nl-kr-request@CS.ROCHESTER.EDU (NL-KR Moderator Brad Miller) (03/30/88)
NL-KR Digest (3/29/88 19:58:34) Volume 4 Number 36 Today's Topics: perfect language Re: Pro-Drop Languages Recent Memoranda in Computer and Cognitive Science Submissions: NL-KR@CS.ROCHESTER.EDU Requests, policy: NL-KR-REQUEST@CS.ROCHESTER.EDU ---------------------------------------------------------------------- Date: Fri, 11 Mar 88 19:17 EST From: HESTVIK%BRANDEIS.BITNET@MITVMA.MIT.EDU Subject: perfect language What is really the policy of NL-KR? I must really wonder after I saw the latest issue. It had a posting from a guy who asked about information about how a perfect language might look like. What is the meaning with such a question? It just doesn't make sense. It is as absurd as to ask for a perfect example of a fish. What makes a fish perfect? It's size? (It must be average in size compared to all other fish.) It's color? It must have a grey color with a few spots, perhaps? I really think the NL-KR moderator should try to keep the level of discussion at a certain level, in order to keep people minimally interested in continuing reading it. After all, reading e-mail DOES consume time. Arild Hestvik ------------------------------ Date: Fri, 25 Mar 88 09:02 EST From: HESTVIK%BRANDEIS.BITNET@MITVMA.MIT.EDU Subject: Re: perfect language *** [From miller@cs.rochester.edu, in reference to why the moderator didn't delete the request about perfect language] ..what the guy wanted to know was what features in a language make it good or bad, e.g. easy to communicate ideas, or whatever. I don't know the answer to this, but that doesn't mean the question is unanswerable, or that it shouldn't be thought about. The result may well relate to designing better computer languages, or international languages, like Esparanto. *** OK - I see. Well then let me answer it. In fact, why don't you include it in NL-KR? NOTHING in a language makes it easier or more difficult to communicate, express meanings etc etc. All languages are EQUAL in this respect (cf. socio-linguistic investigations of Black English in the 60's). The fact that some languages have vocabularies that others don't have will of course make it difficult to talk about computer architecture in Warlbiri or about different types of snow in English. But this is just a fact related to the lexicon of that language. Any language can freely expand its lexicon to cover any thinkable concept understandable by humans (cf. Rosch color-term research (1973)). Human languages do NOT have the design that would make them efficient from the point of comparison with artifical languages. Artificial languages are designed to be easy to parse, write syntax and semantics for. For example, in propositional logic, vacuous quantification is OK: Ax(P(y)) is a syntactically wellformed expression, and receive a perfect semantic interpretation. However, something similar in natural language is BAD: *Who did Mary hit Bill. Natural languages are in fact very clumsily designed from the point of view of formal languages. The rules and principles involved are computationally intractable. It is extremely dubious that insights into natural language can help design better computer languages. As for lingua franca's, like Esperanto, my comment is: Who cares? We already have one or two international languages: English being one of them. ------------------------------ Date: Wed, 23 Mar 88 13:13 EST From: Peter Whitelock <pete@aiva.ed.ac.uk> Subject: Re: Pro-Drop Languages Shouldn't somebody mention the distinction between pro-drop languages and zero anaphora languages? In the former, only subjects can be dropped, and this seems to correlate pretty well with the presence of subject agreement morphemes. In the latter, any argument of a verb can be dropped, and this seems to correlate pretty well with what's usually called Free Word Order, but is better called Free Constituent Order (FCO) i.e. clause bounded scrambling. So does anyone know of a language which has FCO but not zero anaphora, or which has fixed constituent order but allows say, null object anaphors? Elisabet Engdahl suggested that Brazilian Portuguese might be one of the latter. Afrikaans and even South African English allows zero prepositional objects - is this true in other Germanic languages? Pete Whitelock ------------------------------ Date: Fri, 25 Mar 88 11:03 EST From: Robert Rubinoff <rubinoff@linc.cis.upenn.edu> Subject: Re: Pro-Drop Languages In article <302@aiva.ed.ac.uk> pete@uk.ac.ed.aiva (Peter Whitelock,F SB x226E,,) writes: >Constituent Order (FCO) i.e. clause bounded scrambling. So does anyone >know of a language which has FCO but not zero anaphora, Well, Latin has free word/constituent order but is only "pro-drop"; it definitely doesn't allow omission of non-subject elements. And I think that Warlpiri, which has almost complete free word order, isn't even pro-drop. What does seem plausible is that a language can be pro-drop only if there is a way to figure out what the subject was. This doesn't seem like a constraint on possible grammars, just a constraint on which ones would be useful for communication. In a language that doesn't indicate the subject by the verb inflection, people would stop omitting the subject because they would be misunderstood too often. (And then there would be so few sentences without subjects that people would start using them even when they're not needed, e.g. in sentences like "it's raining".) In languages with zero-anaphora, like Japanese, there must be some general constraints that allow people to figure out the referent of the "omitted" phrases, so you don't need inflection on the verb to recover the subject. Or perhaps Japanese speakers just accept more vagueness in their communication; after all "someone hit Bill" is a perfectly good sentence in English, even though we may not have any idea who the someone is. If this is correct, then what you wouldn't expect to find is a language that has null subjects but requires all other NPs to be present and doesn't indicate the subject by inflection on the verb. Robert ------------------------------ Date: Fri, 25 Mar 88 15:18 EST From: Gordon Fitch <gcf@actnyc.UUCP> Subject: Re: Pro-Drop Languages In article <302@aiva.ed.ac.uk> pete@uk.ac.ed.aiva (Peter Whitelock,F SB x226E,,) writes: } .... Afrikaans and even South African English allows zero prepositional }objects - is this true in other Germanic languages? } I thought prepositions without objects were just reverting to their original role as adverbs. As: I went up the stairs. I went up. As I recall in Latin and Greek* poetry even normally bound prepositional prefixes are occasionally found separated, with an adverbial function. *Classical ------------------------------ Date: Fri, 25 Mar 88 23:43 EST From: Jay Sekora <sekora-jay@CS.YALE.EDU> Subject: Re: Pro-Drop Languages In article <302@aiva.ed.ac.uk> pete@uk.ac.ed.aiva (Peter Whitelock,F SB x226E,,) writes: >...has fixed constituent order but allows say, null object anaphors? >Elisabet Engdahl suggested that Brazilian Portuguese might be one of the >latter. Afrikaans and even South African English allows zero prepositional >objects - is this true in other Germanic languages? > >Pete Whitelock Well, English, for one. My ex-girlfriend (from rural Illinois) says things like "Would you like to come with?" That's the only example I can remember at the moment, but she seemed to be able to do that with a large number of prepositions. -jay --- Jay Sekora sekora-jay@yale.UUCP sekjaya@yalevm.bitnet ------------------------------ Date: Mon, 28 Mar 88 18:18 EST From: Rich Wales <wales@maui.cs.ucla.edu> Subject: Re: Pro-Drop Languages Hungarian frequently drops direct objects, as well as subjects. Verbs in Hungarian are conjugated, not only according to person/number of the subject, but also according to whether or not there is a "defi- nite" direct object. (Roughly speaking, a "definite" direct object is a proper name, or a noun with a definite article or a demonstrative. First- and second- person pronouns are *not* treated as "definite"; however, the *polite* second-person pronouns are in fact third-person forms, and *are* con- sidered "definite".) A definite direct object will generally be omitted entirely, rather than being referenced via a pronoun, since the "definite" verb form already suffices to show that the object exists. Some confusion is still possi- ble, though -- since the "definite" direct object could be singular or plural, and could be either "third-person" or a polite "second-person". There is also a special verb form used where the subject is first person singular, and the direct object is second person (singular or plural). This particular form is not generally encountered except in conversation with family members or close friends, though, since the "polite" second- person pronouns are treated as third-person forms. -- Rich Wales // UCLA CS Dept // wales@CS.UCLA.EDU // +1 (213) 825-5683 3531 Boelter Hall // Los Angeles, California 90024-1596 // USA ...!(ucbvax,rutgers)!ucla-cs!wales ...!uunet!cs.ucla.edu!wales "Sir, there is a multilegged creature crawling on your shoulder." ------------------------------ Date: Tue, 29 Mar 88 09:06 EST From: Rob Bernardo <rob@pbhyf.PacBell.COM> Subject: Re: Pro-Drop Languages In this discussion about languages that allow the absence of certain noun phrase constituents of a clause, we have been considering the inflection of the verb as substitute for such absent noun phrases and we have been considering the vagueness or ambiguity that results. But some posters have been talking about the resulting vagueness wrt the listener being able to figure out what **pronoun** would have been there. Here's an example: In article <10706@shemp.CS.UCLA.EDU> wales@maui.UUCP (Rich Wales) writes: +Hungarian frequently drops direct objects, as well as subjects. +Verbs in Hungarian are conjugated, not only according to person/number +of the subject, but also according to whether or not there is a "defi- +nite" direct object. +... +A definite direct object will generally be omitted entirely, rather than +being referenced via a pronoun, since the "definite" verb form already +suffices to show that the object exists. Some confusion is still possi- +ble, though -- since the "definite" direct object could be singular or +plural, and could be either "third-person" or a polite "second-person". This misses the issue. This issue is not whether the listener could deduce what **pronoun** would have gone there, but rather whether the listener could deduce the **referent** of this would-be noun phrase is. In the example above, even if the verb had different inflections for the "number" and "person" as well as "definiteness" of the direct object, or even if there was a direct object pronoun, the listener still might not be able to figure out the referent of the (implicit or explicit) direct object. Verb inflections, pronouns and even full noun phrases capture some information about the referent that the speaker assumes to be sufficient for the listener to deduce the referent. I think many of us are being led astray by generative grammar. We start thinking in terms of how some underlying structure gets converted to the structure of the uttered sentence and we start thinking that the listener must reverse this process to understand the sentence. I suppose some people believe that, but, since it is not obvious, it would behoove the poster to state his/her assumptions in this case. -- Rob Bernardo [backbone]!pacbell!rob [backbone]!ptsfa!rob rob@PacBell.COM residence: (415) 827-4301 (R Bar JB, Concord, CA) business: (415) 823-2417 (Pacific Bell, San Ramon, CA) ------------------------------ Date: Mon, 21 Mar 88 01:36 EST From: yorick@nmsu.csnet Subject: Recent Memoranda in Computer and Cognitive Science .ds CH .nr PS 9 .ps 9 .nr VS 11 .vs 11 .ce 8 .ps 11 \fBRecent Memoranda in Computer and Cognitive Science\fR .ps For copies of the technical reports listed below write to: Memoranda Series Computing Research Laboratory Box 30001 New Mexico State University Las Cruces, New Mexico 88003 USA .fi Fass, D. (1987), Collative Semantics: Lexical Ambiguity Resolution and Semantic Relations (with Particular Reference to Metonymy), MCCS-86-59. An account of coherence is proposed which attempts to clarify the relationship between semantic relation and the resolution of lexical ambiguity. The account of coherence, semantic relations, and lexical ambiguity resolution is embodied in Collactive Semantics, which is a domain-independent semantics for natural language processing. A natural language program called meta5 uses CS; an example is given of how meta5 discriminates semantic relations and the account of coherence presented are expansions and improvements of descriptions published elsewhere (Fass 1987, to appear). Much of the paper explores the relationship of metonymy to literal, metaphorical, and anomalous semantic relations, and its role in lexical ambiguity resolution. The paper also explores the definition of coherence presented, which is that coherence is the synergism of knowledge, where synergism is the interaction of two or more discrete agencies to achieve an effect of which none is individually capable. Mc Kevitt, P. (1987), Natural Language Interfaces in Computer Aided Instruction - What Happened Before and After the 80s AICAI "Coup", MCCS-86-66. There are many advantages and problems in using computers for teaching. Hopefully Artificial Intelligence (AI) will provide solutions to Computer Aided Instruction (CAI) problems. It is evident from the history of CAI that early tutoring systems lacked intelligent interfaces. AI did not play a significant role in CAI until the Artificially Intelligent Computer Aided Instruction (AICAI) "coup" in the eighties. It is important to discover why this happened and the implications of uniting AI and education. Moreover, Natural Language Interfaces (NLIs)carve their own niche in CAI and if AI is to be applied to education, then Computational Natural Language Processing (CNLP) is necessarily a part of that application. The Hard Core Natural Language (HCNL) problems of today must be solved if we are to build robust interfaces. Natural language dialogue and belief systems must be integrated effectively to overcome HCNL. Mc Kevitt, P. (1987), Object and Action Frames for Transfer Semantics, MCCS-86-69. It is a characteristic of computer operating systems that they contain actions or commands which transfer objects such as files and directories from one state to another. In formalizing the domain of operating systems we should build representations of actions which circumscribe the transfer of objects in the system. Transfer Semantics is a knowledge scheme that embodies such representations. Knowledge structures called object frames are used to represent numerous objects. Action frames describe the effects of actions in terms of preconditions, postconditions, actions and actors. Preconditions denote possible or preferred sets of objects that an action will affect. Postconditions relate the state of object sets after an action has occurred. Actions include the particular actions that cause transfer. An actor is the person (or user) who performs some action. The power of Transfer Semantics lies in the inference rules that manipulate action frames. It is applied to the UNIX and TOPS-20 operating systems in a developing program called OS Consultant. OS Consultant will be used by new users to learn operating system concepts. McKevitt, P. (1986), Formalization in an English interface to a UNIX database, MCCS-86-73. When building knowledge representation schemes for particular domains we should realize abstractions of relationships from those domains. Computer operating systems involve numerous actions of commands which can transfer data from one state to another. It is this process of transfer that should be formalized in any representation of such actions. Transfer Semantics already exists as a knowledge representation scheme for operating system commands. Yet, no such scheme is powerful without inferencing. Axiomatic semantic techniques have been applied in exploring the logical foundations of computer programming. We can use axiomatic semantics as a language for constructing abstract formalizations of inference rules for Transfer Semantics. In particular, commands in operating systems such as UNIX* can be formalized by this method. Ballim, A. (1987), A Proposed Language for Representing and Reasoning about Nested Beliefs, MCCS-87-77. An outline of a language for reasoning with nested beliefs is proposed. The language is an extension of standard first order logic, incorporating the notion of a \fIviewpoint\fR (a collection of beliefs about a topic that are held by some person) The syntax, semantics, special operators and rules of inference for a \fIthree-valued\fR language are introduced. An extension of this language is proposed that allows individual viewpoints to be considered. An example of reasoning using the extended language is shown. Slator, B. & Wilks, Y. (1987), Towards Semantic Structures from Dictionary Entries, [i[CRL[xi], MCCS-87-96. Dictionary definitions from a large machine-readable source (Longman's Dictionary of Contemporary English), are converted into semantic structures (a frame-based knowledge representation), suitable for knowledge based parsing. Besides exploiting the special encodings contained in the dictionary, the text of the definition entry itself is analyzed, to extract genus and differentia terms. This additional information further enriches the semantic structures. The output of the program is a lexicon of word-sense frames, with each frame explicitly or implicitly positioned in multiple, pre-existing, hierarchies. These frames constitute a text-specific knowledge source for use by a Preference Semantics parser for text. Pollack, J. (1987), On Connectionist Models Of Natural Language Processing, MCCS-87-100. The interpretation of natural language requires the cooperative application of many systems of knowledge, both language specific knowledge about word use, word order and phrase structure, and ``real-world'' knowledge about stereo-typical situations, events, roles, contexts, and so on. And even though these knowledge systems are nearly decomposable, enabling the circumscription of individual knowledge areas for study, this discomposibility does not easily extend into the realm of computation; that is, one cannot construct a psychologically realistic natural language processor by merely conjoining various knowledge-specific processing modules serially or hierarchically. Part of this thesis describes efforts in building a natural language processing system with modular knowledge sources but strongly interactive processing. Language intepretation takes place on an activation network which is dynamically created from input, recent context, and long-term knowledge. Initially ambiguous and unstable, the network settles on a single intepretation, using a parallel, analog, relaxation process. This work has become associated with a research paradigm undergoing a strong resurgence of interest: Connectionism, which seeks to study computational cognition on neurally-inspired models. While there have been several other efforts at applying connectionism to natural language understanding, all have suffered from severe limitations in computational power. A major portion of this thesis, therefore, is dedicated to understanding and overcoming these limitations. It is argued that these limitations result from some of the underlying assumptions of connectionism, and that dynamic reconfiguration of these networks is necessary. The construction of a Turing Machine out of connectionist primitives demonstrates the efficiency, and thus practical power, of multiplicative connections. As a bridge between the state of a connectionist system and the weights that determine its pattern of connectivity, multiplicative connections are one method of achieving dynamic reconfiguration. Connectionist systems with multiplicative connections have been proposed, but are usually abandoned for lack of a method to control them. Accordingly, we have developed a learning procedure to adjust weights in Cascaded Networks, which are connectionist networks with constrained multiplicative connections. Wilks, Y., Fass, D., Guo, C., McDonald, J., Plate, T. & Slator, B. (1987), A Tractable Machine Dictionary as a Resource for Computational Semantics [i]CRL[xi], MCCS-87-105. This paper distinguishes and then investigates the merits of the position in computational semantics that the semantic structure of language text and of knowledge representations share common organising principles. Better understanding of such principles may come from analysing the semantic structure and estracting the semantic information from dictionaries, a particular kind of text. Dictionaries have particular promise because (a) the semantic structure of text may be more exposed in them than in other forms of text and (b) many are now in machine-readable form and are amenable to analysis by large-scale computational methods. Some convergence is identified between the view of computational semantics presented, computational lexicography, and knowledge acquisition in terms of common issues and problems they share. This convergence is illustrated using some work by the CRL natural language group that attempts to extract semantic information form a particular dictionary, the Longman Dictionary of Contemprorary English, and use that semantic information in two kinds of computational semantics that reflect the general position on computational semantics set forth. Plate, T. (1987), A design for the simulation of connectionist models on coarse grained parallel computer, [i]CRL[xi], MCCS 87-106. The simulation of connectionist models on coarse-grained, loosely coupled MIMD (Multiple Instruction Multiple Data) parallel computers is discussed. It is argued that the issues involved lead naturally to the choice of a small set of design principles, which are then presented. A parallel simulator based on these design principles can efficiently simulate connectionist models that conform to the commonly accepted `rules' of connectionist models. If these design principles are followed, then the connection scheme of the machine relatively unimportant, but the hypercube scheme is the most conventional and flexible. A working general-purpose connectionist model simulator based on these principles, called Hycon, has been built. A function expressing the speedup achieved by using a parallel machine is given, demonstrating that the proposed design principles scale was well to medium sized parallel machines. It is shown how the design principles could be simply extended to the simulation of connectionist models on larger machines. Fass, Dan. (1987), Collative Semantics: An Overview of the Current Meta5 Program, [i]CRL[xi], MCCS 87-112. This paper describes the construction and operation of the current version of a natural language program called meta5. The program contains an implementation of Collactive Semantics (CS), a recently developed domain-independent semantics for natural language processing. CS investigates the interplay between lexical ambiguity and semantic relations, i.e., the interrelations between the representation and resolution of lexical ambiguity on the one hand and the discrimination and representation of seven different kinds of semantic relation (literal, metonymic, metaphorical, anamalous, redundant, inconsistent, and novel) on the other. The program is written in Prolog and consists of a lexicon of 460 word senses, a small grammar, and semantic routines which embody the two processing components of CS, called `collation' and `screening'. Meta5 processes single sentences that contain lexical ambiguity, discriminates and represents the semantic relations between pairs of word senses (during which the program actually recognizes metaphors, metonymies, etc.), and resolve the lexical ambiguity in those sentences. Barnden, J. (1987), Avoiding Some Unwarranted Entailments Between Nested Attitude Reports, MCCS-87-113. The avoidance of unwarranted entailments among propositional attitudes or among propositional attitude reports has been subject to much study, usually under the heading of `opacity'. However, thereis a particular type of unwarranted entailment that has been given relatively little attention, although it arises in various different formal methods for treating attitudes. Intuitively, the unwarranted entailments result from a formal treatment imputing its own explications of predicational aspects of attitudes to the agents holding the attitudes. The problem is especially subtle and difficult to address in the case of nested attitudes (beliefs about beliefs, and so on), where what is explicated is itself an attitude, and the explications are in terms of the arcane theoretical concepts used by the formal treatment. The present paper extends and refines a previously published analysis of the problem, and also proposes a formal approach that appears to avoid difficulties. The approach is entirely first-order and extensional. It is based almost entirely on formal devices that have been proposed by other authors. The crux is the use of an abstractional device to `package' the scheme's own explications in such a way that they are not imputed to attitude-holding agents. The approach uses logic terms that denote agents' representational objects or templates derived from those objects, although it embodies as few commitments as possible to the detailed nature of such objects and templates. The approach points to the need for greater attention to the variety of readings that can be applied to the predicational aspects of attitude reports as opposed to their potentially-referring phrases, for validation of approaches on nested-attitude cases, and for greater sensitivity to natural language utterances that themselves mention features of a formal approach under study. ------------------------------ End of NL-KR Digest *******************