Материал: искусственный интеллект

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

to create internally, these external objects1. These objects can have different levels of complexity; some of them can act as a constitutive element in the construction of other contents or being formed by other more fundamental elements. Specifically, internal objects can create different levels of what we call “quality dimensions” (as a basis space) and external objects are formed by and move in these quality dimensional spaces. In other words, external objects are defined with respect to internal objects which will form a type of internal space depending on the specific problem to model.

Regarding our specific hypotheses, these internal objects and/or quality dimensions can be emotions (usually based on belief) or reasons (usually based on facts), while external objects will be a convex space (for each object) defined by these emotions and reasons (Fig. 2b). Both, internal spaces and external objects create what in other frameworks is called “state of mind”, usually represented by w. However, in our approach, it is not a simple state; w will correspond to a complex internal space and external objects which are embodied into another “decision space” (Fig. 3a). This decision space could be a similar space as defined in QC. This is not; however, the end of the story, this decision space is again embodied into a bigger “meaning space” (sentence space), where decisions/behaviours/thoughts/judgments/concepts among others, can project into meanings (Fig. 3b). The sentence meaning space is created by other level of internal objects, in this case “internal values”, so any decision/behaviour/ thought/judgment/concept would have a meaning corresponding to different superposition of values (Fig. 3c). Thus, in this framework, meaning will be the projection/ interaction of convex external objects into a first layer of internal object space which through a decision/behaviour/thought/judgment/concept project another meaning in a second layer of internal object space (thanks to reasoning). Since each person would have a unique internal space in each layer (even if they share the same axes, the topology can be different), the meaning is intrinsically related to subjectivity. In this sense, different contexts would trigger different or similar meanings (projections/ interactions) and in consequence decision/behaviour/thought/judgment/concept would have different or similar meanings, depending on the type of interaction.

If the interaction corresponds to a rational process, the meaning would be relatively fixed across different situations, but if it is emotional/intuitive, the meaning will change with the situation (unfixed meaning). Additionally, the beliefs in the emotive component and the facts on the rational part of an external object can be known or unknown by the subject. If they are known, the subject can access and report these beliefs or facts, in some way that question (in the “decision space”) about them only fill a thirdperson lack of information, however when they are not known, the subject needs to create or re-create them in order to report internally (first-person) and/or externally (third-person) something about them. In that situation, elements about the potential report would be also part of the creation, re-creation or co-creation of these beliefs or facts. As suggested in [15, 33, 34] and others, classical probability (CP) would mainly

1We avoid the term representation because in the literature it has been invoked with many different connotations.

Fig. 3. (a) Decision Space would correspond to the usual QC description. These decisions can be previously known or unknown by the subject. w the psychological state, is defined based on emotional and rational quality dimensions. (b) Sentence Space is a second level of meaning definition (for sentences). In this example the basis is defined regarding some values: Compassion, Love and Loyalty. Sentence meaning can be fixed (static) or unfixed (dynamical).

(c) Decision and sentence/meaning interaction. Decisions would be explained in a sentence meaning space. In this example, two different decisions, one for Alice (save the man) and other for Bob (save the dog), hidden slightly similar meanings. (d) The big picture: Quality dimensions would be embodied into the other complex decision and meaning spaces.

apply when subjects known about their contents (classical-deterministic reasoning) while quantum probability (QP) will work better when they are unknown (non- classical-determinist reasoning, or intuitive reasoning).

Thus, both descriptions would be part of, at least, two kind of intelligence, that all together and in a compositional way, would converge to a more complete description of human intelligence, something that we preliminary call a compositional approach of human cognition or compositional intelligence and it is expected to be deeper developed in future works as a compositional quantum cognitive approach or even a more general compositional contextual model of cognition.

Summarizing, our framework is a complex three space interaction, where external objects can be created by emotional and rational quality dimensions (internal objects), which can be known or unknown by the subject (following CP or QP respectively),

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

having an impact in the evocation, creation or co-creation of fixed or unfixed meanings, which in turn expose emotional/intuitive or rational reasoning (Figs. 1a and 3).

5 Implementation of a Moral Test

Now, we have all the elements, at least conceptually, to describe a research program of implementing a moral test. This strategy has one theoretical and three experimental research steps. The main goal in our implementation is keeping the simplicity of the written exchange of questions and answers and analysing them using non-commutative models of language at different levels, looking for bias meaning and “subjectivity” from previous knowledge.

5.1Step I. Mathematical Model Definition

Moral dilemmas seem better than simple day to day questions, especially to test some of the requirements and hypotheses stated above. For instance, manipulating contexts in the same moral dilemma will allow us to measure the sequential change in responses and how they follow a classical probability combination when only fixed thinking is present, non-classical probability combination in only intuitive or unfixed thinking and finally how the dynamic of responses changed when the dilemma confronts both emotional and rational thoughts. For example, in experiments were subject were asked to justify their moral judgments, most of them failed to properly explain their choices, revealing high intuitive (emotional) components, instead of rational ones [22] and suitable to be described by QC more than classical frameworks. Hence, we suggest that if the subject rationalizes the dilemma, or situations are manipulated to strongly confront emotional (intuitive) versus rational thoughts, the outcome/meaning would change, and the decision or judgment will be more difficult.

In order to do that, the kind of moral dilemmas required in our implementation should confront emotional and rational resources to judge actions, in a way where utilitarian decisions should not be intuitive, while the probable human answer would be predictable, and only after reasoning (emotional or rational), the non-predictable option would emerge as a possibility. So, our set of moral dilemmas and paradigm need to be slightly different from the existing ones in the current literature.

According to the intelligence definition above, purely emotional or rational judgments would not be intelligent decisions in these contexts; instead, an intelligent decision would balance both possibilities until reaching the best compromise for each subject (in his/her space of meaning). It means that measuring the probability of each answer (as usually in QC) is not enough; it is also needed to compute the meaning behind of each answer. This can be done asking directly to the subject why he/she took one or another decision (asking for justification), what is closely related to a phenomenological approach [35] and computing their answers thanks a natural language model.

Thus, a more complex mathematical model is needed, both to define external and internal objects and the interactions to compute and compare meanings across different subjects, sentences and even machines. It requires the developing of both a novel

experimental paradigm (variations of experiment in Appendix A) and adaptation of CCMM into an ideally more complete compositional model of cognition, what was conceptually defined in Sect. 4.2 and where mathematical definitions are expected as previous requirement to next steps.

5.2Step II: A Toy Example, Moral Dilemmas and Context Effects

A simple example of experiment would have a structure like this: (1) Definition of a conceptual semantic space on quality dimensions and space of meaning. It means to define a set of internal objects (quality dimensions) like emotions and reasons (facts) which will correspond to the axes, and the projections of “external” concepts into these axes, forming a convex space for each concept (Fig. 2). (2) Then, a set of different versions of the same dilemma where only a few concepts are changed to confront different degrees of emotional and rational values. (3) Each dilemma would have one introduction to the story and context, one question associated with the agreement of two or more different actions (Question 1), another question associated with the personal decision (Question 2) and a final “why” question to justify their choice (Question 3). Question 1 and Question 2 can be exchanged to show order effects and/or constructive affective effects. One example is in Appendix A.

The moral dilemma in Appendix A is a variant with three different questions and can be manipulated in many different ways to confront emotional and rational thoughts. So, first, a set of different versions of the same moral dilemma would allow us to describe the understanding of each context and perhaps even looking for contextuality, and secondly, measuring the time reaction for each version will be a way to quantify the dynamic of each judgment (easy, difficult, fast, and slow). If our hypotheses are correct, it is expected that the answers will depend of the modifications in the dilemma formulation and degree of conflict among dilemmas (rational vs. emotional).

Other variations of these ideas and maybe a better way to test the QC phenomenon can be incorporated in the final experiment, for example, using conjunction, disjunction, among other effects observed in QC. Additionally, after each moral dilemma, we can ask for the degree of emotional arousal using standard cognitive tests.

5.3Step III. Quantification of Meaning in Moral Dilemmas

The conceptual convex space defined above should be a conceptual subjective space from which the meaning of crucial words (external objects) in each dilemma will be previously defined according to the individual categorization of words in a fixed or flexible quality dimensional space (internal objects) of at least two layers.

It can be done in many different ways. For example, some emotional and rational quality dimensions can be defined as internal objects and each subject would be able to determine concepts (external objects) according to different values of these quality dimensions (Figs. 2 and 3). Each subject will be asked to evaluate (in a scale of points) some words with respect to these dimensions, building a convex space for each crucial word and subject. Other words can be defined thanks affective lexicon from [36]. Another option is to directly ask the meaning of each crucial word, for instance: What does a dog mean to you? Or what does a man mean to you? Etc. Consequently, the

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

answer can be taken as the direct meaning of each concept and use it to reduce sentences. However, this second option is less simple than the first option.

With the semantic space for each subject, the answer to the “why” question can be quantified in terms of the sentence meaning (interaction or projection into word and sentence space) for each sentence. These answers and sentences contain the essential features of a moral thought and the meaning is expected to change from simple meaning to more complexes, depending on how complex is the dilemma (more or less degree of emotional and rational confrontation).

One consequence of this analysis would be the possibility to predict or at least correlate the final decision according to the subjective meaning of individual words, with respect to each dilemma. At the same time, general meanings can be inferred even when decisions can be completely different. For example, if “Bob likes Dogs”, it is possible that Bob would save the dog instead of the man, while if Alice, who hypothetically does not like dogs, would likely do the opposite. However, if a quality dimension is defined with respect to emotions vs. reasons, and sentence meaning with respect to values like love, loyalty among others (Fig. 3), the final meaning for Bob saving the Dog could be similar than the meaning of saving the man for Alice (Fig. 3b). In other words, apparently different decisions would have similar meanings, and the opposite can be also true: same decisions hide completely different ones.

The way how these meanings evolve from context to context is what we refer here as one general property of high-level cognition associated with QC effects and subjectivity, which in the end can be quantified and compared thanks CCMM. In other words, QC and CCMM will be used to compute some kind of subjectivity in a way that is compared across subjects, making this approach a novel tool to complement the phenomenological program suggested by Varela in [35].

5.4Step IV. Application for AI

The last step would be the implementation of moral dilemmas in AI programs using QC to search for non-commutative structures and CCMM to compute their “why” answers. Thanks to some variations in CCMM and QC (step 1), the decisions and meaning of answers would be directly computed and the structure of meaning across human compared with the meaning across different instances of AI programs, developing a way to compare subjective features across humans and machines.

This implementation is not something trivial and will require a big effort in both, previous validation of moral dilemmas in human (preliminary steps) and then adaptations for AI. For example, it is expected to arrange similar experiences than step II, and simulate different instances to compute the same effect (if there is or not) comparing changes in the kind of structure for different versions of our dilemmas. Concretely, AI would be adapted to answer the same experimental set-ups for human, but any modification in order to facilitate or not the answers of the software will be avoided. Then, QC effects will be quantified in the same way than humans and the meaning of their answers will be “computed” using the same strategy than in human experiments, from the semantic space defined by the machine.

Therefore, the first attempt would be searching for a simple “understanding” of context, i.e. if AI programs can distinguish and differentiate versions of the same moral