- Email: [email protected]
Natural language in the Common Model of Cognition
Available online at www.sciencedirect.com Available online at www.sciencedirect.com
ScienceDirect ScienceDirect
Procedia online Computer 00 (2019) 000–000 Available at Science www.sciencedirect.com Procedia Computer Science 00 (2019) 000–000
www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia
ScienceDirect Procedia Computer Science 145 (2018) 699–709
Postproceedings of the 9th Annual International Conference on Biologically Inspired Cognitive Postproceedings of the 9thBICA Annual International Conference Inspired Cognitive Architectures, 2018 (Ninth Annual MeetingonofBiologically the BICA Society) Architectures, BICA 2018 (Ninth Annual Meeting of the BICA Society)
Natural language in the Common Model of Cognition Natural language in the Common Model of Cognition Philip C. Jackson, Jr. * Philip C. Jackson, Jr. * 0F
0F
TalaMind LLC, PMB #363, 55 E. Long Lake Rd., Troy, MI 48085, U.S.A. TalaMind LLC, PMB #363, 55 E. Long Lake Rd., Troy, MI 48085, U.S.A.
Abstract Abstract This paper discusses how natural language could be supported in further developing the Common Model of Cognition following the approach toward human-level artificial intelligence. These thoughts arethe presented as Model answersoftoCognition questionsfollowing posed by ThisTalaMind paper discusses how natural language could be supported in further developing Common Peter Lindes, Paul Rosenbloom, and reviewers of thisintelligence. paper, followed a description of the TalaMind system, the TalaMind approach toward human-level artificial Theseby thoughts are presented as answersdemonstration to questions posed by and general discussion of theoretical and strategic issues for the Common of Cognition. Petera Lindes, Paul Rosenbloom, and reviewers of this paper, followed by aModel description of the TalaMind demonstration system, and a general discussion of theoretical and strategic issues for the Common Model of Cognition. © 2019 The Authors. Published by Elsevier B.V. © 2019 2018 The Authors. Published by Elsevier Elsevier B.V. This is an open accessPublished article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) © The Authors. by B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the the CC scientific committee of the 9th Annual International Conference on Biologically Inspired This is an open access article under BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the scientific committee of the 9th Annual International Conference on Biologically Inspired Cognitive Peer-review under responsibility of the scientific committee of the 9th Annual International Conference on Biologically Inspired Cognitive Architectures. Architectures. Cognitive Architectures. Keywords: natural language, language of thought, mentalese, TalaMind, Common Model of Cognition Keywords: natural language, language of thought, mentalese, TalaMind, Common Model of Cognition
1. Questions to consider 1. Questions to consider In an email thread for the ‘Language’ subgroup of the Common Model of Cognition, Peter Lindes [23] wrote: In an email thread for the ‘Language’ subgroup of the Common Model of Cognition, Peter Lindes [23] wrote: “I think we should be thinking about what are the architectural implications of language processing as one “I think weofshould be thinking about what areinvolve the architectural implications of language processing as one application the Common Model. This might questions like the following: application of the Common Model. This might involve questions like the following: How much of language processing is part of perception and how much is done in other parts of the model? How much of language processing is part of perception and how much is done in other parts of the model?
* Corresponding author. address:author. [email protected] * E-mail Corresponding URL: http://www.talamind.com E-mail address: [email protected] URL: http://www.talamind.com 1877-0509 © 2019 The Authors. Published by Elsevier B.V. This is an open access under the CC by BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) 1877-0509 © 2019 The article Authors. Published Elsevier B.V. Peer-review under responsibility of the committee of the(https://creativecommons.org/licenses/by-nc-nd/4.0/) 9th Annual International Conference on Biologically Inspired Cognitive This is an open access article under thescientific CC BY-NC-ND license Architectures. Peer-review under responsibility of the scientific committee of the 9th Annual International Conference on Biologically Inspired Cognitive Architectures.
1877-0509 © 2018 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the scientific committee of the 9th Annual International Conference on Biologically Inspired Cognitive Architectures. 10.1016/j.procs.2018.11.051
700 2
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
How are words and their meanings stored and retrieved? How are words and their meanings assembled into meaningful interpretations of sentences? What architectural mechanisms are needed for this process? How can linguistic knowledge be represented and retrieved by the architecture? How can language processing become an acquired skill that is performed rapidly and without the agent being able to describe the details of its processing? How can situational context bias retrievals so that appropriate senses of words are retrieved? What architectural mechanisms can allow processing to become a skill in procedural memory and still allow context to bias retrievals?” Lindes suggested papers answering these questions be submitted to the 2018 Fall Symposium on the Common Model of Cognition. In addition, Paul Rosenbloom [29] suggested the following questions: “(1) Which consensus aspects of man-like language processing directly map onto the common model as currently specified; (2) which would suggest straightforward extensions; and (3) which provide major challenges.” Linde’s and Rosenbloom’s questions are of interest to me, relevant to my previous paper [15] about the Common Model, and relevant to the research approach I advocate toward human-level AI [14]. So, the following pages give my answers to these questions, and to some additional questions asked by anonymous reviewers of a draft version of this paper. Throughout this paper, in referring to the Common Model of Cognition (CM), I refer to the description given by Laird, Lebiere, and Rosenbloom [21] of what was previously called the Standard Model of the Mind. I will abbreviate this reference as LLR. 2. Language and the Common Model of Cognition To begin, some general remarks: LLR hypothesize that natural language processing may be implemented using knowledge and skills plus the cognitive architecture of the Common Model. In effect, it appears natural language would be a potential application developed using the standard components of CM (which are working memory, procedural memory, declarative long-term memory, perception, and motor components). LLR express a “strong commitment” that no additional architectural modules are necessary for language processing, though some architectural primitives (such as a phonological loop) can be included. There are implicit languages involved in CM: An implementation of CM presumably has a symbolic language for specifying production rules which can be applied in a cognitive cycle. There are also presumably symbolic languages for specifying items of information in declarative memory or working memory, and a symbolic language for specifying motor actions that can be performed in an external environment. These symbolic languages are essentially formal languages of some sort. My understanding is that they are not envisioned in CM as having the syntactic or semantic complexity of natural language. In this respect, CM would maintain the tradition of AI research that has been followed for decades, which is to attempt to achieve AI using formal languages, and attempt to support natural language processing as an application. The TalaMind thesis [14] advocates a very different direction, which involves developing an AI system using an internal language (called Tala) based on the unconstrained syntax of a natural language (English), and taking a principled approach toward supporting the unconstrained semantics of natural language. Tala is used as a symbolic language for representing information and procedures. In Tala, natural language expressions are symbolic data structures that represent natural language syntax and can refer to cognitive categories for semantics. The TalaMind cognitive cycle uses pattern-matching of Tala expressions for information and procedures. The thesis proposes an architecture called TalaMind for systems to achieve human-level AI, which is further discussed below. For concision, a system with a TalaMind architecture is called a ‘Tala agent’.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
701 3
The above statements are the key differences between the TalaMind approach and the consensus direction of cognitive architectures reflected in the initiative for a Common Model of Cognition. Indeed, it has been a consensus assumption of AI research in general, almost since its inception, that natural language should be supported indirectly as an application, using simpler formal languages to support inference and representation of semantics. The TalaMind approach advocates departing from that assumption, as I’ll explain throughout this paper. My hope is that eventually a new consensus will form to develop a future generation of the Common Model to achieve human-level AI via the TalaMind approach. Although the TalaMind approach maps onto the cognitive cycle of CM, it would be a major extension to the current structure of CM modules. Fully implementing the TalaMind approach will be a major challenge for AI researchers. I will say more, later in this paper, about why this challenge should be accepted, in discussing strategic and theoretical issues for the Common Model of Cognition. 3. Consensus views of natural language Rosenbloom [29] asked about consensus aspects of human-like language processing, and some discussion of this topic is appropriate. There are different consensi in linguistics about different topics related to natural language, as well as debates and lack of consensus about various topics. TalaMind follows a consensus perspective of research on cognitive linguistics and cognitive semantics [5]. TalaMind follows consensus views about natural language syntax for English, advocating support for the full, unconstrained syntax of English, and also advocating support for ungrammatical English expressions (§3.2.1, §3.3, §3.4). † However, TalaMind is not related to previous work in linguistics at the level of cognition corresponding to the current Common Model, e.g. research on sentence comprehension in relation to working memory [22]. The TalaMind approach does not appear to conflict with such research. Rather, it focuses on different topics. There is not a consensus view in modern linguistics about how word senses exist and should be represented – this remains an unresolved topic, philosophically and scientifically as well as technically (the writings of Peirce and Wittgenstein are still relevant). Much modern work on computational linguistics is corpus-based and does not use word senses (i.e., word meanings and definitions). Adam Kilgarriff, a respected lexicographer, wrote a paper twentyone years ago, saying he did not believe in word senses [17]. However, Kilgarriff [20] clarified his position and continued to support research on word sense disambiguation (WSD) [6]. A sub-community within computational linguistics conducts research on WSD, reported in annual SemEval workshops. The TalaMind approach subscribes to the views of cognitive semantics that word senses exist with a radial, prototypical nature; that words may develop new meanings over time, and old meanings may be deprecated; that word meanings are often metaphorical or metonymical and may involve mental spaces and conceptual blends ‡; that commonsense reasoning and encyclopedic knowledge may be needed for disambiguation, relative to situations in which words are used, and that the meanings of words and sentences may depend on the intentions of speakers. Word meanings can have natural language definitions at the linguistic level of a TalaMind architecture, and also have senses that are cognitive categories represented at the archetype level using methods such as conceptual spaces, image schemas, radial categories, etc. There is not a consensus among cognitive scientists about whether humans have an innate ‘language of thought’ (mentalese) comparable to a natural language like English. Researchers have reasoned for and against this idea (§2.2.1, §4.2.1). However, Jackendoff [12] argued that concepts must be expressed internally as sentences in a mental language: Since the brain is finite, concepts from the set possible to express by natural language must be represented internally within the mind as structures in a language. 1F
2F
†
Throughout this paper the § notation is used to refer to chapters and sections in [14]. For example, §3.2.1 refers to Chapter 3, section 2.1. These references can be directly accessed from the Table of Contents in [14] and via hyperlinks within the thesis. ‡ See [8] and §3.6.7.8, §3.6.7.9.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
702 4
Jackendoff called these structures ‘sentential concepts’. The set of concepts expressible by natural language goes beyond concepts easily expressed in propositional logic or first-order logic – it includes the full range of thoughts expressible by natural language (§2.3.1, §3.2.1, §3.3, §3.6). The expressive capabilities of natural languages should be matched by expressive capabilities of mentalese, or else by Jackendoff's argument the mentalese could not be used to represent concepts expressed in natural language: The ability to express arbitrarily large, structured sentential concepts is plausibly just as important in a mentalese as it is in English. The general-purpose ability to metaphorically tie concepts together across arbitrary, multiple domains is plausibly just as important in a mentalese as it is in English. This is not to say mentalese would have the same limitations as spoken English, or any particular spoken natural language. In mentalese, sentences could have complex structures not physically permitted in speech. This is discussed below in the section ‘Words, meanings, sentences’. 4. Cognitive plausibility Two anonymous reviewers of a previous version of this paper wrote: “Tala expressions are list structures akin to parse trees without any explicit limitations so their cognitive plausibility would need to be established.” “What would be great is if the author could actually tell us what this natural language is that he is proposing be directly supported by the architecture. I assume it can’t be a specific natural language like English, because that can’t be what all people are born with, but is it nothing more than a general tree or graph structured representation language, or does it try to capture some forms of language universals in an interesting manner?” The cognitive plausibility of a language of thought based on natural language can be supported in at least two ways, with references to cognitive linguistics and cognitive neuroscience: The first way considers Jackendoff’s argument; the second considers the phenomenon of ‘self-talk’. Jackendoff’s argument suggests (but does not prove) humans are born with an innate language of thought which enables representing any thought they express with a natural language like English, and supports learning a natural language. An ideal Common Model of Cognition would replicate and support this innate language, but we may be decades or centuries from being able to specify the innate mentalese of human infants well enough to replicate it as computational language. So, the TalaMind approach advocates that a cognitive architecture modeling human-level intelligence should support English § as an innate, internal language of thought, using Tala expressions. This is cognitively plausible by virtue of representing an ability of the human brain after it has learned a natural language. The syntax and semantics of English have been subjects for many decades of re-search, which can support endeavors to replicate them computationally. It still won’t be easy, of course. It could be that people are not born with an innate language of thought, but are just born with a very complex bioneural network (the human brain) that supports learning the full syntax and semantics of a natural language like English. In this case, Jackendoff’s argument still implies that the brain eventually achieves a state which supports internally representing sentential concepts expressible in English. So in this case, it is still cognitively plausible for a cognitive architecture modeling human-level intelligence to support representing a natural language as an internal language of thought – again the cognitive plausibility is justified by Jackendoff’s argument. Representing the syntax and semantics of what may be called a ‘natural language of thought’ is likely to be easier than representing the neural network of the human brain, for which research may continue in parallel. Mental discourse (self-talk, inner speech) is a feature people often ascribe to their minds, and a psychological phenomenon which has been remarked upon for centuries. It seems we each have the ability to mentally hear some 3F
§
Any other comparable natural language could be used, e.g. French, German, Dutch, Japanese, etc., or multiple natural languages could be supported. To simplify discussion, I refer only to English.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
703 5
of our thoughts expressed internally in natural language. This topic is discussed in section 2.1 of [3]. Baars and Gage write that inner speech is not just for verbal rehearsal but provides an individual’s “running commentary” on current issues, and is related to linguistic and semantic long-term memory. The fact that we hear inner speech suggests (but does not prove) some thoughts are represented internally in a language of thought with the expressiveness of natural language. So it supports the cognitive plausibility of a natural language mentalese. 5. Internal use of language by AI systems There is not a consensus based on analysis and discussion among AI scientists that an AI system cannot use a natural language like English as an internal language for representation and processing. Rather, this has been an assumption by AI scientists over the decades, who have used programming languages and formal logic languages like predicate calculus to support natural language processing, with very limited success: No AI system can yet understand natural language as well as a five year-old child. In 1955, McCarthy [25] proposed that his research in the Dartmouth summer project would focus on the relationship of intelligence and language. He noted that every formal language yet developed lacked important features of English, such as the ability for speakers to refer to themselves and make statements about progress in problem-solving. He proposed to create a computer language that would have properties similar to English. The artificial language would allow a computer to solve problems by making conjectures and referring to itself. Concise English sentences would have equivalent, concise sentences in the formal language. McCarthy’s proposed artificial language would allow statements about physical events and objects, and hopefully enable computers to be programmed for learning how to play games and perform tasks. Although McCarthy proposed in 1955 to develop a formal language with properties similar to English, his subsequent work did not take this direction, though it appears in some respects he continued to pursue it as a goal. Beginning in 1958 his papers concentrated on use of predicate calculus for representation and inference in AI systems, while discussing philosophical issues involving language and intelligence. McCarthy was far from alone in such efforts: AI research on natural language understanding has attempted to translate natural language into a formal language such as predicate calculus, frame-based languages, conceptual graphs, etc., and then to perform reasoning and other forms of cognitive processing with expressions in the formal language. Some approaches have constrained and ‘controlled’ natural language, so that it may more easily be translated into formal languages, database queries, etc. Since progress has been very slow in developing natural language understanding systems by translation into formal languages, Jackson [14] investigated whether it may be possible and worthwhile to perform cognitive processing directly with unconstrained natural language. The Tala language responds to McCarthy’s 1955 proposal for a formal language that corresponds to English, though probably not in the way he envisioned. It enables a TalaMind system to formulate statements about its progress in solving problems. Tala can represent complex English sentences involving self-reference and conjecture. Short English expressions have short correspondents in Tala, a property McCarthy sought for a formal language. The TalaMind approach can address theoretical questions not easily addressed by more conventional approaches. For instance, it supports reasoning in mathematical contexts, but also supports reasoning about people who have self-contradictory beliefs. Tala provides a language for reasoning with underspecification and for reasoning with sentences that have meaning yet which also have nonsensical interpretations. Tala sentences can declaratively describe recursive mutual knowledge. Tala facilitates representation and conceptual processing for learning by analogical, causal and purposive reasoning, learning by self-programming, and imagination via conceptual blends.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
704 6
6. Discussion of further questions 6.1. Language and perception The previous remarks support an answer to Lindes’ [23] question: How much of language processing is part of perception and how much is done in other parts of the model? There is certainly important processing of natural language that occurs in perception for processing speech, and also for processing visual information that can help disambiguate speech, e.g. when someone points to something to indicate what they mean. However, in the TalaMind approach much processing of natural language semantics occurs in the cognitive cycle, and involves working memory, procedural memory, and declarative long-term memory. Arguably such processing should occur in any cognitive system that attempts to understand natural language, and to create natural language expressions for communication: Cognition should be involved in reasoning to create natural language expressions – it can’t all be perception. 6.2. Words, meanings, sentences Next, I’ll discuss Lindes’ [23] questions: How are words and their meanings stored and retrieved? How can linguistic knowledge be represented and retrieved by the architecture? In the TalaMind approach words and their meanings are essentially stored in declarative and procedural memory, but there is additional structure involved. The TalaMind architecture has three levels, called the linguistic, archetype, and associative levels, adapted from Gärdenfors’ paper on inductive inference [9]. At the linguistic level, the architecture includes the Tala language, a conceptual framework for managing concepts expressed in Tala, and conceptual processes that operate on concepts in the conceptual framework to produce intelligent behaviors and new concepts. The archetype level is where cognitive categories are represented using methods such as conceptual spaces, image schemas, radial categories, etc. The associative level would typically interface with a real-world environment and supports connectionism, Bayesian processing, etc. In general, the thesis does not prescribe research choices at the archetype and associative levels. The TalaMind architecture is actually a broad class of architectures, open to further design choices at each level (§1.5, §2.2.2). The architecture is open at the three conceptual levels, e.g. permitting predicate calculus, conceptual graphs, and other symbolisms in addition to the Tala language at the linguistic level, and permitting integration across the three levels, e.g. potential use of deep neural nets at the linguistic and archetype levels. The conceptual framework at the linguistic level has additional structure, beyond that indicated by the CM terms ‘working memory’, ‘procedural memory’, and ‘declarative long-term memory’: A prototype design for a conceptual framework and conceptual processes is described below in the section about the TalaMind demonstration system. Such architectural mechanisms support addressing Lindes’ [23] questions: How are words and their meanings assembled into meaningful interpretations of sentences? What architectural mechanisms are needed for this process? How can linguistic knowledge be represented and retrieved by the architecture? If a natural language expression is heard or seen in the environment, then in the TalaMind approach there will be conceptual processes for constructing Tala expressions representing alternative syntactic and semantic interpretations of the serial word expression, to understand which interpretation is intended, and reason with the interpretation. These processes may result in asking for clarification, of course. A Tala expression is a multi-level list structure representing the dependency parse-tree (syntax) of a natural language sentence or phrase (with extensions, e.g. supporting executable concepts). When these structures are
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
705 7
created within a Tala agent, or exchanged between subagents of a Tala agent, processing them need not involve parsing English sentences as linear strings of symbols. Such internal processing also need not involve disambiguation: Tala expressions can include pointers to word senses and referents, to support internal processing. Pointers can also be used to represent re-entrant links and lattices in Tala expressions (§5.3.9.1). The TalaMind approach relies on word senses being represented either via natural language definitions, or via concepts represented using methods of cognitive semantics such as conceptual spaces, image schemas, etc. The approach is also open at the TalaMind architecture’s associative level to other methods from computational linguistics for representing and disambiguating word senses, viz. [11], [19]. Further discussion of the nature of meaning and representation is given in §2.2.2, §3.6, §4.2.8, §5.4. 6.3. Language acquisition Lindes’ [23] next question was: How can language processing become an acquired skill that is performed rapidly and without the agent being able to describe the details of its processing? The TalaMind approach does not directly address this question, since natural language representation and processing are built into the architecture of a cognitive system, as a feature of its cognitive cycle. In the TalaMind approach, an agent could describe the details of its processing (though it might have to refer to its associative level as a black box). Some discussion is given in the thesis (§4.2.4) about the human inability to describe how we understand natural language, e.g. in the context of solving Searle’s Chinese Room problem. This is an interesting topic for future research, related to artificial consciousness while learning and developing new executable concepts: We are aware of our physical actions when we learn to ride a bicycle, but later these physical actions become subconscious and automatic. Such a process could be involved in learning a new natural language. 6.4. Situational Context Lindes [23] asked two related questions: How can situational context bias retrievals so that appropriate senses of words are retrieved? What architectural mechanisms can allow processing to become a skill in procedural memory and still allow context to bias retrievals? I think the general answer to this is to represent subject domains ** in a way that supports access to interrelated word senses. If the words and phrases in a situational context indicate that a specific subject domain is involved, then the subject domain could help guide retrieval of appropriate word senses. Methods that have been developed for corpus-based computational linguistics might be applied, viz. [2], [11]. This is a topic for future research, of course. 4F
7. TalaMind demonstration system To give some further substance for my remarks, and further illustrate the TalaMind approach, I’ll next describe the design, processing, and output of the TalaMind prototype demonstration system [14]. The demonstration system is a functional prototype in which two Tala agents, named Ben and Leo, interact in a simulated environment. Each Tala agent has its own TalaMind conceptual framework and conceptual processes. To the human observer, a simulation is displayed as a sequence of English sentences, in effect a story, describing interactions between Ben and Leo, their actions and percepts in the environment, and their thoughts. The story that
**
Gärdenfors [10] discusses how subject domains help children learn word meanings, and describes the use of conceptual spaces to model semantics of nouns, adjectives, and verbs.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
706 8
is simulated depends on the initial concepts Ben and Leo have, their initial percepts of the simulated environment, and how their executable concepts process their perceptions to generate goals and actions, leading to further perceptions and actions at subsequent steps of the story. The demonstration system includes a prototype design for a conceptual framework and conceptual processes at the linguistic level of a TalaMind architecture, for each Tala agent. The conceptual framework includes prototype representations of perceived reality, subagents, a Tala lexicon, encyclopedic knowledge, mental spaces and conceptual blends, scenarios for nested conceptual simulation, executable concepts, grammatical constructions, and event memory. (A prototype routine converts Tala expressions into English text output displayed by the simulation, creating some typographical errors in the output.) Table 1.‘Discovery of bread’ story simulation output
Time step
Event
1...1
Leo has excess grain.
1...2
Leo tries to eat grain.
1...1 1...4 1...7 1...8 1...8 1...8 1...17 1...20 1...21 1...23 1...28 1...29 1...29 1...29 1...29 1...29 1...30 1...33 1...33 1...36 1...37
Leo thinks Leo has excess grain. Leo asks Ben can you turn grain into fare for people?. Ben examines grain.
Ben thinks wheat grains resemble nuts.
Ben imagines an analogy from nuts to grain focused on food for people.
Ben thinks grain perhaps is an edible seed inside an inedible shell.
Ben mashs grain.
Ben thinks dough is too gooey. Ben bakes dough.
Ben tries to eat flat bread.
Ben thinks people would prefer eating thick, soft bread over eating flat bread.
Ben thinks how can Ben change the flat bread process so bread is thick and soft?.
Ben thinks what other features would thick, soft bread have?
Ben thinks thick, soft bread would be less dense.
Ben thinks thick, soft bread might have holes or air pockets.
Ben thinks air pockets in thick, soft bread might resemble bubbles in bread.
Ben thinks Ben might create bubbles in bread by adding beer foam to dough.
Ben mixs the dough with beer foam. Ben bakes dough.
Leo says bread is edible, thick, soft, tastes good, and not gooey. Ben says Eureka!
The prototype conceptual processes include interpretation of executable concepts with pattern-matching, variable binding, conditional and iterative expressions, transmission of internal speech acts between subagents, nested conceptual simulation, conceptual blending, and composable interpretation of grammatical constructions.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
707 9
In the prototype system, only the linguistic level of a TalaMind architecture is implemented, for both Tala agents. They communicate by exchanging Tala concepts, which are displayed as English sentences spoken by Ben and Leo. It was not possible within the timeframe and resources available for the thesis to implement the archetype or associative levels of a TalaMind architecture. Abstracting out these levels facilitated creation of the prototype. For the thesis, two stories were simulated in which Ben is a cook and Leo is a farmer. The first is a story in which Ben and Leo discover how to make bread. In the second story, Ben and Leo agree to an exchange of wheat for bread and then perform the exchange. In each case, the stories that are simulated are essentially predefined: what happens depends on the initial goals, knowledge, and executable concepts that Ben and Leo have within their conceptual frameworks. Chapter 6, section 4 of [14] discusses how these story simulations illustrate the potential of the TalaMind approach to support human-level AI. I wrote the TalaMind prototype demonstration software in JScheme and Java. 8. The ‘discovery of bread’ story simulation Initially in the simulation, neither Ben nor Leo know how to make bread, nor even what bread is, nor that such a thing as bread exists. We may imagine Leo is an ancient farmer who raises goats and grows wheat grasses for the goats to eat, but does not yet know how to eat wheat himself. Ben is an ancient food and drink maker, who knows about cooking meat and making beer, perhaps from fermented wheat grass. The discovery of bread simulation includes output from a pseudorandom ‘discovery loop’: After removing shells from grain Ben performs a random sequence of actions to make grain softer for eating. This eventually results either in the discovery of dough, or in making grain a “ruined mess”. In the first case, Ben proceeds to discover how to make flat bread, and then leavened bread. In the second case, he says the problem is too difficult, and gives up. Table 1 on the previous page shows a condensed example of output for the first case, omitting several less important steps in the simulation due to page limits for this paper. Each step of the form “Ben thinks…” is an internal speech act produced by a subagent of Ben communicating to another subagent of Ben, using the Tala mentalese as an interlingua. The net effect of this internal dialog is to allow Ben to perform most of the discovery of bread conceptual processing. These internal dialogs also support semantic disambiguation by Ben and Leo of each other’s utterances. Of course, it is not claimed that the story describes how humans actually discovered bread. †† Appendix B of [14] discusses processing at each step of the story. 5F
9. Theoretical and strategic issues for the Common Model of Cognition If the researchers and developers of the Common Model wish to realize Newell’s long-term goals for unified theories of cognition ‡‡ , and also thereby to achieve human-level artificial intelligence, then they must create a system that has the ability to somehow represent and develop the full range of human thoughts which can be expressed linguistically. No existing formal language can do this. A natural language like English already has the ability to do this perhaps as well as any artificial, formal language could – cf. [30]. As discussed above, there is no theoretical reason why the syntax and semantics of a natural language like English cannot be used by an AI system as its language of thought, to help achieve human-level AI. There are several theoretical advantages for using natural language as a language of thought for an AI system. A major strength of natural language is its provision of syntax and semantics that can support meta-reasoning, analogical reasoning, causal and purposive reasoning, as well as abduction, induction, and deduction, in any domain. 6F
††
The story simplifies the process for making bread, and omits steps of threshing and winnowing grain, describing just a single step “pounding grain”. The use of beer foam to leaven bread does have an historical basis: Pliny the Elder wrote that the people of Gaul and Spain used the foam from beer to leaven bread and “hence it is that the bread in those countries is lighter than that made elsewhere” – see [4], IV, Book XVIII, p. 26. ‡‡ See [28], pp. 15-19.
708 10
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
So, if the creators of the Common Model wish to realize Newell’s long-term goals for unified theories of cognition, they should follow the TalaMind approach, or something very much like it, to develop a future generation of the Common Model. There are other routes that can be taken toward human-level AI: Some may attempt to replicate the full range of human thoughts entirely with deep neural nets. Some may attempt to artificially replicate the circuitry of the human brain. Some may continue developing and relying only on simpler formal logic languages that can represent different aspects of human reasoning – an approach that has failed to achieve human-level AI for six decades. Such approaches may in theory be successful, but it is likely they will take more time and effort than the TalaMind approach to achieve human-level AI. If they are successful, the neural network approaches will produce black boxes lacking the explainability needed to support beneficial, ethical human-level AI [16]. If they are successful, the formal language approaches may produce systems with internal languages that are difficult for people to understand, again impeding beneficial AI. 10. Conclusion To realize Newell’s long-term goals for unified theories of cognition, and to achieve human-level artificial intelligence, we should abandon the assumption that natural language must be supported as an application, using only simpler formal languages. It is cognitively plausible that natural language representation and processing are core functionalities of human-level intelligence, needed for internal representation of thoughts. This may be necessary but not sufficient: For example, it appears cognitively plausible that spatial-temporal reasoning and visualization, artificial consciousness (§3.7.6), and other ‘higher-level mentalities’ (§2.1.2) need to be treated as core functionalities. References [1] Aleksander, Igor., and Helen Morton. (2007) “Depictive Architectures for Synthetic Phenomenology”, in Antonio Chella and Riccardo Manzotti (eds) Artificial Consciousness: 67-81, Imprint Academic. [2] Agirre, Eneko, and Philip Edmonds, eds. (2007) Word Sense Disambiguation – Algorithms and Applications. Springer. [3] Baars, Bernard J., and Nicole M. Gage. (2007) Cognition, Brain, and Consciousness – Introduction to Cognitive Neuroscience. Elsevier. [4] Bostock, John, and Riley, Henry T. (1856) The Natural History of Pliny. London: H. G. Bohn. [5] Evans, V., and Green, M. (2006) Cognitive Linguistics – An Introduction. Lawrence Erlbaum Associates. [6] Evans, Roger, Alexander Gelbukhy, Gregory Grefenstettez, Patrick Hanks, Miloš Jakubícek, Diana McCarthy, Martha Palmer, Ted Pedersen, Michael Rundell, Pavel Rychlý, Serge Sharoff, and David Tugwell. (2016) “Adam Kilgarriff’s legacy to computational linguistics and beyond.” CICLing 2016: Computational Linguistics and Intelligent Text Processing: 3-25. Springer. [7] Fauconnier, Gilles. (1994) Mental Spaces: Aspects of Meaning Construction in Natural Language. Cambridge University Press. [8] Fauconnier, Gilles, and Mark Turner. (2002) The Way We Think – Conceptual Blending and the Mind’s Hidden Complexities. Basic Books. [9] Gärdenfors, Peter. (1995) “Three levels of inductive inference.” Studies in Logic and the Foundations of Mathematics, 134: 427-449. Elsevier. [10] Gärdenfors, Peter. (2017) “Semantic knowledge, domains of meaning and conceptual spaces”, in Peter Meusburger, Benno Werlen, and Laura Suarsana (eds) Knowledge and Action: 203-219. Springer. [11] Ide, Nancy, Aurélie Herbelot, and Lluís Màrquez. (2017). Proceedings of the 6th Joint Conference on Lexical and Computational Semantics. Association for Computational Linguistics. [12] Jackendoff, Ray. (1989) “What is a concept that a mind may grasp it?” Mind & Language, 4 (1 and 2): 68-102. Wiley-Blackwell. [13] Jackendoff, Ray. (1992) Languages of the Mind – Essays on Mental Representation, MIT Press. [14] Jackson, Philip C. (2014) Toward Human-Level Artificial Intelligence – Representation and Computation of Meaning in Natural Language. Ph.D. Thesis, Tilburg University, The Netherlands. [15] Jackson, Philip C. (2017) “Toward human-level models of minds.” AAAI Fall Symposium Series Technical Reports, FS-17-05: 371-375.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
709 11
[16] Jackson, Philip C. (2018). “Toward beneficial human-level AI… and beyond.” AAAI Spring Symposium Series Technical Reports, SS-18-01: 48-53. [17] Kilgarriff, Adam. (1997) “I don’t believe in word senses”. Computers and the Humanities, 31: 91-113. [18] Kilgarriff, Adam. (1999) “95% replicability for manual word sense tagging.” Proceedings of the Ninth Conference of the European Chapter of the Association for Computational Linguistics (EACL), Bergen, Norway: 277–288. [19] Kilgarriff, Adam, and Martha Palmer, eds. (2000) Special Issue on SENSEVAL: Evaluating Word Sense Disambiguation Programs. Computers and the Humanities, 34: 1-2. Kluwer. [20] Kilgarriff, Adam. (2007) “Word senses”, in Eneko Agirre and Philip Edmonds (eds) Word Sense Disambiguation – Algorithms and Applications: 29-46. Springer. [21] Laird, John E., Christian Lebiere, and Paul S. Rosenbloom. (2017) “A Standard Model of the Mind: Toward a common computational framework across artificial intelligence, cognitive science, neuroscience, and robotics.” AI Magazine, Winter 2017, 38 (4): 13-26. [22] Lewis, Richard L., Shravan Vasishth, and Julie A. Van Dyke (2006). “Computational principles of working memory in sentence comprehension.” Trends in Cognitive Science, 10 (10): 447-454. [23] Lindes, Peter. (2018) “Common-model-language.” Email to CM language group, June 26, 2018. [24] McCarthy, John. (2008). “The well-designed child.” Artificial Intelligence, 172 (18): 2003-2014. [25] McCarthy, John, Marvin L. Minsky, Nathaniel Rochester, and Claude E. Shannon. (1955) “A proposal for the Dartmouth Summer Research Project on Artificial Intelligence” in Ronald Chrisley and Sander Begeer, eds. (2000) Artificial Intelligence: Critical Concepts in Cognitive Science, 2: 44-53. Routledge Publishing. [26] Newell, Allen. (1973) “You can’t play 20 questions with nature and win: projective comments on the papers of this symposium”, in William G. Chase (ed) Visual Information Processing, 283-308. Academic Press. [27] Newell, Allen, and Herbert A. Simon. (1976) “Computer science as empirical inquiry: Symbols and search.” Communications of the ACM, 19 (3): 113–126. [28] Newell, Allen. (1990) Unified Theories of Cognition. Harvard University Press. [29] Rosenbloom, Paul. S. (2018).” Common-model-language”. Email to CM language group, June 26, 2018. [30] Sowa, John F. (2007) “Fads and fallacies about logic.” IEEE Intelligent Systems, March 2007, 22 (2): 84-87.
ScienceDirect ScienceDirect
Procedia online Computer 00 (2019) 000–000 Available at Science www.sciencedirect.com Procedia Computer Science 00 (2019) 000–000
www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia
ScienceDirect Procedia Computer Science 145 (2018) 699–709
Postproceedings of the 9th Annual International Conference on Biologically Inspired Cognitive Postproceedings of the 9thBICA Annual International Conference Inspired Cognitive Architectures, 2018 (Ninth Annual MeetingonofBiologically the BICA Society) Architectures, BICA 2018 (Ninth Annual Meeting of the BICA Society)
Natural language in the Common Model of Cognition Natural language in the Common Model of Cognition Philip C. Jackson, Jr. * Philip C. Jackson, Jr. * 0F
0F
TalaMind LLC, PMB #363, 55 E. Long Lake Rd., Troy, MI 48085, U.S.A. TalaMind LLC, PMB #363, 55 E. Long Lake Rd., Troy, MI 48085, U.S.A.
Abstract Abstract This paper discusses how natural language could be supported in further developing the Common Model of Cognition following the approach toward human-level artificial intelligence. These thoughts arethe presented as Model answersoftoCognition questionsfollowing posed by ThisTalaMind paper discusses how natural language could be supported in further developing Common Peter Lindes, Paul Rosenbloom, and reviewers of thisintelligence. paper, followed a description of the TalaMind system, the TalaMind approach toward human-level artificial Theseby thoughts are presented as answersdemonstration to questions posed by and general discussion of theoretical and strategic issues for the Common of Cognition. Petera Lindes, Paul Rosenbloom, and reviewers of this paper, followed by aModel description of the TalaMind demonstration system, and a general discussion of theoretical and strategic issues for the Common Model of Cognition. © 2019 The Authors. Published by Elsevier B.V. © 2019 2018 The Authors. Published by Elsevier Elsevier B.V. This is an open accessPublished article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) © The Authors. by B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the the CC scientific committee of the 9th Annual International Conference on Biologically Inspired This is an open access article under BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the scientific committee of the 9th Annual International Conference on Biologically Inspired Cognitive Peer-review under responsibility of the scientific committee of the 9th Annual International Conference on Biologically Inspired Cognitive Architectures. Architectures. Cognitive Architectures. Keywords: natural language, language of thought, mentalese, TalaMind, Common Model of Cognition Keywords: natural language, language of thought, mentalese, TalaMind, Common Model of Cognition
1. Questions to consider 1. Questions to consider In an email thread for the ‘Language’ subgroup of the Common Model of Cognition, Peter Lindes [23] wrote: In an email thread for the ‘Language’ subgroup of the Common Model of Cognition, Peter Lindes [23] wrote: “I think we should be thinking about what are the architectural implications of language processing as one “I think weofshould be thinking about what areinvolve the architectural implications of language processing as one application the Common Model. This might questions like the following: application of the Common Model. This might involve questions like the following: How much of language processing is part of perception and how much is done in other parts of the model? How much of language processing is part of perception and how much is done in other parts of the model?
* Corresponding author. address:author. [email protected] * E-mail Corresponding URL: http://www.talamind.com E-mail address: [email protected] URL: http://www.talamind.com 1877-0509 © 2019 The Authors. Published by Elsevier B.V. This is an open access under the CC by BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) 1877-0509 © 2019 The article Authors. Published Elsevier B.V. Peer-review under responsibility of the committee of the(https://creativecommons.org/licenses/by-nc-nd/4.0/) 9th Annual International Conference on Biologically Inspired Cognitive This is an open access article under thescientific CC BY-NC-ND license Architectures. Peer-review under responsibility of the scientific committee of the 9th Annual International Conference on Biologically Inspired Cognitive Architectures.
1877-0509 © 2018 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the scientific committee of the 9th Annual International Conference on Biologically Inspired Cognitive Architectures. 10.1016/j.procs.2018.11.051
700 2
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
How are words and their meanings stored and retrieved? How are words and their meanings assembled into meaningful interpretations of sentences? What architectural mechanisms are needed for this process? How can linguistic knowledge be represented and retrieved by the architecture? How can language processing become an acquired skill that is performed rapidly and without the agent being able to describe the details of its processing? How can situational context bias retrievals so that appropriate senses of words are retrieved? What architectural mechanisms can allow processing to become a skill in procedural memory and still allow context to bias retrievals?” Lindes suggested papers answering these questions be submitted to the 2018 Fall Symposium on the Common Model of Cognition. In addition, Paul Rosenbloom [29] suggested the following questions: “(1) Which consensus aspects of man-like language processing directly map onto the common model as currently specified; (2) which would suggest straightforward extensions; and (3) which provide major challenges.” Linde’s and Rosenbloom’s questions are of interest to me, relevant to my previous paper [15] about the Common Model, and relevant to the research approach I advocate toward human-level AI [14]. So, the following pages give my answers to these questions, and to some additional questions asked by anonymous reviewers of a draft version of this paper. Throughout this paper, in referring to the Common Model of Cognition (CM), I refer to the description given by Laird, Lebiere, and Rosenbloom [21] of what was previously called the Standard Model of the Mind. I will abbreviate this reference as LLR. 2. Language and the Common Model of Cognition To begin, some general remarks: LLR hypothesize that natural language processing may be implemented using knowledge and skills plus the cognitive architecture of the Common Model. In effect, it appears natural language would be a potential application developed using the standard components of CM (which are working memory, procedural memory, declarative long-term memory, perception, and motor components). LLR express a “strong commitment” that no additional architectural modules are necessary for language processing, though some architectural primitives (such as a phonological loop) can be included. There are implicit languages involved in CM: An implementation of CM presumably has a symbolic language for specifying production rules which can be applied in a cognitive cycle. There are also presumably symbolic languages for specifying items of information in declarative memory or working memory, and a symbolic language for specifying motor actions that can be performed in an external environment. These symbolic languages are essentially formal languages of some sort. My understanding is that they are not envisioned in CM as having the syntactic or semantic complexity of natural language. In this respect, CM would maintain the tradition of AI research that has been followed for decades, which is to attempt to achieve AI using formal languages, and attempt to support natural language processing as an application. The TalaMind thesis [14] advocates a very different direction, which involves developing an AI system using an internal language (called Tala) based on the unconstrained syntax of a natural language (English), and taking a principled approach toward supporting the unconstrained semantics of natural language. Tala is used as a symbolic language for representing information and procedures. In Tala, natural language expressions are symbolic data structures that represent natural language syntax and can refer to cognitive categories for semantics. The TalaMind cognitive cycle uses pattern-matching of Tala expressions for information and procedures. The thesis proposes an architecture called TalaMind for systems to achieve human-level AI, which is further discussed below. For concision, a system with a TalaMind architecture is called a ‘Tala agent’.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
701 3
The above statements are the key differences between the TalaMind approach and the consensus direction of cognitive architectures reflected in the initiative for a Common Model of Cognition. Indeed, it has been a consensus assumption of AI research in general, almost since its inception, that natural language should be supported indirectly as an application, using simpler formal languages to support inference and representation of semantics. The TalaMind approach advocates departing from that assumption, as I’ll explain throughout this paper. My hope is that eventually a new consensus will form to develop a future generation of the Common Model to achieve human-level AI via the TalaMind approach. Although the TalaMind approach maps onto the cognitive cycle of CM, it would be a major extension to the current structure of CM modules. Fully implementing the TalaMind approach will be a major challenge for AI researchers. I will say more, later in this paper, about why this challenge should be accepted, in discussing strategic and theoretical issues for the Common Model of Cognition. 3. Consensus views of natural language Rosenbloom [29] asked about consensus aspects of human-like language processing, and some discussion of this topic is appropriate. There are different consensi in linguistics about different topics related to natural language, as well as debates and lack of consensus about various topics. TalaMind follows a consensus perspective of research on cognitive linguistics and cognitive semantics [5]. TalaMind follows consensus views about natural language syntax for English, advocating support for the full, unconstrained syntax of English, and also advocating support for ungrammatical English expressions (§3.2.1, §3.3, §3.4). † However, TalaMind is not related to previous work in linguistics at the level of cognition corresponding to the current Common Model, e.g. research on sentence comprehension in relation to working memory [22]. The TalaMind approach does not appear to conflict with such research. Rather, it focuses on different topics. There is not a consensus view in modern linguistics about how word senses exist and should be represented – this remains an unresolved topic, philosophically and scientifically as well as technically (the writings of Peirce and Wittgenstein are still relevant). Much modern work on computational linguistics is corpus-based and does not use word senses (i.e., word meanings and definitions). Adam Kilgarriff, a respected lexicographer, wrote a paper twentyone years ago, saying he did not believe in word senses [17]. However, Kilgarriff [20] clarified his position and continued to support research on word sense disambiguation (WSD) [6]. A sub-community within computational linguistics conducts research on WSD, reported in annual SemEval workshops. The TalaMind approach subscribes to the views of cognitive semantics that word senses exist with a radial, prototypical nature; that words may develop new meanings over time, and old meanings may be deprecated; that word meanings are often metaphorical or metonymical and may involve mental spaces and conceptual blends ‡; that commonsense reasoning and encyclopedic knowledge may be needed for disambiguation, relative to situations in which words are used, and that the meanings of words and sentences may depend on the intentions of speakers. Word meanings can have natural language definitions at the linguistic level of a TalaMind architecture, and also have senses that are cognitive categories represented at the archetype level using methods such as conceptual spaces, image schemas, radial categories, etc. There is not a consensus among cognitive scientists about whether humans have an innate ‘language of thought’ (mentalese) comparable to a natural language like English. Researchers have reasoned for and against this idea (§2.2.1, §4.2.1). However, Jackendoff [12] argued that concepts must be expressed internally as sentences in a mental language: Since the brain is finite, concepts from the set possible to express by natural language must be represented internally within the mind as structures in a language. 1F
2F
†
Throughout this paper the § notation is used to refer to chapters and sections in [14]. For example, §3.2.1 refers to Chapter 3, section 2.1. These references can be directly accessed from the Table of Contents in [14] and via hyperlinks within the thesis. ‡ See [8] and §3.6.7.8, §3.6.7.9.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
702 4
Jackendoff called these structures ‘sentential concepts’. The set of concepts expressible by natural language goes beyond concepts easily expressed in propositional logic or first-order logic – it includes the full range of thoughts expressible by natural language (§2.3.1, §3.2.1, §3.3, §3.6). The expressive capabilities of natural languages should be matched by expressive capabilities of mentalese, or else by Jackendoff's argument the mentalese could not be used to represent concepts expressed in natural language: The ability to express arbitrarily large, structured sentential concepts is plausibly just as important in a mentalese as it is in English. The general-purpose ability to metaphorically tie concepts together across arbitrary, multiple domains is plausibly just as important in a mentalese as it is in English. This is not to say mentalese would have the same limitations as spoken English, or any particular spoken natural language. In mentalese, sentences could have complex structures not physically permitted in speech. This is discussed below in the section ‘Words, meanings, sentences’. 4. Cognitive plausibility Two anonymous reviewers of a previous version of this paper wrote: “Tala expressions are list structures akin to parse trees without any explicit limitations so their cognitive plausibility would need to be established.” “What would be great is if the author could actually tell us what this natural language is that he is proposing be directly supported by the architecture. I assume it can’t be a specific natural language like English, because that can’t be what all people are born with, but is it nothing more than a general tree or graph structured representation language, or does it try to capture some forms of language universals in an interesting manner?” The cognitive plausibility of a language of thought based on natural language can be supported in at least two ways, with references to cognitive linguistics and cognitive neuroscience: The first way considers Jackendoff’s argument; the second considers the phenomenon of ‘self-talk’. Jackendoff’s argument suggests (but does not prove) humans are born with an innate language of thought which enables representing any thought they express with a natural language like English, and supports learning a natural language. An ideal Common Model of Cognition would replicate and support this innate language, but we may be decades or centuries from being able to specify the innate mentalese of human infants well enough to replicate it as computational language. So, the TalaMind approach advocates that a cognitive architecture modeling human-level intelligence should support English § as an innate, internal language of thought, using Tala expressions. This is cognitively plausible by virtue of representing an ability of the human brain after it has learned a natural language. The syntax and semantics of English have been subjects for many decades of re-search, which can support endeavors to replicate them computationally. It still won’t be easy, of course. It could be that people are not born with an innate language of thought, but are just born with a very complex bioneural network (the human brain) that supports learning the full syntax and semantics of a natural language like English. In this case, Jackendoff’s argument still implies that the brain eventually achieves a state which supports internally representing sentential concepts expressible in English. So in this case, it is still cognitively plausible for a cognitive architecture modeling human-level intelligence to support representing a natural language as an internal language of thought – again the cognitive plausibility is justified by Jackendoff’s argument. Representing the syntax and semantics of what may be called a ‘natural language of thought’ is likely to be easier than representing the neural network of the human brain, for which research may continue in parallel. Mental discourse (self-talk, inner speech) is a feature people often ascribe to their minds, and a psychological phenomenon which has been remarked upon for centuries. It seems we each have the ability to mentally hear some 3F
§
Any other comparable natural language could be used, e.g. French, German, Dutch, Japanese, etc., or multiple natural languages could be supported. To simplify discussion, I refer only to English.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
703 5
of our thoughts expressed internally in natural language. This topic is discussed in section 2.1 of [3]. Baars and Gage write that inner speech is not just for verbal rehearsal but provides an individual’s “running commentary” on current issues, and is related to linguistic and semantic long-term memory. The fact that we hear inner speech suggests (but does not prove) some thoughts are represented internally in a language of thought with the expressiveness of natural language. So it supports the cognitive plausibility of a natural language mentalese. 5. Internal use of language by AI systems There is not a consensus based on analysis and discussion among AI scientists that an AI system cannot use a natural language like English as an internal language for representation and processing. Rather, this has been an assumption by AI scientists over the decades, who have used programming languages and formal logic languages like predicate calculus to support natural language processing, with very limited success: No AI system can yet understand natural language as well as a five year-old child. In 1955, McCarthy [25] proposed that his research in the Dartmouth summer project would focus on the relationship of intelligence and language. He noted that every formal language yet developed lacked important features of English, such as the ability for speakers to refer to themselves and make statements about progress in problem-solving. He proposed to create a computer language that would have properties similar to English. The artificial language would allow a computer to solve problems by making conjectures and referring to itself. Concise English sentences would have equivalent, concise sentences in the formal language. McCarthy’s proposed artificial language would allow statements about physical events and objects, and hopefully enable computers to be programmed for learning how to play games and perform tasks. Although McCarthy proposed in 1955 to develop a formal language with properties similar to English, his subsequent work did not take this direction, though it appears in some respects he continued to pursue it as a goal. Beginning in 1958 his papers concentrated on use of predicate calculus for representation and inference in AI systems, while discussing philosophical issues involving language and intelligence. McCarthy was far from alone in such efforts: AI research on natural language understanding has attempted to translate natural language into a formal language such as predicate calculus, frame-based languages, conceptual graphs, etc., and then to perform reasoning and other forms of cognitive processing with expressions in the formal language. Some approaches have constrained and ‘controlled’ natural language, so that it may more easily be translated into formal languages, database queries, etc. Since progress has been very slow in developing natural language understanding systems by translation into formal languages, Jackson [14] investigated whether it may be possible and worthwhile to perform cognitive processing directly with unconstrained natural language. The Tala language responds to McCarthy’s 1955 proposal for a formal language that corresponds to English, though probably not in the way he envisioned. It enables a TalaMind system to formulate statements about its progress in solving problems. Tala can represent complex English sentences involving self-reference and conjecture. Short English expressions have short correspondents in Tala, a property McCarthy sought for a formal language. The TalaMind approach can address theoretical questions not easily addressed by more conventional approaches. For instance, it supports reasoning in mathematical contexts, but also supports reasoning about people who have self-contradictory beliefs. Tala provides a language for reasoning with underspecification and for reasoning with sentences that have meaning yet which also have nonsensical interpretations. Tala sentences can declaratively describe recursive mutual knowledge. Tala facilitates representation and conceptual processing for learning by analogical, causal and purposive reasoning, learning by self-programming, and imagination via conceptual blends.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
704 6
6. Discussion of further questions 6.1. Language and perception The previous remarks support an answer to Lindes’ [23] question: How much of language processing is part of perception and how much is done in other parts of the model? There is certainly important processing of natural language that occurs in perception for processing speech, and also for processing visual information that can help disambiguate speech, e.g. when someone points to something to indicate what they mean. However, in the TalaMind approach much processing of natural language semantics occurs in the cognitive cycle, and involves working memory, procedural memory, and declarative long-term memory. Arguably such processing should occur in any cognitive system that attempts to understand natural language, and to create natural language expressions for communication: Cognition should be involved in reasoning to create natural language expressions – it can’t all be perception. 6.2. Words, meanings, sentences Next, I’ll discuss Lindes’ [23] questions: How are words and their meanings stored and retrieved? How can linguistic knowledge be represented and retrieved by the architecture? In the TalaMind approach words and their meanings are essentially stored in declarative and procedural memory, but there is additional structure involved. The TalaMind architecture has three levels, called the linguistic, archetype, and associative levels, adapted from Gärdenfors’ paper on inductive inference [9]. At the linguistic level, the architecture includes the Tala language, a conceptual framework for managing concepts expressed in Tala, and conceptual processes that operate on concepts in the conceptual framework to produce intelligent behaviors and new concepts. The archetype level is where cognitive categories are represented using methods such as conceptual spaces, image schemas, radial categories, etc. The associative level would typically interface with a real-world environment and supports connectionism, Bayesian processing, etc. In general, the thesis does not prescribe research choices at the archetype and associative levels. The TalaMind architecture is actually a broad class of architectures, open to further design choices at each level (§1.5, §2.2.2). The architecture is open at the three conceptual levels, e.g. permitting predicate calculus, conceptual graphs, and other symbolisms in addition to the Tala language at the linguistic level, and permitting integration across the three levels, e.g. potential use of deep neural nets at the linguistic and archetype levels. The conceptual framework at the linguistic level has additional structure, beyond that indicated by the CM terms ‘working memory’, ‘procedural memory’, and ‘declarative long-term memory’: A prototype design for a conceptual framework and conceptual processes is described below in the section about the TalaMind demonstration system. Such architectural mechanisms support addressing Lindes’ [23] questions: How are words and their meanings assembled into meaningful interpretations of sentences? What architectural mechanisms are needed for this process? How can linguistic knowledge be represented and retrieved by the architecture? If a natural language expression is heard or seen in the environment, then in the TalaMind approach there will be conceptual processes for constructing Tala expressions representing alternative syntactic and semantic interpretations of the serial word expression, to understand which interpretation is intended, and reason with the interpretation. These processes may result in asking for clarification, of course. A Tala expression is a multi-level list structure representing the dependency parse-tree (syntax) of a natural language sentence or phrase (with extensions, e.g. supporting executable concepts). When these structures are
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
705 7
created within a Tala agent, or exchanged between subagents of a Tala agent, processing them need not involve parsing English sentences as linear strings of symbols. Such internal processing also need not involve disambiguation: Tala expressions can include pointers to word senses and referents, to support internal processing. Pointers can also be used to represent re-entrant links and lattices in Tala expressions (§5.3.9.1). The TalaMind approach relies on word senses being represented either via natural language definitions, or via concepts represented using methods of cognitive semantics such as conceptual spaces, image schemas, etc. The approach is also open at the TalaMind architecture’s associative level to other methods from computational linguistics for representing and disambiguating word senses, viz. [11], [19]. Further discussion of the nature of meaning and representation is given in §2.2.2, §3.6, §4.2.8, §5.4. 6.3. Language acquisition Lindes’ [23] next question was: How can language processing become an acquired skill that is performed rapidly and without the agent being able to describe the details of its processing? The TalaMind approach does not directly address this question, since natural language representation and processing are built into the architecture of a cognitive system, as a feature of its cognitive cycle. In the TalaMind approach, an agent could describe the details of its processing (though it might have to refer to its associative level as a black box). Some discussion is given in the thesis (§4.2.4) about the human inability to describe how we understand natural language, e.g. in the context of solving Searle’s Chinese Room problem. This is an interesting topic for future research, related to artificial consciousness while learning and developing new executable concepts: We are aware of our physical actions when we learn to ride a bicycle, but later these physical actions become subconscious and automatic. Such a process could be involved in learning a new natural language. 6.4. Situational Context Lindes [23] asked two related questions: How can situational context bias retrievals so that appropriate senses of words are retrieved? What architectural mechanisms can allow processing to become a skill in procedural memory and still allow context to bias retrievals? I think the general answer to this is to represent subject domains ** in a way that supports access to interrelated word senses. If the words and phrases in a situational context indicate that a specific subject domain is involved, then the subject domain could help guide retrieval of appropriate word senses. Methods that have been developed for corpus-based computational linguistics might be applied, viz. [2], [11]. This is a topic for future research, of course. 4F
7. TalaMind demonstration system To give some further substance for my remarks, and further illustrate the TalaMind approach, I’ll next describe the design, processing, and output of the TalaMind prototype demonstration system [14]. The demonstration system is a functional prototype in which two Tala agents, named Ben and Leo, interact in a simulated environment. Each Tala agent has its own TalaMind conceptual framework and conceptual processes. To the human observer, a simulation is displayed as a sequence of English sentences, in effect a story, describing interactions between Ben and Leo, their actions and percepts in the environment, and their thoughts. The story that
**
Gärdenfors [10] discusses how subject domains help children learn word meanings, and describes the use of conceptual spaces to model semantics of nouns, adjectives, and verbs.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
706 8
is simulated depends on the initial concepts Ben and Leo have, their initial percepts of the simulated environment, and how their executable concepts process their perceptions to generate goals and actions, leading to further perceptions and actions at subsequent steps of the story. The demonstration system includes a prototype design for a conceptual framework and conceptual processes at the linguistic level of a TalaMind architecture, for each Tala agent. The conceptual framework includes prototype representations of perceived reality, subagents, a Tala lexicon, encyclopedic knowledge, mental spaces and conceptual blends, scenarios for nested conceptual simulation, executable concepts, grammatical constructions, and event memory. (A prototype routine converts Tala expressions into English text output displayed by the simulation, creating some typographical errors in the output.) Table 1.‘Discovery of bread’ story simulation output
Time step
Event
1...1
Leo has excess grain.
1...2
Leo tries to eat grain.
1...1 1...4 1...7 1...8 1...8 1...8 1...17 1...20 1...21 1...23 1...28 1...29 1...29 1...29 1...29 1...29 1...30 1...33 1...33 1...36 1...37
Leo thinks Leo has excess grain. Leo asks Ben can you turn grain into fare for people?. Ben examines grain.
Ben thinks wheat grains resemble nuts.
Ben imagines an analogy from nuts to grain focused on food for people.
Ben thinks grain perhaps is an edible seed inside an inedible shell.
Ben mashs grain.
Ben thinks dough is too gooey. Ben bakes dough.
Ben tries to eat flat bread.
Ben thinks people would prefer eating thick, soft bread over eating flat bread.
Ben thinks how can Ben change the flat bread process so bread is thick and soft?.
Ben thinks what other features would thick, soft bread have?
Ben thinks thick, soft bread would be less dense.
Ben thinks thick, soft bread might have holes or air pockets.
Ben thinks air pockets in thick, soft bread might resemble bubbles in bread.
Ben thinks Ben might create bubbles in bread by adding beer foam to dough.
Ben mixs the dough with beer foam. Ben bakes dough.
Leo says bread is edible, thick, soft, tastes good, and not gooey. Ben says Eureka!
The prototype conceptual processes include interpretation of executable concepts with pattern-matching, variable binding, conditional and iterative expressions, transmission of internal speech acts between subagents, nested conceptual simulation, conceptual blending, and composable interpretation of grammatical constructions.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
707 9
In the prototype system, only the linguistic level of a TalaMind architecture is implemented, for both Tala agents. They communicate by exchanging Tala concepts, which are displayed as English sentences spoken by Ben and Leo. It was not possible within the timeframe and resources available for the thesis to implement the archetype or associative levels of a TalaMind architecture. Abstracting out these levels facilitated creation of the prototype. For the thesis, two stories were simulated in which Ben is a cook and Leo is a farmer. The first is a story in which Ben and Leo discover how to make bread. In the second story, Ben and Leo agree to an exchange of wheat for bread and then perform the exchange. In each case, the stories that are simulated are essentially predefined: what happens depends on the initial goals, knowledge, and executable concepts that Ben and Leo have within their conceptual frameworks. Chapter 6, section 4 of [14] discusses how these story simulations illustrate the potential of the TalaMind approach to support human-level AI. I wrote the TalaMind prototype demonstration software in JScheme and Java. 8. The ‘discovery of bread’ story simulation Initially in the simulation, neither Ben nor Leo know how to make bread, nor even what bread is, nor that such a thing as bread exists. We may imagine Leo is an ancient farmer who raises goats and grows wheat grasses for the goats to eat, but does not yet know how to eat wheat himself. Ben is an ancient food and drink maker, who knows about cooking meat and making beer, perhaps from fermented wheat grass. The discovery of bread simulation includes output from a pseudorandom ‘discovery loop’: After removing shells from grain Ben performs a random sequence of actions to make grain softer for eating. This eventually results either in the discovery of dough, or in making grain a “ruined mess”. In the first case, Ben proceeds to discover how to make flat bread, and then leavened bread. In the second case, he says the problem is too difficult, and gives up. Table 1 on the previous page shows a condensed example of output for the first case, omitting several less important steps in the simulation due to page limits for this paper. Each step of the form “Ben thinks…” is an internal speech act produced by a subagent of Ben communicating to another subagent of Ben, using the Tala mentalese as an interlingua. The net effect of this internal dialog is to allow Ben to perform most of the discovery of bread conceptual processing. These internal dialogs also support semantic disambiguation by Ben and Leo of each other’s utterances. Of course, it is not claimed that the story describes how humans actually discovered bread. †† Appendix B of [14] discusses processing at each step of the story. 5F
9. Theoretical and strategic issues for the Common Model of Cognition If the researchers and developers of the Common Model wish to realize Newell’s long-term goals for unified theories of cognition ‡‡ , and also thereby to achieve human-level artificial intelligence, then they must create a system that has the ability to somehow represent and develop the full range of human thoughts which can be expressed linguistically. No existing formal language can do this. A natural language like English already has the ability to do this perhaps as well as any artificial, formal language could – cf. [30]. As discussed above, there is no theoretical reason why the syntax and semantics of a natural language like English cannot be used by an AI system as its language of thought, to help achieve human-level AI. There are several theoretical advantages for using natural language as a language of thought for an AI system. A major strength of natural language is its provision of syntax and semantics that can support meta-reasoning, analogical reasoning, causal and purposive reasoning, as well as abduction, induction, and deduction, in any domain. 6F
††
The story simplifies the process for making bread, and omits steps of threshing and winnowing grain, describing just a single step “pounding grain”. The use of beer foam to leaven bread does have an historical basis: Pliny the Elder wrote that the people of Gaul and Spain used the foam from beer to leaven bread and “hence it is that the bread in those countries is lighter than that made elsewhere” – see [4], IV, Book XVIII, p. 26. ‡‡ See [28], pp. 15-19.
708 10
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
So, if the creators of the Common Model wish to realize Newell’s long-term goals for unified theories of cognition, they should follow the TalaMind approach, or something very much like it, to develop a future generation of the Common Model. There are other routes that can be taken toward human-level AI: Some may attempt to replicate the full range of human thoughts entirely with deep neural nets. Some may attempt to artificially replicate the circuitry of the human brain. Some may continue developing and relying only on simpler formal logic languages that can represent different aspects of human reasoning – an approach that has failed to achieve human-level AI for six decades. Such approaches may in theory be successful, but it is likely they will take more time and effort than the TalaMind approach to achieve human-level AI. If they are successful, the neural network approaches will produce black boxes lacking the explainability needed to support beneficial, ethical human-level AI [16]. If they are successful, the formal language approaches may produce systems with internal languages that are difficult for people to understand, again impeding beneficial AI. 10. Conclusion To realize Newell’s long-term goals for unified theories of cognition, and to achieve human-level artificial intelligence, we should abandon the assumption that natural language must be supported as an application, using only simpler formal languages. It is cognitively plausible that natural language representation and processing are core functionalities of human-level intelligence, needed for internal representation of thoughts. This may be necessary but not sufficient: For example, it appears cognitively plausible that spatial-temporal reasoning and visualization, artificial consciousness (§3.7.6), and other ‘higher-level mentalities’ (§2.1.2) need to be treated as core functionalities. References [1] Aleksander, Igor., and Helen Morton. (2007) “Depictive Architectures for Synthetic Phenomenology”, in Antonio Chella and Riccardo Manzotti (eds) Artificial Consciousness: 67-81, Imprint Academic. [2] Agirre, Eneko, and Philip Edmonds, eds. (2007) Word Sense Disambiguation – Algorithms and Applications. Springer. [3] Baars, Bernard J., and Nicole M. Gage. (2007) Cognition, Brain, and Consciousness – Introduction to Cognitive Neuroscience. Elsevier. [4] Bostock, John, and Riley, Henry T. (1856) The Natural History of Pliny. London: H. G. Bohn. [5] Evans, V., and Green, M. (2006) Cognitive Linguistics – An Introduction. Lawrence Erlbaum Associates. [6] Evans, Roger, Alexander Gelbukhy, Gregory Grefenstettez, Patrick Hanks, Miloš Jakubícek, Diana McCarthy, Martha Palmer, Ted Pedersen, Michael Rundell, Pavel Rychlý, Serge Sharoff, and David Tugwell. (2016) “Adam Kilgarriff’s legacy to computational linguistics and beyond.” CICLing 2016: Computational Linguistics and Intelligent Text Processing: 3-25. Springer. [7] Fauconnier, Gilles. (1994) Mental Spaces: Aspects of Meaning Construction in Natural Language. Cambridge University Press. [8] Fauconnier, Gilles, and Mark Turner. (2002) The Way We Think – Conceptual Blending and the Mind’s Hidden Complexities. Basic Books. [9] Gärdenfors, Peter. (1995) “Three levels of inductive inference.” Studies in Logic and the Foundations of Mathematics, 134: 427-449. Elsevier. [10] Gärdenfors, Peter. (2017) “Semantic knowledge, domains of meaning and conceptual spaces”, in Peter Meusburger, Benno Werlen, and Laura Suarsana (eds) Knowledge and Action: 203-219. Springer. [11] Ide, Nancy, Aurélie Herbelot, and Lluís Màrquez. (2017). Proceedings of the 6th Joint Conference on Lexical and Computational Semantics. Association for Computational Linguistics. [12] Jackendoff, Ray. (1989) “What is a concept that a mind may grasp it?” Mind & Language, 4 (1 and 2): 68-102. Wiley-Blackwell. [13] Jackendoff, Ray. (1992) Languages of the Mind – Essays on Mental Representation, MIT Press. [14] Jackson, Philip C. (2014) Toward Human-Level Artificial Intelligence – Representation and Computation of Meaning in Natural Language. Ph.D. Thesis, Tilburg University, The Netherlands. [15] Jackson, Philip C. (2017) “Toward human-level models of minds.” AAAI Fall Symposium Series Technical Reports, FS-17-05: 371-375.
Philip C. Jackson / Procedia Computer Science 145 (2018) 699–709 P. C. Jackson / Procedia Computer Science 00 (2019) 000–000
709 11
[16] Jackson, Philip C. (2018). “Toward beneficial human-level AI… and beyond.” AAAI Spring Symposium Series Technical Reports, SS-18-01: 48-53. [17] Kilgarriff, Adam. (1997) “I don’t believe in word senses”. Computers and the Humanities, 31: 91-113. [18] Kilgarriff, Adam. (1999) “95% replicability for manual word sense tagging.” Proceedings of the Ninth Conference of the European Chapter of the Association for Computational Linguistics (EACL), Bergen, Norway: 277–288. [19] Kilgarriff, Adam, and Martha Palmer, eds. (2000) Special Issue on SENSEVAL: Evaluating Word Sense Disambiguation Programs. Computers and the Humanities, 34: 1-2. Kluwer. [20] Kilgarriff, Adam. (2007) “Word senses”, in Eneko Agirre and Philip Edmonds (eds) Word Sense Disambiguation – Algorithms and Applications: 29-46. Springer. [21] Laird, John E., Christian Lebiere, and Paul S. Rosenbloom. (2017) “A Standard Model of the Mind: Toward a common computational framework across artificial intelligence, cognitive science, neuroscience, and robotics.” AI Magazine, Winter 2017, 38 (4): 13-26. [22] Lewis, Richard L., Shravan Vasishth, and Julie A. Van Dyke (2006). “Computational principles of working memory in sentence comprehension.” Trends in Cognitive Science, 10 (10): 447-454. [23] Lindes, Peter. (2018) “Common-model-language.” Email to CM language group, June 26, 2018. [24] McCarthy, John. (2008). “The well-designed child.” Artificial Intelligence, 172 (18): 2003-2014. [25] McCarthy, John, Marvin L. Minsky, Nathaniel Rochester, and Claude E. Shannon. (1955) “A proposal for the Dartmouth Summer Research Project on Artificial Intelligence” in Ronald Chrisley and Sander Begeer, eds. (2000) Artificial Intelligence: Critical Concepts in Cognitive Science, 2: 44-53. Routledge Publishing. [26] Newell, Allen. (1973) “You can’t play 20 questions with nature and win: projective comments on the papers of this symposium”, in William G. Chase (ed) Visual Information Processing, 283-308. Academic Press. [27] Newell, Allen, and Herbert A. Simon. (1976) “Computer science as empirical inquiry: Symbols and search.” Communications of the ACM, 19 (3): 113–126. [28] Newell, Allen. (1990) Unified Theories of Cognition. Harvard University Press. [29] Rosenbloom, Paul. S. (2018).” Common-model-language”. Email to CM language group, June 26, 2018. [30] Sowa, John F. (2007) “Fads and fallacies about logic.” IEEE Intelligent Systems, March 2007, 22 (2): 84-87.