Next: 3. Applied Laboratory Work (Post-1988) Up: About the Natural Language Laboratory Previous: 1. About Natural Language


2. Theoretical Laboratory Work (Post-1988)

2.1. Levels of Language

2.1.1. Morphology

Pollard and Sag (1987, 1994) have proposed a program for organizing the lexicon hierarchically by lexical types. Lexical items inherit information which is not specific to them by according to their lexical type. In addition, relationships among words (for example, that between the 3rd person singular of the verb and its base) are expressed by lexical rules. Lexical rules may add information or transform information.

Paul McFetridge is initiating a project in information-based morphology to explore these proposals with particular focus on how lexical rules may be constrained and implementing a lexicon which is elaborated as needed rather than completely compiled when loaded.

2.1.2. Syntax

Much of the laboratory's research in syntax aims to integrate computational and linguistic formalisms and ideas. The majority of research concerns unification-based grammars which are a family of grammars whose central operation is unification. Two descriptions can unify as long as they do not contain conflicting information. The result contains all the information that is in either description.

Unification-based grammars are declarative and lexical. Their declarative nature makes them relatively easy to understand and modify, since a grammar states relationships instead of giving procedures for their computation. Their lexical nature reduces grammar development to lexicon development because syntactic classes contain information about the classes with which they combine, which means that only a small number of grammar rules and principles are needed, and that those rules and principles are schematic and fixed. However the drawback is large lexicons with complex lexical entries. To avoid redundancy and to capture linguistic generalizations in such lexicons, lexical types and lexical rules have been introduced. Lexical types eliminate "vertical redundancy" by factoring out general information shared by whole syntactic classes, organizing lexical entries into multiple hierarchies based on the information they share, and inheriting down general information. Lexical rules eliminate "horizontal redundancy" by identifying lexical entries that have common, specific patterns which cannot be inherited via lexical types. In such cases, a lexical rule fills in the specific information.

2.1.2.1. Head-Driven Phrase Structure Grammar

Head-driven phrase structure grammar (Pollard & Sag, 1987, 1994) is one of the best known unification-based formalisms. A distinctive feature of the grammar is that attribute-value matrices are used to represent lexical entries and grammar rules and principles. A second distinctive feature is its head-driven nature: the head constituent of a phrase is a central notion. Otherwise, the grammar borrows freely from previous unification-based grammars. The treatment of syntactic categories, syntactic features, and some of the principles are from generalized phrase structure grammar (Gazdar et al., 1985). A number of the lexical rules are similar to those in lexical-functional grammar (Kaplan & Bresnan, 1982).

Chart parsing implementations of head-driven phrase structure grammar have been developed in Lisp by Paul McFetridge (McFetridge & Cercone, 1990) and in Prolog by Fred Popowich and Carl Vogel (1990, 1991a). Chart parsing is a type of parsing in which all syntactic structures which are built are placed on a single graph structure called a chart. A successful parse of a sentence is a list of edges and nodes that includes the start node of the chart, all the words in the sentence, and the end node.

2.1.2.2. Tree Unification Grammar

Tree unification grammar (Popowich, 1989a, 1989b, 1993) resembles head-driven phrase structure grammar in certain respects. Both formalisms use signs as information structures and provide facilities for establishing the relationships between signs. Tree unification grammar incorporates a greater degree of lexicalization, using one rule as opposed to the four or five rules described in Pollard and Sag (1987, 1994). Head-driven phrase structure grammar also relies on the use of universal grammar principles which apply in conjunction with the grammar rules. Many of these principles make use of functionally dependent values: they contain calls to relatively complex functions like collect-indices, selectively-combine-semantics, order-constituents and bip (application of the binding inheritance principle). In tree unification grammar, the relationships between the information contained in different signs is stated in terms of unification, or very simple list operations, within the lexical templates.

Tree unification grammar also differs from head-driven phrase structure grammar in the kind of derivational structures that it allows. In tree unification grammar the structures are binary, consisting of a specified functor and argument. For the structures used by head-driven phrase structure grammar, which are called phrasal signs, there is no theoretical limit on the branching factor. In the phrasal signs of head-driven phrase structure grammar, it is very awkward to describe specific relationships relevant to constraints on reflexive pronouns. Tree unification grammar specifies three different reflexive relationships contained in its R-GPR, R-PR and R-NPR lexical templates (Popowich, 1989a). The different relationships between semantic and reflexive information can be captured more elegantly in the formulation of head-driven phrase structure grammar introduced in Pollard (1989), but it is still not as concise as the tree unification grammar characterization that we have described.

A parser has been written that can use tree unification grammar in an NLP system (Popowich, 1989b). The prototype for this system will form the basis of a larger system for developing and testing natural language grammars that can be used in a natural language interfaces. The larger system has been developed by Susan Kindersley (Popowich & Kindersley, 1991).

2.1.2.3. Parsing as Inheritance Reasoning

Vogel and Popowich (1990) have proposed that parsing with a unification-based grammar can be viewed as reasoning over an inheritance network. This research is of a system-building nature in comparison to the related research of Nebel and Smolka (1989) who explore the relationship between terminological reasoning systems and unification grammars with respect to the computational properties of computing subsumption in either system.

The aim of the research is to explore the extent of the relationship which begins with a number of similarities observed between unification-based grammar and inheritance reasoning. First, unification and inheritance reasoning are closely related: the monotonic unification operator can directly model strict inheritance, and extended unification algorithms can be used to describe defeasible or mixed strict and defeasible inheritance (Bouma, 1990).

Second, lexical types and inheritance reasoning are closely related: lexical types allow common information to be shared by lexical entries, while inheritance has frequently been used in NLP to allow the sharing of information in the lexicon (Flickinger et al., 1985; Evans & Gazdar, 1990). Lexical information is organized in an inheritance hierarchy and the information associated with a specific lexical entry is determined by inheritance over this hierarchy. Information relating to grammar rules and principles can also be captured in a similar manner (Pollard, 1985; Fraser & Hudson, 1990). The construction of an inheritance hierarchy allows the concise statement of linguistic information, be it lexical, phonological, morphological, syntactic or semantic.

Third, inheritance need not be restricted to descriptive hierarchies such as a hierarchy of lexical types, but can also be used to establish relationships in derivational structures such as parse trees. Inheritance can thus be used in parsing and in generation.

The result of this research is a pedagogical tool for describing the structure and application of head-driven phrase structure grammar and, produced as an artifact, a cleanly specified if inefficient alternative parsing strategy. More efficient parsers for head-driven phrase structure grammar do exist: Proudian and Pollard (1985) developed a chart parser for an early version of the formalism, and there are McFetridge's and Popowich's chart parsers too. A nice side effect of the inheritance reasoner is that it provides an elegant way to represent and process unknown words.

2.1.2.4. Static Discontinuity Grammar

Veronica Dahl, Director of the Logical and Functional Programming Group at SFU, has developed static discontinuity grammar (Dahl, 1986, 1989), which belongs to the family of discontinuous grammars also developed by Dahl, which are in turn a type of logic grammar. Logic grammars differ from traditional grammars in several respects including the form of their grammar symbols, which may include arguments representing trees or semantic interpretations, and the existence of processors based on specialized theorem-provers which give grammar rules a procedural aspect (Abramson & Dahl, 1989).

Discontinuous grammars can deal with many language phenomena involving discontinuity (i.e., the relating of constituents which may be widely separated in a sentence or phrase), for example unbounded dependency (see section 2) and free word order.

The significance of static discontinuity grammar is that it transparently turns high level linguistic specifications into executable code. Such specifications can include two of the major tenets of modern linguistics: discontinuity, and the modular constraining of discontinuous constituents through various linguistic principles and filters, as in government-binding theory. Popowich has developed a left-corner parser of static discontinuity grammar. Dahl and Popowich (1990) describes the parser, parsing efficiency, and the tree admissibility interpretation of static discontinuity grammars.

Scattered Context Grammars (Greibach & Hopcroft, 1969) are a specialization of SDGs in which "atoms" are used as grammar symbols instead of logical "terms". Popowich (1994a) introduces a chart parsing algorithm for Scattered Context Grammars.

2.1.2.5. Government-Binding Theory

Government-binding theory is a model of grammar descended from extended standard theory (Chomsky, 1982, 1986). Most of the descriptive research into the formal syntax of the different languages of the world is within the government-binding formalism.

Dahl has worked on merging government-binding theory into static discontinuity grammar (Dahl, 1988). The merging of one theory within the other is possible because the two have certain similarities: ``both attempt to achieve maximal coverage with minimal expressive machinery, both can describe fairly free types of movement, both tend toward abstraction and generalization'' (Abramson & Dahl, 1989, p. 153).

Dahl, Popowich and Rochemont (1993) are conducting further research into a static discontinuity grammar-based implementation of aspects of government-binding theory.

2.1.3. Semantics

The Natural Language Laboratory has done work on the semantics of natural language from the viewpoint of analysis, representation and generation. Fass has worked on the analysis of metaphor and metonymy, Cercone has developed a representation for discourse concepts, and Pattabhiraman has examined the role of salience in natural language generation.

2.1.3.1. A Representation for NLP Using Extended Semantic Networks

Cercone et al's (1992) extended semantic network representation can represent states, events, actions, logical and natural language quantifiers, expressible intentionality, modalities, adverbials, comparatives (explicit and implicit), and the meanings of complex concepts such as walking. The concept representation is hierarchical and recursive rather than dependent upon primitives. Formal methods for interpretation and manipulation of the network constructs have been developed and implemented. Complex concepts are efficiently processed. Lambda-abstracted predications are uniformly handled in the representation so that sentences such as ``loving one's neighbours is a virtue'' are understood.

Work has concentrated on two main problems: modifiers, and inferencing and reasoning strategies. Iterated modifier analysis is used advantageously to account for comparative and descriptive modifiers. In an implementation, we have refined our treatment of English modifiers.

A means of promoting inferences is being developed through the use of supra-organizations which will be imposed on the network database to allow rapid access to facts contextually pertinent to concepts.

2.1.3.2. A Stratified Meaning Representation for Natural Language Discourse

The stratified model, developed by Strzalkowski (1986), uses a lambda-categorial language for meaning representation. The model applies a series of disambiguating transformations to a discourse fragment before assigning it a final representation. The transformations include morphological analysis, lexical ambiguity resolution, syntactic parsing, computation of extra-sentential dependencies, and pragmatic evaluation in discourse context. The computation of extra-sentential dependencies is performed on an intermediate, formal representation of the discourse fragment. The transformations are done using Montague's (Thomason, 1974) style of coupling syntactic and semantic processing but not his use of an intensional logic the intermediate representation used is a lambda-categorial language (a typed predicate calculus language with a lambda-operator). The final representation of discourse content, while not decided upon, will have a form resembling that of "abstract situation" from Barwise and Perry's (1983) situation semantics. The model has been used to investigate two discourse problems: first, the resolution of anaphoric and cataphoric reference across sentences (Strzalkowski & Cercone, 1986), and second, the representation and manipulation of non-singular concepts (Strzalkowski & Cercone, 1989). Presently, there is no implementation of the stratified model.

This work has so far been restricted to cases of anaphora in two sentence stories. In these stories, cases have been selected where a reference, if it can be computed at all, has a unique antecedent. A set of rules has been formalized for computing the references in these cases, and the transformation which encompasses these rules, and the rules themselves, have been related to the stratified model.

Cross-sentence cataphoric references have also been studied and some solutions specified for certain cases. Another discourse problem that has been studied using the same methods is repairing references in discourse. An example of reference repair is given in the following three sentence discourse: ``I bought a black coffee in the pub. I brought it back to my office to drink. The cream in my office was off and it (the coffee) tasted bad.''

This work (Strzalkowski & Cercone, 1989) generalizes the treatment of the concepts given in Strzalkowski and Cercone (1986). Non-singular concepts are abstract entities that embrace a variety of smaller or larger collections of instances. They are usually referred to using bare plural noun phrases (such as `birds', `alligators' and `presidents'), definite singular noun phrases with a generic interpretation (``the alligator,'' ``the president''), mass nouns (`water', `gold'), and functional uses of definite descriptions (``the number of students,'' ``the temperature'').

Non-singular definite descriptions appear in a number of reasoning puzzles, for example, the following syllogism adapted from Partee (1972):
The president is elected every four years.
The president is Bush.
Thus, Bush is elected every four years.

This erroneous conclusion arises because we assign a non-singular interpretation to ``The president'' instead of a singular one.

Strzalkowski and Cercone's solution is a representation based on a partially ordered structure of levels in which the objects of the same relative singularity are assigned to the same level. Their choice of representation has been motivated by the following main concerns:
(1) the representation should systematically distinguish between those language terms that are used to refer to objects of different singularity, that is, those classified within different but related levels of the model;
(2) the representation should capture certain types of inter-sentential dependencies in discourse, most notably anaphoric-type cohesive links;
(3) finally, the representation should serve as a basis for defining a formal semantics of discourse paragraphs that would allow for capturing the exact truth conditions of sentences involving non-singular terms, and for computing inter-level inferences.

Some initial progress has been made on (1) and (2) but (3) is currently under investigation. Strzalkowski and Cercone believe that their approach promotes computational feasibility, because they avoid the identification of general terms like `temperature', `water', etc., with intensions, that is, functions over possible worlds. In their theory, the concept of non-singularity has a local (often subjective) character.

2.1.3.3. Metaphor and Metonymy Interpretation

Non-literal language, of which metaphor and metonymy are types, includes phenomena whose meaning cannot be obtained by direct composition of their constituent words (see Fass et al., 1991, 1992a). The relationship between metaphor and metonymy has been much disputed (Fass, 1988a).

Fass (1988b, 1991a, 1991b, 1994, 1997) has developed a computational approach for distinguishing selected examples of metonymy from metaphor and from literalness and anomaly in short English sentences. Consider, for example, a waitress in a restaurant saying ``The ham sandwich is waiting for his check.'' This sentence contains a metonymy while ``The car drank gasoline'' contains a metaphor. The difference between a metonymy and a metaphor, according to Lakoff and Johnson (1980), is that in metonymy one entity stands for another whereas in metaphor one entity is viewed as resembling another. This difference can be seen in the above sentences: ``The ham sandwich'' stands for the male person who ordered the ham sandwich, while ``The car'' is viewed as resembling an animate drinker in that both use up a liquid, take in the liquid through an aperture, etc.

In Fass's approach, literalness is distinguished because it satisfies contextual constraints that the non-literal others all violate. Metonymy is discriminated from metaphor and anomaly in a way that supports Lakoff and Johnson's (1980) view of metonymy and metaphor, permits chains of metonymies (Reddy, 1979), and allows metonymies to co-occur with instances of either literalness, metaphor, or anomaly. Metaphor is distinguished from anomaly because the former contains a relevant analogy, unlike the latter.

The approach to metaphor and metonymy is part of collative semantics (Fass, 1988c), a semantics for natural language processing, and has been implemented in a computer program called meta5. Collative semantics is also concerned with the representation and resolution of lexical ambiguity. The meta5 program contains a lexicon of 500 word senses, a small grammar, and routines containing collative semantics.

2.1.3.4. Salience in Language Generation

The role of salience in decisions in natural language generation has been examined. The salience of an entity in natural language generation represents its prominence. It is therefore used in several decisions in natural language generation as a measure of preference that justifies the choices committed. Salience is a multi-aspect notion that includes vividness and topicality. The interactions among salience-influenced decisions can be characterized as synergistic, competitive, complementary, asynchronous, concurrent and weighted. The problems involved in making these interactions operational in natural language generation systems are being examined, in particular, from the viewpoint of efficient, maximally-deterministic generation. Computational mechanisms are being developed to capture the influences of salience in natural language generation decisions such as content selection, lexical selection, tactical linearization and the generation of pronouns (Pattabhiraman & Cercone, 1990a, 1991a, 1992b).

Pattabhiraman is also interested in the problem of evaluating theories and systems in natural language generation (Pattabhiraman & Cercone, 1990b) and is currently working on developing a coherent classification of natural language generation systems, and spelling out a set of evaluation criteria that are relevant to each kind. The classification of natural language generation systems is based on a number of dimensions such as objectives, stages of evolution and approaches to the modelling of linguistic knowledge.

2.1.4. Pragmatics

SFU's work in pragmatics has been concerned with dialogue between a natural language interface and a user when one of them has faulty or missing information, specifically, how to keep the dialogue going and how to repair the dialogue. This work can also be viewed as introducing robustness into NLP systems because they attempt to keep a dialogue going in situations where it has broken down. The main problem in this work, which has been overlooked by many other researchers, is fault finding from the dialogue: is the faulty or missing information in the database of the natural language interface or with the user?

2.1.4.1. Predicting and Explaining Query Failure

Database systems don't contain all the information needed to model their target domain. Users' questions usually contain assumptions about the domain. A null answer arises when or more assumptions of a question are incorrect. For example, ``which graduate students have taught CMPT681?'' (example due to Bonnie Webber) assumes that graduate students can teach that course or ``which text book is used for CMPT898?'' assumes that CMPT898 is a course on which a text book is used. A null response -- "no" -- is not very helpful. What is needed is more informative responses.

Joseph has developed a way of predicting null responses to database queries in a way that avoids wasteful search of secondary memory and a way of generating meaningful explanations of why no result was found (Joseph & Aleliunas, 1991). A distinction is made between two main kinds of null value: ``property is inapplicable'' as in a male person's maiden name (which violates an obvious semantic constraint on male persons), and ``value at present unknown.'' This research developed from Kao's work (Kao et al., 1988; Cercone et al., 1990) which analyzed the reasons for the null responses once the database access resulted in a no response.

Joseph uses an inference procedure to detect empty responses to queries and to give a reason why a response was empty. The inference procedure uses three sources of information: the query itself (an expression in relational calculus), the RM/T* catalogue which describes the structural and functional dependencies of the actual database, and a knowledge base whose rules are semantic constraints on the database. There is a general rule of inference for each operation of relational calculus (e.g., join, projection, selection). As soon as the inference procedure determines that one of the attribute-domain sets is empty, further database searching is unnecessary and the query must return an empty response. The inference procedure's actions are strictly controlled by the structure of the query.

2.1.4.2. A Belief-Based View of Ill-Formed Input

Ill-formed input refers to ``lexical problems such as misspelled words, sentential problems such as missing words or phrases and bad word order, semantic problems such as anomalous or self contradictory sentences, and contextual problems such as incoherent requests or continuations'' (Allen, 1983). Users of natural language interfaces produce ill-formed input which must be handled by the interfaces. Perhaps the biggest problem with ill-formed input is locating it. We have tackled this problem by treating ill-formed input as a clash of beliefs between a user and an interface: the ill-formed input is then viewed as the false or incorrect beliefs of either the user or interface. To see how this works, consider what happens when a natural language interface parses a sentence containing the item `recieve'. `Recieve' is not in the interface's lexicon, hence there is a conflict of belief. There are two obvious causes of the conflict: first, `recieve' is a spelling error by the user (i.e., the user has false beliefs about the spelling of `receive'), and second, `recieve' is an unknown word to the interface (i.e., the interface has missing beliefs about the word `recieve'). This central idea about clashing beliefs has been incorporated into a framework we have developed for treating ill-formed input in which an interface and its users each has a ``belief model'' about itself and other participants with which it engages in dialogue (Fass et al., 1990; Fass & Hall, 1991). Each model contains beliefs of that participant about its language and its knowledge of the domain. The clashes of beliefs lead to a far more extensive classification of situation in which ill-formed input occurs than previous approaches to ill-formed input.

2.1.5. Language as a Medium

Users prefer different media at different stages of learning to use a system and for different applications. Naive users learn to use new systems faster with natural language, while experienced users prefer control keys because they are faster (Napier et al., 1989). Graphical displays may be preferred for certain tasks, e.g., clicking on a region of a map to get information about that region. What is needed are user interfaces that can handle different media and different applications.

Chris Groeneboer and Nick Cercone (1991) have developed a modular architecture for a multi-modal, multi-application user interface. The architecture has five layers consisting of: user, modalities, the interface itself, applications, and the underlying operating system. Requests and responses undergo transformations as they pass through the layers. User requests pass down through the layers to the appropriate application. System responses pass up through the layers to the user in an appropriate modality for presentation. The layered architecture is designed to make changes easy, for example, adding a new modality, application or user. Control resides with the interface itself which contains a set of experts: a modality expert, an applications expert, a representation expert, an interaction expert, and a coordinator which organizes the other experts. Each expert has access to two knowledge bases specific to its expertise, one that contains general domain-independent information, and another that contains specific domain-dependent information. For example, the interaction expert has a domain-independent knowledge base about human-computer interaction in general while the domain-dependent knowledge base contains user models which in turn contain information about specific users.

Groeneboer and Cercone have also developed an underlying philosophy of interface design based on the constructivist paradigm. Within this paradigm, knowledge is viewed as constructed by a user to fit with the user's experiences of the world. Attention is paid to the processes by which users organize their experience. The significance of this paradigm for interface design is that attention is paid to how users construct the interface environments they want.

2.2. Particular Languages

2.2.1. Chinese

Xing Pue and Paul McFetridge (1995, 1996, 1997) have done research on Chinese linguistics, particular complementation in Chinese.

2.2.2. English

Natural language interfaces for English and machine translation systems for English-Spanish have been developed.

2.2.3. German

Computer assisted language learning programs have been developed for German (Heift & McFetridge, 1993; Heift, 1998).

2.2.4. Spanish

Machine translation systems for English-Spanish have been developed.

Computer assisted language learning programs are being developed for Spanish (McFetridge & Heift, 1994).



Next: 3. Applied Laboratory Work (Post-1988) Up: About the Natural Language Laboratory Previous: 1. About Natural Language


  • Last modified 24 July 1998 (Dan Fass <fass@cs.sfu.ca>)