Материал: искусственный интеллект

However, common approaches on human intelligence assume only a rational and logical nature of human thinking. Against that, many different cognitive experiments have been shown how human thinking can be “easily wrong” or “illogical”, what is associated with some cognitive fallacies [14]. Cognitive fallacies are typically wrong assumptions about the dynamics of cognition and judgments. Experimental examples are usually wrong answers to apparently simple questions, which contradict classical probabilistic frameworks [15]. These answers are apparently due to fast and intuitive processing before some extra level of information processing, evidencing that the dynamic for holistic information processing can be completely different from logical or rational processing of the same information. These cognitive features show interesting properties of concept combinations, human judgments and decision making under uncertainty. For example, one of these cognitive fallacies is the ordering question effect: the idea that the order in which questions are presented should not affect the final outcome or response. However, in psychology and sociology, this effect is well known and called question bias. For example, analyses in 70 surveys from 651 to 3006 participants demonstrated not only that question order affects the final outcome, but also this effect can be predicted by quantum models of cognition, revealing a noncommutative structure [16]. Another example is the conjunction of two concepts like the Pet-fish problem, where the intersection of individual concept probabilities do not explain the observed probability for a typical Pet-Fish like a goldfish [17]. Recently, the explanation of these phenomena were demonstrated using the mathematical framework of quantum mechanics; also known as the quantum cognition (QC) programme [9]. QC uses non-classical mathematical probabilities to successfully explain and predict part of these cognitive behaviours. One of these predictions is the recently proved constructive effect of affective evaluations [18], where preliminary ratings of negative adverts influence the rating of following positive adverts and vice versa. It suggests that the construction of some new affective content has intrinsic connection with QC formalism. In consequence, understanding human intelligence is not possible with only the classical idea of logical and rational kind of intelligence, it is also needed to explain and incorporate emotional and intuitive intelligence (no classical reasoning), apparently better described by QC.

In this way, we will suggest a compositional human intelligence (Fig. 1a) composed of a rational reasoning and also intuitive and emotional one, considering intelligence as the whole “living” body engaged with the environment [10] (brain-body- environment system). Compositional complexes intelligences would be different in the way how they manage both (or even more) rational and emotional processes of information associated with internal and external resources.

2.3In Search of the General Principles of Intelligence

The order-effect of stimuli presentations, conjunction, disjunction, decision making under uncertainty, contextuality, among others, can be understood as an intrinsic property of human judgment: contextual dependencies as inherent to previous knowledge experiences. In other words, humans and animals can understand/distinguish among different contexts and act accordingly, thanks to previously learned experiences. Therefore, one reasonable, but not an intuitive assumption, is to consider these effects as

126 C. M. Signorelli and X. D. Arsiwalla

a general property (or the minimal requirement) of high-level cognition, i.e. some types of bias would be a reaction to contextual dependencies, as wording or framing, implicit meanings, question order, etc. For instance, exchanging words like “kill” instead of “save” in moral judgment, triggered different answers even if the outcomes of each dilemma were the same in both situations [19]. If some changes in the formulation of dilemmas trigger different contextual internal meanings and they evoke different answers according to different contexts, it would mean that high-level cognition can understand and distinguish different context and respond differently to each one in a way that is not always optimal (rationally speaking).

A true understanding of different contexts (in the sense of [20]) implies the existence of a minimal kind of meaning. This understanding of context due to meaning (e.g. BUP does not mean the same than PUB) would require a minimal noncommutative structure of cognition connected with our suggested principle of balance internal and external resources. We hypothesized this structure as part of an inherent neural network construction in the brain, where structure-function relationship would be dynamic, highly flexible and context-dependent [10]. Emotion and rational reasoning, in this sense, would correspond to the outcome of interactions among many components, each one related to neural assemblies or distributed networks, which combine, influence, shape and constrain one another [1, 21].

Fig. 1. (a) A proposal for Compositional Intelligence. Context is constructed by composition of external and internal objects; the interaction evokes and/or creates meaning from which intuitive/emotional and rational reasoning would emerge. (b) Moral Test and Processes required. Some processes needed for moral thought are stated as examples, among many other possible processes.

To sum up, the idea of balancing internal and external resources as a preliminary proposal for a general intelligence, implies at least four requirements (Fig. 1): (i) a noncommutative structure (mathematically and operationally), (ii) understanding of context, which implies different kind of contexts effects, framing, wording, question ordereffects, etc., (iii) meaning projection in at least emotional and rational meaning spaces, which implies a basic notion of subjectivity (see discussion on Sect. 4), and (iv) behaviours, judgments and decisions based on these meanings (reasoning would emerge from these meaning interactions).

3 Towards a Moral Test

Even if these previous requirements were defined regarding human cognition, it is expected that any kind of true intelligence can understand different contexts and evoke, create or recreate meaning of their internal and external resources. Particularly, in humans, our definition of human intelligence implies that humans can integrate different types of information and manage a balance between internal and external social resources through intuitive, emotional and rational reasoning, which, in turn, would emerge from meaning interactions. Thus, the next question is how to measure and quantify these requirements, both in humans and machines.

One way to answer this question is searching for situations where humans need to explicitly use both emotional and rational resources to solve complex problems. One example is a certain kind of moral dilemmas. Moral dilemmas are controversial situations to study moral principles, where subjects need to judge some actions and sometimes take difficult and even paradoxical decisions. Moral dilemmas are simple, in the sense that they do not require any kind of specific knowledge, but at the same time, some of them can be very complex because they require a deep understanding of each situation, and deep reflection to balance moral consequences, emotions and optimal solutions. No answer is completely correct, they are context dependent and solutions can vary among cultures, subjects, or even across the same subject in particular emotional circumstances. Moreover, morality and ethics are not necessarily associated with a particular religion, political view, education level, age or gender [22, 23], while it seems an intrinsic human condition and a very “relative” (or even biased) feature. People that give the impression to act against the moral establishment, really act according to their “own” moral, apparently developed in completely different social accepted conditions. One simple example is the acceptation of monogamy or polygamy and therefore certain moral attitudes in different societies. Morality, in this way, is the set of internal (particular experience) and external (social culture) values learned by experience, which allows us to behave in our societies on demand of both emotional and rational thoughts, usually reacting in a very intuitive and fast way. Morality also requires many previous processes associated with high-level cognition (Fig. 1b), starting for decision-making to self-reflection, to be able to detect mistakes on these decisions; sense of confidence, to estimate how correct a decision or action is; mental imagery, to create new probable scenarios of action; empathy, to equilibrate individual and social requirements; understanding of context, to adapt moral decisions to the context, among others. Therefore, the main suggestion is that moral and ethics emerge as a way to integrate individual and social regulation (as different types of information) in human species, and apparently also in other animals species [24] (even if human moral can be completely different in comparison with animal moral). Morality is related to both rational and emotional reasoning [25] and it has the peculiarity to be very dependent of the context, wording and framing [19, 23], kind of social community, subjects and probably even emotional states of each subject [26].

Hence, any precise test to measure the distinct human intelligence (or even nonanthropomorphic intelligence) should consider the way of thinking and information processing to develop moral thoughts, independently of what is judged as a correct or

128 C. M. Signorelli and X. D. Arsiwalla

incorrect about these thoughts in our societies. It is how the dynamics of answers and meanings to moral dilemmas change depending on the context, the capacity to justify or not any action, and report our intrinsic experience according to intuitive and rational reasoning, what is really important in human intelligence and what should be measured as an attribute of intelligence (balance between rational and emotional processes).

4 Compositional Quantum Cognition

In order to implement the idea of a Moral test, it is necessary a compositional cognitive model and specifically a model of natural language able to incorporate the noncommutative, bias and commutative properties at different levels and different contexts of natural answers to different moral dilemmas (where rational and emotional thoughts are involved). For one side QC seems a good candidate for cognitive model while CCMM is for natural language. CCMM has been proved a better theoretical framework than only distributional and symbolic models of natural language [27] and more general than some QC models [28]. Even though some of these experiments have been only made in statics text corpus, they suggest a potential richer description for more complex cognitive experiments.

Therefore, in this section and in order to maintain simplicity, the main concepts of CCMM framework will be presented together with a preliminary framework, which expects to integrate QC and CCMM in a common approach. For a more complete mathematical background and description about CCMM and QC, please refer to [8, 29] and [9, 15, 30] respectively.

Fig. 2. Semantic space with emotional and rational basis. (a) One example of Vector Space for words: Dog, Man, Healthy and Injured. The basis is a combination of emotional (Joy vs. Sadness) and rational categorization (Efficient vs. Inefficient). (b) Example of Convex Space for the same words above. One emotional convex space (Joy, Sadness, Anticipation, Surprise) and other rational convex space (Big, Fast, Efficiency). In convex spaces, words correspond to regions of the space, instead of only vectors as it is in vector space.

4.1Categorical Compositional Model of Meaning

Applications of CCMM basically needs: (1) Define a compositional structure as a grammar category (e.g. pre-groups grammar); (2) Define a meaning space as a semantic category, for example vector space or conceptual space of meaning; and finally,

(3) Connect both categories in some way that it is possible to interpret the grammar category into the semantic category (mathematically, one has to define a functor between categories) [8]. Thus, some useful concepts are:

Compositional Structures. In CCMM, compositional structures are certain rules/definitions about how elements compound each other. In other words, how processes, states, effects, among other possible elements, compound. Grammatical types and their composition are described using a pregroup algebra due to Lambek [31]. However, any kind of grammar definition can, in principle, be implemented. Grammar will be interpreted as the way how word meanings interact, defining first primitive types as nouns n, sentences s, and then other types like adjectives nnl, and verbs, for instance, a transitive verb as nrsnl, among other kinds of words to form complex sentences.

Semantic Spaces. Semantic spaces are spaces where individual words are defined with respect to each other. The simplest way is using distributional approaches to define vectors of meaning for each word (Fig. 2a), or even better, defining density matrices. The choice of a basis vector and how to build other words, adjectives and verbs from the basis, is not trivial and it can be done in many different ways. Of course, it will depend on what the experimenter would like to describe and compare.

Conceptual Spaces. The idea of conceptual spaces, recently suggested in [32], is a more cognitively realistic way to define semantic spaces. This approach is called convex conceptual spaces. In short, concepts can be defined by a combination of others primitive features or quality dimensions, building spaces which can be superposed or not, to define regions of similarity. One perceptual example is to define taste based on some features such as: Saline, Sweet, Sour, and Bitter. Then, different kind of food would be described with a certain level of each taste dimension, and where other features like colour, texture, can also be incorporated [29]. Other complex example is defining elements regarding emotional states and factual features (Fig. 2b). Additionally, conceptual spaces require two semantic/meaning spaces, one for words in a quality dimension space (Fig. 2) and other for sentences in a “sentence meaning space” (Fig. 3a), then, the final sentence meaning is an “interaction” between both spaces.

Computing the Meaning of a Sentence. Diagrammatically, the final meaning of one sentence will be the meaning of individual words interacting according to the grammatical structure, defined as a process [8, 29].

4.2Proof of Concept: Compositional Quantum Cognition

In our framework, meaning is the interaction between external and internal objects, understanding external objects/contents as transductions of external stimuli, while internal objects/contents would correspond to the “space of transduction” that will help