obser- One challenge for developing spoken dialogue sys- vation has led us to develop a new method that customizes a tems in multiple domains is facilitating system broad-coverage, domain-independent parser to new domains component communication using a shared domain that allows us to preserve the language information that is ontology. Since each domain comes with its own common across domains, while addressing specialization is- set of concepts and actions relevant to the appli- sues unique to each domain as much as possible and at the cation, adapting a system to a new domain requires same time keeping the development for custom domain adap- customizing components to use the ontological rep- resentations required for that domain. Our research Traditionally, the most common way to adapt a dialogue in multiple domain development has highlighted system to a new domain is to map the words in the system differences in the ontological needs of a general- lexicon directly to the concepts in the domain model, and purpose language interface and a task-specific rea- write a grammar customized for the domain (see for example soning application. Although different domain ap- [Seneff, 1992]). This process yields a parser that can quickly plications have their own ontologies, many aspects and accurately obtain semantic representations of in-domain of spoken dialogue interaction are common across utterances in the form needed by the system reasoning com- domains. In this paper, we present a new method ponents. However, the performance gains in accuracy and of customizing a broad-coverage parser to different efficiency are offset by the lack of portability of the language domains by maintaining two ontologies, one that interpretation components, which often require large portions is generalized for language representation, and an- of the grammar and lexicon to be rewritten to adapt them to other that is customized to the domain, and defining mappings between them. In this way we preservethe broad-coverage language components across A system that has a method of preserving syntactic in- domains as well as produce semantic representa- formation is AUTOSEM [Ros´e, 2000], which uses the syn- tions that are optimally suited to the domain rea- tactic lexicon COMLEX [Macleod et al., 1994] as a source of reusable syntactic information, and manually links sub-categorization frames in the lexicon to the domain-specificknowledge representation. The linking is performed directly Introduction
from syntactic arguments (e.g., subject, object .) to the slots One of the challenges in developing spoken dialogue sys- in a frame-like domain representation output by the parser tems for multiple domains is making the system components and used by the reasoners. Ros´e’s approach speeds up the communicate with each other using a specific domain on- development process for developing tutoring systems in mul- tology. Each domain comes with its own set of concepts tiple domains, but does not provide a mechanism for preserv- and actions that are relevant to the application, so adapting ing general semantic representations across domains.
a system to a new domain means that all components must Our approach introduces an intermediate layer of abstrac- be able to interface with the ontological representations re- tion, a generic language ontology for the parser that is linked quired for that domain. Researchers working with multi- to the lexicon and allows us to preserve general syntactic and component systems typically require all components to use semantic information across domains. Thus we maintain two a shared ontology (e.g., [Goddeau et al., 1994]), or they de- ontologies: domain-independent for the parser, and domain- velop methods to map between the internal representations specific for the knowledge representation. The parser uses the and a communication language which uses a shared ontol- generic language ontology to create meaning representations ogy (e.g., [Gurevych et al., 2003]). Our research in multiple of the input speech which can then be converted into the re- domain development has highlighted differences in the on- quired domain representation. To integrate the linguistic and tological needs of a general-purpose language interface and domain knowledge, we define a set of mappings between the a task-specific reasoning application. While each domain general semantic representations produced by the parser and the domain-specific knowledge sources localized in the back- end application. We also use the mappings to specialize thelexicon to the domain, resulting in substantial improvementin parsing speed and accuracy.
The customization method described here was developed in the process of adapting the TRIPS dialogue system [Allen et al., 2001] to multiple domains, including a transportationrouting system [Allen et al., 1996] and a medication schedul- ing adviser [Ferguson et al., 2002]. We assume a generic di- alogue system architecture [Allen et al., 2000] that includesa speech module, a parser (responsible for syntactic parsingand initial logical form generation), an interpretation manager (responsible for contextual processing and dialogue manage-ment), and a back-end application that is responsible for the Figure 1: Ontology fragment for physical objects in TRIPS general problem-solving behavior of the system.
The paper is organized as follows. First we contrast the needs of linguistic ontologies and domain ontologies, in sec- reasoning in the Pacifica domain, it is not the best choice for tion 2. Next, we describe the organization of our domain- the language ontology, especially in a multi-domain system.
independent lexicon and ontology and how the parser uses In the ontology for physical objects, making the difference it to generate a semantic representation of the language in- between MOVABLE-OBJ and FIXED-OBJ a top-level dis- put that is then converted to representations that can be used tinction is counterintuitive in our medication adviser domain.
by the domain-specific reasoners,in section 3. In section 4, Having a COMMODITY category is also not a good choice we describe how the domain-independent representations are for a general ontology. In a medication adviser domain, this linked with the domain-specific KR ontology via a set of category is not relevant to the task, and people should be clas- transforms. In section 5, we specify how the transforms are sified as living beings. Arguably, this could be solved by in- used to specialize the generic lexicon, which improves speed troducing multiple inheritance and making PERSON a child and accuracy for the parsing of in-domain utterances. Finally, of both COMMODITY and LIVING-BEING. However, this in section 6 we discuss issues involved in using this method does not solve the problem completely, because it is difficult to classify many physical objects consistently as cargos. Forexample, whether a door is a subtype of COMMODITY may Linguistic ontology versus a domain model
be dependent on a specific application.
In the process of developing our dialogue system in multiple Similarly, the action ontology has a set of associated prob- domains, it has become clear that the language and domain- lems. We need to have the type restrictions to limit the parser specific knowledge representation have differing needs. As search space and improve disambiguation accuracy. Yet the an example of a domain-specific ontology, consider our Paci- type restrictions tailored for the Pacifica domain are not well fica transportation domain. In this domain, the user is given suited to other domains. For example, the sentence I moved the task to evacuate people and possibly other cargo before an to a different place clearly does not fall within the Pacifica approaching storm. A fragment of an ontology for physical representations of either MOVE or TRANSPORT. We could objects used by the planner for this domain is shown in Fig- add a new class to the ontology to represent these actions, ure 1. The top-level distinction is made between fixed objects but this would mean adding another sense to the word move, such as geographical-locations, and movable objects such as increasing the ambiguity in the lexicon and making main- vehicles and commodities, which are suitable for transporta- tenance more difficult. The very specific categories, JUST- tion. People are classified as a kind of commodity because TRANSPORT and MOVE-THEN-TRANSPORT, cannot be they are transported as cargo in the Pacifica scenario.
distinguished in the lexicon at all, and they just add confu- The planner knows 3 main actions: MOVE, TRANS- sion to trying to determine which are the correct ontological PORT and LOAD. MOVE is the action when a vehicle is categories for different words in the lexicon.
moved between 2 points, without any loading actions in- From a parsing perspective, a good classification for move- volved. TRANSPORT is the action of transporting a cargo ment verbs would have a general MOTION type which covers between the two points. If the vehicle is not specified explic- all instances of objects being moved, whether they are cargo itly, a suitable vehicle needs to be chosen, and actions neces- or vehicles, with a subtype for transportation covering the sary to load it planned; LOAD is the action of putting a cargo verbs where transportation of cargo is clearly implied, e.g., into a vehicle. Pragmatically, TRANSPORT is further subdi- transport, ship. For physical object classification, we still vided into a JUST-TRANSPORT action, when there’s already want to retain the information on whether some objects (such a vehicle at a place where the cargo is located, and MOVE- as mountains) are inherently fixed or not, but it should not be THEN-TRANSPORT, where a vehicle needs to be moved to included as the topmost division in the hierarchy.
that location first. Ontology definitions representing those ac- Another point of tension between the needs of parsing and reasoning components is the argument structure. From the Although this ontology is optimally suited for planning and point of view of planner, it would be most convenient to have Figure 2: Action ontology fragment in TRIPS Pacifica domain.
the definition of MOVE and TRANSPORT which have a sin- cally for a given domain, so the concepts in the ontology are gle :PATH slot, where the :PATH slot has a complex value organized in ways that are best suited for domain-specific rea- with :SOURCE, :DESTINATION and :VIA slots. This was soning. The KR ontology makes fine-grained distinctions be- the original implementation in the TRAINS-95 system, but it tween concepts as relevant to the domain, so its hierarchical was abandoned because its divergence from linguistic struc- structure is deeper than that of the LF ontology. And because ture made it difficult to implement in practice. For exam- the KR concepts are organized to facilitate reasoning in the ple, the values filling :PATH slots can come from a variety task domain, their representation may be inconsistent with of adverbial modifiers, e.g., to, toward for :DESTINATION, how concepts are expressed linguistically, as is the case with through, via, by for :VIA, and occasionally from a direct ob- the MOVE and TRANSPORT concepts described above.
ject, such as leave Avon for Bath at 5pm. It proved difficult In the remainder of the paper, we describe our method of to handle this kind of variation in the grammar without in- integrating the information in the domain-independent lin- troducing domain-specific rules which were not suitable in guistic ontology with domain-specific KR ontologies to max- other domains. A representation more suitable for a parser imize both system portability and efficient parsing.
is a MOVE type with TO-LOC and FROM-LOC argumentswhich can be filled either by the subcategorized complements From language to general semantic
or by PP-modifiers as they are encountered.
representation
Given these issues, we decided to separate the ontology The LF ontology is close in structure to linguistic form, so it used in parsing (the LF ontology) from the ontology used in can be easily mapped to natural language and used in mul- reasoning (the KR ontology). The design considerations for tiple domains. It classifies entities (i.e., objects, events or the LF and KR ontologies are summarized in Table 1.
properties) primarily in terms of argument structure. Every The LF ontology is designed to be as general as possi- LF type declares a set of linguistically motivated thematic ar- ble and cover a wide range of concepts that are needed to guments, a structure inspired by FRAMENET [Johnson and generate semantic representations of the language input for Fillmore, 2000], but which covers a number of areas where use in a variety of practical domains. Accordingly, the LF FRAMENET is incomplete, such as words related to plan- ontology has a relatively flat structure and makes only lin- ning and problem solving. We use the LF ontology in con- guistically motivated sense distinctions. For example, the LF junction with a generic grammar covering a wide range of ontology distinguishes between LF CONSUMPTION and syntactic structures and requiring minimal changes between LF REMOVING senses of take, since these two senses can be domains. For example, adapting the parser from the trans- distinguished with different syntactic and semantic patterns.
portation to the medical domain required adding LF types for The LF CONSUMPTION sense of take requires a consum- medical terms (our generic hierarchy was incomplete in this able substance as an object and often occurs with a temporal area) and corresponding vocabulary entries, but we did not modifier on the take event, as in I take an aspirin at bedtime, need to change the grammar or existing lexical entries, and while the LF REMOVING sense of take requires a moveable we continue to use the same lexicon in both domains.
entity as object and often occurs with a path modifier on the The LF types in the LF ontology are organized in a single- event, as in The truck took the cargo to the station. The se- inheritance hierarchy. We implement multiple inheritance via mantic representations we generate from the LF ontology are semantic feature vectors associated with each LF type. Each as close to the original linguistic structure as possible, so there LF type has an associated feature set type, which classifies is straightforward mapping between lexical entries and their entities as physical objects, abstract objects, propositions, sit- uations or times. The feature type is associated with a set The KR ontology, on the other hand, is designed specifi- of features that encode basic meaning components used in LF Ontology
KR Ontology
As general as possible, broad coverage of concepts Domain-specific concepts and organization Relatively flat structure and linguistically motivated Deeper structure, with fine-grained distinctions between Simple representations that are close to linguistic form Roles organized for efficient reasoning w/o regard for lin- Table 1: Design considerations for LF and KR ontologies.
semantic restrictions, such as form, origin and function for the same LF type, because, as noted in section 2, we only physical objects. For example, Figure 3 contains LF type def- make the distinctions in the LF ontology that are linguisti- initions used for various types of vehicles (LF Vehicle), food cally motivated and as independent of any given domain as items (LF Food) and people (LF Person). The feature vector possible. For example, cram, fill, pack and load are all linked for LF Vehicle, denoted by :sem, encodes the fact that it is a to LF Filling. It is difficult to further subdivide them in a physical entity, which has the form of an object (as opposed way that will be consistent across domains, therefore the LF to a formless substance), is man-made (origin artifact), func- ontology does not make any further distinctions.1 In the Paci- tions as a vehicle (object-function vehicle), and is a mobile fica domain, it is in fact sufficient to interpret all of them as instances of a LOAD action. However, these words obvi-ously differ in meaning, and in a different domain we may need to make more specific distinctions between them. To retain those distinctions, word senses are treated as leaves of :sem (phys-obj (form object) (origin artifact) the semantic hierarchy. The complete LF type of a word is written as as LF-parent*LF-form, where the LF-parent is the type defined in the LF ontology (for example, LF Filling forload), and the LF-form is the canonical form associated with the word (for example, LF Filling*load for all morphological :sem (phys-obj (origin natural) (object-function comestible) For every word sense in the lexicon, we specify syntactic features (such as agreement, morphology, etc.), LF type, andthe subcategorization frame and syntax-semantics mappings.
To illustrate, consider the verb load in the sense to fill the container. The lexicon entry for load linked to LF Filling :sem (phys-obj (origin human) (intentional +) is shown in Figure 5. It contains two possible mappings be- tween the syntax and the LF: one in which the THEME isrealized as direct object, corresponding to Load the oranges Figure 3: LF type definitions for some objects in the LF on- into the truck, and another in which the THEME is realized as prepositional complement, corresponding to Load the truckwith oranges. In this figure, SUBJ denotes the restrictionon the subject of the verb, DOBJ the restriction on the di- While the distinctions we encode with feature vectors can rect object, and COMP3 the restriction on the prepositional be represented in a multiple inheritance hierarchy, a feature- complement. SEM denotes the selectional restrictions prop- based representation makes it easy to implement an effi- agated from the definition in Figure 4 based on role corre- cient type-matching algorithm based on [Miller and Schubert, spondences. These restrictions are used as follows: Given 1988]. More importantly, using feature vectors allows us to the restriction on the THEME argument in Figure 4, only ob- easily modify the semantic information associated with a lexi- jects marked as (mobility movable) are accepted as a direct cal entry, a property utilized during the customization process object or prepositional with complement of load. Finally, the described below. More detailed description of our feature set SUBJ-MAP, DOBJ-MAP and COMP3-MAP denote the se- can be found in [Dzikovska et al., to appear].
mantic arguments into which these syntactic arguments will The semantic features encode basic meaning components that we use in selectional restrictions on verb arguments. To The parser produces a flattened and unscoped logical form illustrate, consider Figure 4, which shows definitions for a using reified events [Davidson, 1967]. A simplified repre- certain class of words that specify filling events that involve sentation showing the semantic content of Load the oranges containers. Intuitively, LF Filling defines a motion event in into the truck is shown in Figure 6. The representation has which an intentional being (AGENT) loads a movable ob- ject (THEME) into another physical object that can serve asa container (GOAL). The restriction on the agent argument specifies that the entity filling this role must be an intentionalbeing.
1This is consistent with the FrameNet interpretation, where all of In the LF ontology, many similar words can be mapped to these words are instances of the Filling frame.
(LF LF Filling*load)(SUBJ (NP (SEM (phys-obj (intentional +))))) (DOBJ (NP (SEM (phys-obj (mobility movable))))) (COMP3 (PP (ptype into) (SEM (phys-obj (container +)))))) (SUBJ (NP (sem (phys-obj (intentional +))))) (DOBJ (NP (SEM (phys-obj (container +))))) (COMP3 (PP (ptype with) (SEM (phys-obj (mobility movable)))))) Figure 5: Lexicon definitions for the verb load used in (a) load the oranges into the truck (b) load the truck with oranges.
(F e123 LF Filling*load :agent pro1 :theme v1 :goal v2) (IMPRO pro1 LF Person :context-rel +YOU+) Figure 6: The LF representation of the sentence load the or- :sem (situation (cause agentive)):arguments tion to KR representation needs to accomplish two things: it has to convert the domain-independent LF-representationsyntax into the syntax required by the KR, and it has to Figure 4: LF type definitions for LF Motion and LF Filling.
map the general ontological concepts in the LF (LF types) In the lexicon, feature vectors from LF arguments are used to into the domain-specific ontological concepts in the KR (KR generate selectional restrictions based on mappings between types). We illustrate our method by showing how the LF rep- subcategorization frames and LF arguments.
resentations can be transformed into two different knowledgerepresentation languages: the knowledge representation cur-rently used by the TRIPS planner (TRIPS Planner Language),and the KM knowledge representation language [Clark and fier coming from a noun phrase specifier, or a quantifier First we illustrate how the LF representation syntax can be denoting the function of the variable, e.g., “SPEECHACT” converted into the syntax required by either the Trips Planner for speech acts, “F” for forms derived from verb and adver- Language or the KM language. For simplicity, these exam- bial predicates, “IMPRO” for implicit pronouns, such as the ples assume that both of these languages use a definition of implicit subject of an imperative. Note that for simplicity, the the LOAD action with ACTOR, CARGO and VEHICLE slots representation shown there is a simplified version that omits as specified in the Pacifica ontology (described in section 2).
some of the reference and discourse information and focuseson the semantic content of an utterance.
Consider the example sentence Load the oranges into the truck. The target forms that we want to sent to the Planner From general semantic representation to
for both KR languages are given in Figure 8: (a) is the rep-resentation in the TRIPS Planner Language, and (b) is the domain-specific KR
representation in the KM language. Our task is to transform To produce domain-specific KR representations from the the LF representation for this sentence (shown in Figure 6) generic LF representations, we developed a method to cus- into these target representations. The focus of this first exam- tomize parser output. Our method transforms the interme- ple is how we obtain the different syntactic representations diate semantic representation produced by the parser (the required by the different KR languages. However, note that LF representation) into the specific KR representation used the first step in the transformation process is for reference res- by the reasoners. The transformation from LF representa- olution to replace all variables in definite descriptions in the (a) (define-lf-to-kr-transform load-transform-trips :pattern ((?spec ?ev LF Filling :agent ?a :theme ?t :goal ?g) − > ((TYPE ?ev LOAD) (actor ?ev ?a) (cargo ?ev ?t) (container ?ev ?g)))) (b) (define-lf-to-kr-transform load-transform-km :pattern ((?spec ?ev LF Filling :agent ?a :theme ?t :goal ?g) − > (?ev is (a LOAD with (actor ?a) (cargo ?t) (container ?g))))) Figure 7: Transforms between the domain-independent form and domain-dependent forms (a) in TRIPS planner language (b)in the KM language Figure 9: The representation of a MOVE event as a MOVE frame with an embedded PATH frame in the KM language.
Figure 8: The KR representation for Load the oranges into tology. LF Motion is a high-level LF ontology concept that the truck in (a) the TRIPS planner language (b) the KM lan- covers a large class of motion words (motion, move, trans- port, walk, etc.) In the example above, the system will applythe path transform to any descendants of LF Motion unless amore specific transform exists.
LF representation with constants denoting the entities they Conversely, we can define transforms that use adverbial refer to in the domain. In this example, v1 is replaced by OR- modifiers to fill slots in a single frame. For example, we ANGES2 and v2 is replaced by TRUCK3. The details of this can define a transform that converts Go straight to Pittsford, substitution process are beyond the scope of this paper.
where straight is analyzed as modifying to in the LF represen- In order to transform the domain-independent LF represen- tation and creates a MOVE frame with the slot value (Manner tation in Figure 6 into the domain-specific KR representations DIRECT), if this is the form supported by the system reason- in Figure 8, we define mapping rules that specify the pattern that the LF must match, and the corresponding KR represen- In our examples so far, all arguments are optional. For ex- tation. The relevant transform for our example is shown in ample, in Figure 10, we have LF From-loc, LF To-loc and Figure 7. It will map all LF representations using LF type LF Via as semantic arguments for a MOVE event. These are LF Filling into instances of the LOAD concept. This trans- not always overtly expressed in language, and if not all of form specifies that the ACTOR slot of LOAD will be filled them are present in the LF representation, the transform still with the variable from the AGENT argument of LF Filling applies. However, there are some cases in which an argu- (?a), the CARGO slot with variable from the THEME argu- ment must be present for a transform to apply. In such cases, ment (?t) , and the CONTAINER slot with the variable from we specify preconditions on the required arguments. For ex- ample, the TRIPS Pacifica ontology distinguishes between In this example, there is a direct correspondence between MOVE, a request to move vehicles without any additional the arguments of the LF type and the slots of a KR type.
planning required, and TRANSPORT, a request to move car- A more complex case involves transforming a single LF gos, as described in section 2. With this representation, it form into a set of frames or predicates in the target KR lan- only makes sense to interpret motion verbs as instances of guage. For example, our LF representations treat all path TRANSPORT if there is an identifiable cargo present. We adverbials (LF To-loc, LF From-loc, etc.) as separate com- achieve this with the transform in Figure 11, which specifies ponents, which we have found to be the most convenient that the THEME semantic argument that fills the CARGO slot domain-independent implementation for capturing their syn- is required for the transform to apply.
tactic variability. However, for a transportation domain, it is To make transform writing more efficient, we use the lex- more convenient to collect all adverbials into a single path ical (orthographic) form of words in transforms. For exam- frame, because path descriptions are essential for routing.
ple, from the point of view of the LF ontology, medication The transform shown in Figure 10 maps all path adverbials names have similar distributions across syntactic contexts, that modify an LF Motion event into a single path frame in and therefore are represented as leaves under the LF DRUG the KM language. The result of the application is a set of type, e.g., LF DRUG*prozac, LF DRUG*aspirin. However, frames (or a frame with another embedded frame) shown in the KR ontology in the medication adviser domain makes pragmatic distinctions between medications (e.g., prescrip- The transforms use the hierarchical properties of the LF on- tion vs. over-the-counter), but uses medication names as leaf (define-lf-to-kr-transform path-transform-km :pattern ((?spec ?ev LF Motion :lf from-loc ?fl :lf to-loc ?tl :lf via ?va) − > ((?ev is (a MOVE with (path *1))) (*1 is (a PATH with (source ?fl) (destination ?tl) (mid-point ?va)))) Figure 10: A transform to collect all path adverbials into a single path frame in the KM language. *1 denotes an operation ofcreating a new variable for the path frame.
(define-lf-to-kr-transform transport-transform-trips :preconditions ((:obligatory theme)):pattern ((?spec ?ev LF Motion :agent ?a :theme ?t :instrument ?vh) − > ((TYPE ?ev LOAD) (actor ?ev ?a) (cargo ?ev ?t) (vehicle ?ev ?vh)))) Figure 11: The transform for TRANSPORT actions in TRIPS Pacifica domain. The precondition limits the transform applica-tion only to the cases when a cargo slot can be filled.
(define-lf-to-kr-transform drug-transform-medadvisor :pattern ((?spec ?en (:* LF DRUG ?lf-form)) Figure 12: A transform for medications in the medication ad- Figure 13: The logical forms and the resulting KR forms af- viser system, using the TRIPS planner language.
ter applying the transform in Figure 12 (a) when the typeASPIRIN is defined in the KR ontology (b) when the typeZOLOFT is not defined in the KR ontology.
types in the KR hierarchy. We can write a single transformfor all LF DRUG children, shown in Figure 12, that convertsthe LF type to the KR type based on the lexical form specified We pre-process every lexical entry by determining all pos- in the entry. Note that since the transform allows a variable sible transforms that apply to its LF type. For each transform (?lf-form) to determine the resulting KR type, it also requires that applies, a new sense definition is created that is identical the user to specify a default value that will be assigned if the to the old definition but contains a new feature in the semantic class with the name derived from ?lf-form cannot be found in vector, kr-type, with the value of the KR ontology class spec- ified in the transform. Thus, we obtain a (possibly larger) For instance, assume a KR ontology in which the medica- specialized set of lexical entries that specify the KR class to tion ASPIRIN is defined, but the medication ZOLOFT is not which they belong. We then propagate type information from defined. Then given the transform in Figure 12 and the logi- the LF representation into the syntactic arguments, which cre- cal form in Figure 13(a), we obtain an entity of KR type AS- ates tighter selectional restrictions in the lexicon. Finally, we PIRIN, since ASPIRIN is defined in the KR ontology. How- increase the preference values for the senses for which map- ever, since the concept ZOLOFT is not defined in the KR pings were found, so domain-specific entries will be tried first ontology, we obtain an entity of KR type SUBSTANCE, as shown in Figure 13(b), because SUBSTANCE was specified We illustrate the lexicon specialization process with the as the default value in the transform.
verb load. Given the definitions of load and LF Filling inFigure 4, and the transform definitions in Figure 7, the algo-rithm to generate the lexical entry for the verb load proceeds Specializing the lexicon
In addition to using the transforms described above to convert • Fetch the definition of LF Filling and the semantic fea- LF representations to KR representations, we also use them in a pre-processing stage to specialize the lexicon. By inte- • Determine the applicable transform, in this case load- grating the domain-specific semantic information into the lex- icon and grammar, we increase parsing speed and improvessemantic disambiguation accuracy.
• Add KR-type load to the semantic feature vector of load, 2Using the lexical form directly is the simplest possible case, • Lexicon queries the ontology about selectional restric- since it assumes that the KR will use the lexical forms directly as tions on arguments, and determines that the element that concepts. An easy extension is determining the KR type based on fills the CARGO slot needs kr-type COMMODITY; some operator applied to a lexical form, for example, concatenatingit with a pre-defined prefix used in the KR ontology, e.g., MED- • Add KR-type COMMODITY to the semantic feature vec- (LF LF Filling*load)
(SUBJ (NP (SEM (phys-obj (intentional +) (kr-type Person) (origin human) (form solid-object))))
(DOBJ (NP (SEM (phys-obj (mobility movable) (kr-type commodity))))) (COMP3 (PP (ptype into) (SEM (phys-obj (container +) (kr-type Vehicle) (origin artifact) (form object)))))
(LF LF Filling*load)
(SUBJ (NP (sem (phys-obj (intentional +) (kr-type Person) (origin human) (form solid-object))))
(DOBJ (NP (SEM (phys-obj (container +) (kr-type Vehicle) (origin artifact) (form object)))))
(COMP3 (PP (ptype with) (SEM (phys-obj (mobility movable) (kr-type Commodity) ))))
Figure 14: The lexicon entry for the verb load specialized for Pacifica domain (a) for load the oranges into the truck (b) forload the truck with oranges • Apply similar transforms to the rest of the arguments.
(a) (origin human) => (form solid-object) This process creates a new definition of load, shown in Figure 14, that has 2 entries corresponding to the same two senses described in section 3, but with stricter selectional re-strictions on the arguments as the result of domain specializa- Figure 15: Some feature inference rules in the TRIPS lexicon.
tion. Since this specialized lexical entry for load conforms to (a) The domain-independent inference rule in the generic lex- KR requirements, now suitable objects or prepositional com- icon; (b) The domain-dependent inference rule defined in the plements of load must be not only movable, as specified in the LF ontology, but also identified as belonging to the classCOMMODITY in the domain. Comparable transforms apply structions with specialized entries. This makes the system to the relevant nouns, so oranges, people and other cargos more robust for out of domain utterances, because it allows will have a kr-type value that is a subtype of COMMODITY us to find a parse and respond more intelligently even when inserted in their semantic feature vectors.
out of domain words are present. At the same time, using The entries are further specialized with the use of feature feature specialization rules for domain-specific entries allows inference rules. Feature inference is a general mechanism us further restrict search space even for unspecialized entries in the system lexicon used to express dependencies between through tighter restrictions on argument slots. For example, feature values. For example, we have a feature rule associ- the domain-independent entry for load allows any containers ated with our feature system declaring that if something is to fill the GOAL argument of the loading action. The special- marked as a human (origin human), it is also a solid object ized entry restricts the GOAL argument to vehicles. Thus, the (form solid-object), in (Figure 15(a)). During the specializa- verb phrase load the dispenser will be accepted only if there tion process we add the rules that declare dependencies be- is no other interpretation available, because while dispenser tween the values of kr-type features and the values of domain- does not have a specialized entry in Pacifica, it is not marked independent features. For example, in our Pacifica domain we with the domain-independent value (object-function vehicle) have a rule declaring that if something is marked as (kr-type and therefore does not satisfy the restriction.
Person), then it must also have the domain-independent fea- Lexicon specialization considerably speeds up the pars- ture value (origin human), shown in Figure 15(b). When that ing process. We conducted an evaluation comparing pars- value is added to the feature vector in the subject restriction ing speed and accuracy on two sets of 50-best speech lattices in the entry for load, it will trigger the domain-independent produced by our speech recognizer: 34 sentences in the med- rule in 15(a), which causes (form solid-object) to be added to ical domain and 200 sentences in the Pacifica domain. Ta- the feature set. The new features added as the result of spe- ble 2 provides some statistics on the lexicons used in these cialization and feature inference are highlighted with bold in domains. The results presented in Table 3 show that lexi- con specialization considerably increases parsing speed and Our lexicon specialization process is designed to easily in- improves disambiguation accuracy. The times represent the tegrate the specialized and non-specialized entries. If no spe- average parsing time per lattice, and the errors are the num- cialization is found for a lexical entry, it remains in the lexi- ber of cases in which the parser selected the incorrect word con, though with a lower preference, with the assumption that sequence out of the alternatives in the lattice.3 kr-type is assigned an undefined value kr-root, which will sat-isfy any kr-type restriction. It can then participate in any con- 3For the purpose of this evaluation, we considered correct tional restrictions propagated into the entries will effectively select the correct definitions. Table 2 shows that the number of lexicon entries increased from 1947 to 2048 (5%) in our Pacifica domain, and from 1947 to 1954 (0.3%) in our Medi-cation Advisor domain.
In spite of the increased number of lexical entries, we be- lieve that our approach offers significant benefits with respect to portability. It is difficult to quantify the effort of porting a system to a new domain, but as a first approximation, we list • Define a domain model and the KR ontology. This
comes for free in a dialogue system, since the reasoning • Define lexical entries for previously unseen words
Table 3: Average parsing time per lattice in seconds and sen- and add new type to the LF ontology as needed. The
tence error rate for our specialized grammar compared to our number of words that need to be added diminishes as the generic grammar. Numbers in parentheses denote total time lexicon and the LF ontology grow. We are also working on automatic methods to obtain new words from sourcessuch as WordNet and FrameNet.
The amount of work involved in domain customization is • Modify the grammar as needed to cover new syn-
relatively small. The generic lexicon and grammar stay es- tactic constructs. We already have a wide-coverage
sentially the same across domains, and a KR ontology must grammar that can deal with a variety of constructs be defined for the use of back-end reasoners anyway. We need such as conjunction, long-distance dependencies, rela- to write the transforms to connect the LF and KR ontologies, tive clauses and a variety of other phenomena common but as their number is small compared to the total number of in speech. So the development of the grammar is a sense entries in the lexicon and the number of words needed one-time investment that requires diminishing amount of in every domain, this represents an improvement over hand- crafting custom lexicons for every domain.
• Define the LF-KR transforms. This is the most time-
consuming part of the process. However, the size of the Discussion
transform set is proportional in size to the size of the do- In previous work, we have relied on the built-in ontology sup- main, which is considerably smaller than the set of all port in the TRIPS system to provide ontology definitions, lexical entries in the domain, and convenience features subtype unification and other information necessary in the such as using the LF hierarchical structure and the lex- customization process [Dzikovska et al., to appear]. The ical forms help speed up the process considerably. In built-in ontology facilitates efficient processing, but it lim- addition, the separation between the linguistic and do- its the expressiveness of the language to the forms supported main ontologies allows developers to write mappings in in the TRIPS knowledge representation (first-order logic and semantic terms, using descriptive labels assigned to se- frames with simple slots). Using an external ontology such mantic arguments such as AGENT, THEME, etc., at a as KM allows us to utilize the richer representations that sup- level of abstraction that avoids addressing the details of port reasoning, but comes at a performance cost of having to the grammar and subcategorization frames such as those call the external ontology during parsing, which can poten- in COMLEX. They can also take advantage of the hi- tially be expensive. However, with caching the results of the erarchical structure in the domain-independent ontology previous calls, or pre-compiling some of the information, we and write mappings that cover large classes of words.
believe it should not significantly impair system performance.
• Define feature inference rules. This is not required, but
In lexicon specialization, there is a tradeoff between the it is very helpful in improving parsing speed, especially strength of the selectional restrictions and the lexicon size.
when parsing speech inputs where out of domain words The number of distinct lexical entries may increase in the can be present. In the specialization process, we actively specialized lexicon because there is no one-to-one correspon- utilize the hierarchical structure of the KR ontology - a dence between the LF and KR ontologies, so several trans- rule defined for a parent is inherited by all its children, forms may apply to the same LF type depending on the se- so the rules need only be defined for a relatively small mantic arguments that are filled. A new entry is created for every transform that applies, and during parsing the selec- Note that the first three items in the list above (defining a domain grammar, defining new words and defining new choices in which a different pronoun, article or tense form were sub-stituted, e.g., can I tell my doctor and could I tell my doctor were grammar rules) need to be done in any system which is being considered equivalent. However, equally grammatical substitutions employed in a new domain. The advantage of our approach of a different word sense, e.g., drive the people vs. get the people is that instead of trying to change and adapt the lexical entries and the rules from the previous domain to suit the language in the new domain, we can re-use the old lexicon and define [Allen et al., 2001] James Allen, Donna Byron, Myroslava LF-KR transforms and inference rules, the number of which Dzikovska, George Ferguson, Lucian Galescu, and is an order of magnitude smaller than the number of lexical Amanda Stent. Towards conversational human-computer interaction. AI Magazine, 22(4):27–38, 2001.
Another advantage of using a domain-independent se- [Clark and Porter, 1999] P. Clark and B. Porter. KM (1.4): mantinc representation that is later mapped to a domain- Users Manual. http://www.cs.utexas.edu/users/mfkb/km, specific form is that it makes it easier to design domain- independent reference resolution and discourse management components. Many current ontologies do not have explicit Davidson, 1967] Donald Davidson. The logical form of ac- support for encoding information needed for reference and tion sentences. In Nicholas Rescher, editor, The Logic dialogue processing, such as distinguishing between definite of Decision and Action, pages 81–95. University of Pitts- and indefinite descriptions. In our system, we use the LF rep- resentations generated by the parser to encode this informa- Davidson, Essays on Actions and Events, Oxford Univer- tion, which guarantees that the reference module can get it encoded consistently regardless of the specific language used [Dzikovska et al., to appear] Myroslava by the back-end. During reference resolution, these descrip- Mary D. Swift, and James F. Allen. Customizing mean- tions are replaced with references to constants or sets known ing: building domain-specific semantic representations to the back-end reasoners, thus insulating them from the de- from a generic lexicon. In Harry Bunt, editor, Computing tails necessary for reference resolution and discourse process- Meaning, Volume 3, Studies in Linguistics and Philosophy.
Kluwer Academic Publishers, to appear.
[Ferguson et al., 2002] G.M. Ferguson, J.F. Allen, N.J. Blay- Conclusion
lock, D.K. Byron, N.W. Chambers, M.O. Dzikovska,L. Galescu, X. Shen, R.S. Swier, and M.D. Swift. The Our development of spoken dialogue systems in multiple do- medication advisor project: Preliminary report. Techni- mains has highlighted the need for a method of adapting a cal Report 766, Computer Science Dept., University of broad-coverage parser to different domain applications that allows us to preserve general language information common across domains and also produce interpretations customized Goddeau et al., 1994] D. Goddeau, E. Brill, J. Glass, to the needs of each domain. We have developed the method C. Pao, M. Phillips, J. Polifroni, S. Seneff, and described here to meet this need. By maintaining two on- Galaxy: A human-language interface to on- tologies, one that is generalized for language representation, and another that is customized to the domain, and defining 707–710, Yokohama, Japan, September 1994.
mappings between them, we can preserve our broad-coverage http://www.sls.lcs.mit.edu/ps/SLSps/icslp94/galaxy.ps.
language components across domains as well as produce se- [Gurevych et al., 2003] Iryna Gurevych, Stefan Merten, and mantic representations that are optimally suited to the domain Robert Porzel. Automatic creation of interface specifica- tions from ontologies. In Proceedings of the HLT-NAACLWorkshop on The Software Engineering and Architecture Acknowledgments
of Language Technology Systems (SEALTS), May 2003.
This material is based upon work supported by the Office of Naval Research under grant number N00014-01-1-1015 and semantic and syntactic coding of predicate-argument the Defense Advanced Research Projects Agency under grant In Proceedings ANLP-NAACL 2000, Seattle, number F30602-98-2-0133. Any opinions, findings, conclu- sions or recommendations expressed in this material are thoseof the author(s) and do not necessarily reflect the views of [Macleod et al., 1994] Catherine Macleod, Ralph Grishman, on Sharable Natural Language Resources, Nara, August References
[Allen et al., 1996] James F. Allen, Bradford W. Miller, Eric K. Ringger, and Teresa Sikorski. A robust system for natural spoken dialogue. In Proceedings of the 1996 general reasoning. In Tom M. Smith, Reid G.; Mitchell, Annual Meeting of the Association for Computational Lin- editor, Proceedings of the 7th National Conference on Artificial Intelligence, pages 161–165, St. Paul, MN, [Allen et al., 2000] James Allen, Donna Byron, Myroslava Dzikovska, George Ferguson, Lucian Galescu, and [Ros´e, 2000] Carolyn Ros´e. A framework for robust seman- tic interpretation. In Proceedings 1st Meeting of the North NLENG: Natural Language Engineering, Cam- American Chapter of the Association for Computational bridge University Press, 6(3):1–16, 2000.
[Seneff, 1992] Stephanie Seneff. TINA: A natural language system for spoken language applications. ComputationalLinguistics, 18(1):61–86, March 1992.

Source: http://www.ida.liu.se/labs/nlplab/ijcai-ws-03/papers/dzikovska.pdf