Making AI meaningful again

TL;DR

This paper critiques current AI approaches, especially in language processing, highlighting underlying issues and proposing an alternative perspective that incorporates philosophical insights to make AI more meaningful.

Contribution

It identifies fundamental problems in current language-centric AI and introduces a novel approach that integrates philosophical considerations.

Findings

Current AI faces serious conceptual problems.

Deep learning successes have not addressed core issues.

Philosophy offers valuable insights for AI development.

Abstract

Artificial intelligence (AI) research enjoyed an initial period of enthusiasm in the 1970s and 80s. But this enthusiasm was tempered by a long interlude of frustration when genuinely useful AI applications failed to be forthcoming. Today, we are experiencing once again a period of enthusiasm, fired above all by the successes of the technology of deep neural networks or deep machine learning. In this paper we draw attention to what we take to be serious problems underlying current views of artificial intelligence encouraged by these successes, especially in the domain of language processing. We then show an alternative approach to language-centric AI, in which we identify a role for philosophy.

Tables

Table 1. Table 1: Minimal desiderata for a real-world AI language processing system

Property	System	Example
Exactness	needs to be able to be exact where necessary and not always restricted to the merely approximative	in the insurance domain: automated validation and payment of a claim
Information security	needs to avoid insecurities of the sort which arise, for example, when even slight perturbations lead to drastically erroneous outputs	in autonomous driving: avoid harmful consequences of adversarially manipulated traffic signs
Robustness	needs to be able to work reliably in a consistent way even given radical changes of situation and input, or to detect critical changes and report on its own inability to cope	in any domain: content not understood by the system is marked for inspection by a human; an alert can be generated if necessary
Data parsimony	needs to be trainable with thousands to millions of data points (rather than billions to trillions  –  magnitudes which rarely occur in reality)	in the domain of business correspondence: automation of letter-answering on the basis of just a few thousand examples per class of letter
Semantic fidelity	needs to be able to incorporate contextual interpretations of input situations	in sentiment analytics: the Declaration of Independence should not be classified as hate speech
Inference	needs to be able to compute the consequences of given inputs in a way that allows the system to distinguish correlation from causality (thus requiring the ability to reason with time and causation)	in medical discharge summaries: determination of required actions on the basis of text input, for example in automated processing
Prior knowledge usage	needs to be able to use prior knowledge to interpret situations	in claims management: understanding that issuing a declaration of inability to pay implies earlier receipt of a payment request

Equations

$$\text{text}⇝\Gamma$$

$$\Gamma \curvearrowright \Delta$$

$$\Delta ⊢{\varphi }_i\in \Omega ,\forall i=1\dots n$$

$$\text{text}⇝\Gamma \curvearrowright \Delta ⊢{\varphi }_i\in \Omega ,\forall i=1\dots n,$$

Full text

Keywords: Artificial intelligence, deep neural networks, semantics, logic, Basic Formal Ontology (BFO)

1 The current paradigm of AI: Agnostic Deep Neural Networks (dNNs)

An AI application is a computer program that can create an output in response to input data in a way that is similar to the ways humans react to corresponding environmental stimuli. In what follows we will focus on AI applications that work with natural language input, where the currently dominant paradigm is provided by what is called agnostic deep machine learning.111Also referred to as ‘brute force’ learning The latter is a subfield of applied mathematics in which input-output-tuples of data are used to create stochastic models, in a process often (somewhat simplistically) referred to as ‘training’. The inputs are connected to outputs probabilistically, which means that there is a certain (a priori unknown but measurable) likelihood that a given input will be associated with a given output. The models are referred to as ‘stochastic’ because they work by utilizing the fact that the data on which they draw is probabilistic in this sense. The models are, in addition, ‘agnostic’  –  which means that they do not rely on any prior knowledge about the task or about the types of situations in which the task is performed, and they are often “end to end,” which means that they are meant to model an entire process such as answering a letter or driving a car. The models are, finally, ‘deep’ in the sense that their architecture involves multiple layers of networks of computational units (thus not, for example, because of any depth in their semantics)

For agnostic deep learning to be useable in creating an AI application, a number of conditions must be satisfied:

A sufficient body of training data must be available in the form of tuples of input and output data. These are digital mappings of, respectively, a situation in response to which an action is required, and an action of the corresponding sort [18]. A classical AI-application in this sense is the spam filter, whose initial output data were created using annotations, in this case adding the label “spam” to email inputs. 2. 2.

Computers must receive the training material in digital form, so that it can be processed using the computing resources available today [7]. 3. 3.

The annotated training tuples must be reasonably consistent (noise-poor)  –  that is, similar inputs should lead to similar outputs. This is because machine learning requires repetitive patterns  –  patterns that have arisen in a recurring, rather than erratic, process. The behaviour of human email users when identifying spam forms a repetitive process of the needed sort. The reason for this is that users of email have a motive to become experts in successful identification of spam, since they are aware of the high costs of failure. The movement of the oil price over time, in contrast, is an example of an erratic process. This is because the input data pertaining to geopolitical and economic events bear no consistent relation to the output data, for example the price of Brent crude. 4. 4.

The data input must be abundant, since a machine-learning algorithm is a stochastic model that needs to represent as much as possible of the variance which characterises the situation in which the model is to be used. Because in language applications the overall complexity of the relationship between input and output is typically very high, the models will need many parameters. For mathematical reasons these parameters can only be estimated (through the type of optimisation process otherwise called “training”) on the basis of huge data sets. If the training sets are too small, there is a high chance that novel input data will not have the properties of the data sampled in the training distribution. The model will then not be able to produce an adequate output under real production conditions.

Most of the AI applications in current use, for example in product recommendation or advertisement placement, draw on situations where these four conditions are satisfied. To establish the training set for the first spam filters, developers needed to collect millions of input-output data tuples where inputs are emails received by humans and outputs are the classifications of these emails by their respective recipients either as spam or as valid email. They then trained a machine-learning model using these data tuples and applied the result to new emails. The goal is that the model should replicate the human reaction it has been trained with, which means: identifing spam in a way that matches, or even outperforms, the behaviour of a typical human.

In classification applications such as this, it is only knowledge of a very simple type  –  knowledge captured by simple input-output-tuples  –  that is given to the machine by its mathematician or AI-engineer trainers. In some cases, application developers may wish to improve the model that is generated by the algorithm from the data by selecting for training purposes only those tuples that have certain desired properties (as when, in building training models for autonomous cars, they in effect seek to approximate the positive aspects of the driving behaviour of mature females rather than that of teenage males). The performance of the machine is in such cases designed to surpass that of the average human, because the trainers of the model select only the most desired sorts of responses from what may be a much more considerable variance exhibited in actual behaviour. Designers may also select data that have been somehow validated by experts for correctness, creating what is called a “gold standard” set of annotations. Because the engineer uses prior knowledge about data quality when making such selections, this is equivalent to an  –  albeit minimalistic  –  usage of prior knowledge in machine learning.

When such strategies are followed, machine learning with neural networks can out-perform even the strongest human performance with regard to both efficiency and effectiveness,222Increasing efficency means: reducing unit production costs; increasing effectiveness means: achieving higher desired quality per production unit. and we can now distinguish three types of cases in which such better-than-human performance is achievable:

dNNs with higher efficiency than is obtainable by humans: when the behaviour that is modelled consists of truly repetitive processes with narrow scope and with data that can easily be represented in digital form: for example in complex industrial automation tasks, following a pattern that has been the driver of engineering since the industrial revolution. 2. 2.

dNNs with higher effectiveness than humans: in hypothesis-based pattern identification, for example in the recent identification by a dNN of a correlation between retinal patterns and cardiovascular risk factors [37]. 3. 3.

AI with both a higher efficiency and effectiveness than with humans: achieved in reinforcement learning, a method used in certain narrowly defined situations of the sort that arise in games (for example in GO [42] or First Person Shooters [21]) and in contexts that can be framed like games [45].

Examples of applications under each of these headings that will be possible in the near future are: (i) driving a car on a highway under (near-)average weather conditions, (ii) scientific pattern-search applications, for example in biology or astronomy, (iii) maintenance robotics as in industrial nuclear plant waste removal.

Unfortunately, each of these types of situations is highly restrictive, and none occurs where we are dealing with natural language input.

2 Applying agnostic Deep Neural Networks in the field of language understanding

To understand how modern agnostic deep-neural-network AI works in the language domain, consider the most prominent production example, which is that of machine translation as illustrated by Google Translate.333https://translate.google.com/ A recent publication authored by Google Brain444This is the official name of Google’s AI department. While Google’s machine-learning engineers are certainly among the world’s leading representatives of their craft, the name nonetheless reveals a certain hubris. and Google Research with the title “Attention is all you need” [48] provides a representative example. The stochastic models described in this paper were trained for the translation of English to German and of English to French. To train Transformer  –  which is the best-performing “big” model described in the paper  –  the authors encoded the language material at their disposal using the method of byte-pair encoding, which encodes each single-sentence input into an encoding vector of 1024 real numbers (rounded to a certain number of decimal places555This encoding approach is used (with variations on how the vector is created) by all dNNs since “word2vec” [32].). This is a complexity-reducing encoding, which means (very roughly) that it treats each sentence simply as a series of meaningless signs.666In this entire text, “meaning” signifies the relevance to the actions and thoughts that humans attribute to the stimuli that they encounter in sensation. For a non-English speaker, an English sentence, too, is a series of meaningless stimuli. For an English speaker, in contrast, the sentence is immediately interpreted as meaningful. This allows the encoding process to retain certain important features of the input sentences because relevant sentence patterns are repeated in many sentences in a similar way, and these sentences are shown to the algorithm.777For example, the algorithm learns to translate the German word ‘Mehl’ into ‘flour’ because this pair is repeated many times in training sentences. But it will fail to translate “Wir haben Mehl Befehl gegeben zu laufen” into the adequate “We ordered Mehl to run”. It rather gives out the nonsensical “We have ordered flour to run” (result produced on Jan. 7, 2019). The translation fails because there are not enough training examples to learn the martial usage of surnames without title.

But at the same time, it necessarily leads to the discarding of many subtleties of these sentences. This is because the embedding of the sentence loses relations not only between words within the sentence but also between sentences. For a further feature of the experiments reported is that the models used are trained with quite small amounts of training data: 36 million sentence pairs for English-French, and only 4.5 million for English-German. The models are completely agnostic: they have no knowledge of linguistics, for example, because they have no knowledge of anything at all. Rather, they just try to mimic the human translations (or rather the corresponding sets of vectorially simplified input-output pairs) they learned from. The principal problem with this approach, however, is that embedding into a linear vector of encoding real numbers  –  no matter how long the vector is  –  leads to the discarding of all information pertaining to the contexts of the input sentences. That this has adverse consequences becomes clear when we reflect that, in all language interpretation processes, even for single sentence inputs, humans use prior knowledge to contextualise the sentences they receive. As an example, consider how a typical reader of this text would contextualise the single sentence: “In the beginning was the word.”888To a reader without knowledge of the Bible this sentence (John, 1,1) will seem strange or unintellegible. It is impossible to enumerate all such contextual constellations and include them as annotated features to training sets for stochastic models in amounts sufficient for machine learning.

2.1 Results thus far

How well, then, do these models do? Transformer, specifically, creates a model that achieves a sentence-level score of 28.4 for English-German and 41.8 for English-French using the BLEU metric, which measures on a scale from 0 to 100 the degree of matching of the machine-translation with a human gold-standard translation [36]. A score of 100 can never be achieved because there are always several valid translations for any given sentence and not all of them can be in the gold-standard set. But 75-85 could be achieved in theory. Such a score would be excellent, and it would correspond to the translation abilities of an average bilingual speaker. The scores achieved by Transformer, in contrast, which are reported as the state-of-the art in machine translation, are low.999To illustrate the limitations of the approach, Hofstadter used input sentences with a high degree of cross-contextualisation (see “The Shallowness of Google Translate”, The Atlantic, January 30, 2018). Text by Hofstadter: In their house, everything comes in pairs. There’s his car and her car, his towels and her towels, and his library and hers. Google Translate: Dans leur maison, tout vient en paires. Il y a sa voiture et sa voiture, ses serviettes et ses serviettes, sa bibliothèque et les siennes. Translated back into English by Google: In their house everything comes in pairs. There is his car and his car, their napkins and their napkins, his library and their’s. The reason for this shallowness of this so-called “neural machine translation” is that the vector space it uses is merely morpho-syntactical and lacks semantic dimensions.

2.2 General limitations of machine learning

Major limitations of current deep-learning paradigms have been identified already (for example in [29]). They include first of all a set of quite general problems affecting stochastic models of any sort  –  not only deep neural nets but also traditional regression and classification approaches [18], including graph-based stochastic models (Bayesian Networks).

The first of these limitations turns on the huge data need of stochastic models, which may employ millions of parameters. Transformer, for example, has 213 million parameters and needs at a minimum billions of data tuples to become useful even for the sorts of rough translation produced by Google Translate. This limitation is already of considerable importance given that, leaving aside the resources of internet giants such as Google, there are few real-world examples of data available in the sorts of quantities needed to deal with complex outcomes using any sort of stochastic approach.

Second, all stochastic models require a stable environment. The quality of their output depends on how well they reflect the real-world input-output relationship they are aiming to represent. Where this relationship is erratic, there can be no good model (consider again the oil price example mentioned above). But even where the relationship is stable, the model will quickly become invalid if the input-output relationship changes on either side even in some minor way. This is because the model does not generalise. Once fed with data as input that do not correspond to the distribution it was trained with, the model will fail. And it will not alert the user, because it will not know that it is failing.101010So-called deterministic AI models [41] do not generalize, either, but they report their failures. This explains why stochastic spam filters and similar applications are so vulnerable to changing situations, and why they so often need re-training. And the more complex an application, the more demanding will be the re-training of its network that is required upon change of input constellations (for example when new types of sensors are introduced in driverless cars). The costs for such re-training will vary, of course, with the complexity of the input and the accuracy requirements of the network.

But there is a third group of limitations, turning on the fact that the output of all stochastic models is, by definition, approximative. Models of this sort can yield only the most probable output for any given input and model, and this output often falls below even the average human output. For many imaginable useful purposes, however, the output should be at least as reliable as the behaviour not of the average but of a qualified subset of human reference samples. This is very hard to achieve in language-focused applications using dNNs only. Unlike machines, humans are able spontaneously and immediately to attribute meaning to the world they experience. This is because the human species has evolved with a complex set of dispositions to react immediately in highly specific ways to specific sorts of external stimuli.

Human beings are, along many dimensions, tuned to the environments in which they live. The entities that we experience are spontaneously assigned meanings that reflect their relevance to our survival, meanings that are assigned using mechanisms hardwired into our brains. The belief that stochastic models can learn to make decisions without benefit of prior hardwiring of this sort is as naive as the old tabula rasa theories that were once the staple of empiricist philosophers and of their empirical psychologist followers. Such views were criticised by [15] in his ecological theory of perception111111Indeed they were criticised, 200 years earlier, by Immanuel Kant in 1781 in his Critique of Pure Reason., and they were experimentally refuted in the works on infant cognition of [4], [16], [23], [24], [25], who demonstrated that infants  –  and primates [38])  –  possess a large body of categorical and structural knowledge about the world of solid objects long before they even start acquiring the grammar of their mother tongue [28]. Indeed, it seems that language acquisition presupposes the working of a common set of ontological distinctions on the side of language learners, including the distinction between objects and processes, between individuals and categories, between natural and accidental properties of objects, and so forth.

Even the theory of Bayesian models for concept learning based on similarity acknowledges (i) the need for a prior genus-individual distinction to explain the mechanics behind generalization and (ii) the existence of a prior meta-heuristic linking membership in a class to property instantiation [46, 47]. As [39] formulates the matter, categorization relies on inferences about the causal role of putative essences in producing observable features. Such features are, in other words, merely secondary, derivative; and all the naive knowledge brought to bear by the infant follows from the natural and universal supposition that things belong to classes sharing similar properties [31, 44]. Even children as young as 3 believe that the ‘insides’ of objects are relevant in determining class membership [12, 14, 23]. According to [4] (p. 207), experiments on object recognition suggest that there is an object tracking system in the infant  –  a system that tracks three-dimensional, bounded, and coherent physical entities, and fails to track perceptually specified figures that have a history of non-cohesion. And what holds of infant cognition in general holds also of infant language learning and language competence in particular, where the capability of object tracking grounds the use of nouns and pronouns. Indeed, part of the background source of this empirical work on infant ontology was Chomsky’s work [6] on innate universal grammar. [13] make explicit reference to these ideas when they assert that they are able to “determine how languages and conceptual systems are constrained by examining the forms and meanings that children construct, and which errors they fail to make” ([13]; compare [33], p. 47).

For our purposes here, it is crucial that the AI applications running on today’s computers can simulate at best only small fragments of the hard-wired human capabilities revealed in such research. This means that they can simulate only small fragments of the semantics underlying human language use. As we shall see, neural networks do in this respect have limitations no less severe than those of traditional logic-based AI approaches to the modeling of human cognition. The formal ontologies used in the latter, however, allow us to overcome some of the limitations of the former because they involve direct representations of the sorts of objects, processes and attributes (and associated nouns, verbs and predicates) used by human beings in perceiving, acting and speaking. The training of neural networks is the attempt to build simulacra of relation-rich content of this sort out of gigantically large numbers of features represented using numerical input vectors or matrices. The training algorithms estimate what amounts to a very large polynomial (this is what a neural network is) with the help of an optimization procedure. But building the simulacra in this way seems to be infeasible even for simple ontologies of the RDF-sort made up of structures of the type: entity A – relates to – entity B [17].

2.3 Limitations applying specifically to deep neural networks (dNNs)

As humans process sensory input data, they assign meanings to the objects and events which the data represent (and from which the sensory content originates), experiencing these objects and events as belonging to a specific sort of categorical structure. dNNs, in contrast, do not use any of the target-derived properties of the input data that humans spontaneously use when they assign meaning to the data which they receive through experience. The result is a tremendous brittleness of dNN capabilities. [34] describe how high-performance neural networks developed for image classification can be nudged into a complete misclassification of images when the input material is mixed with a perturbation image. For example, what is at first correctly classified by the system as a flagpole is classified as a labrador after the system is very slightly perturbed. Perturbations of an analogous sort do not cause problems for humans at all. As [22] showed, dNNs used in image recognition work merely by learning certain surface-statistical regularities from images: the green grass that forms the typical background of a cow, for example, is contrasted with the grey of asphalt that forms the typical background of a car. They can be perturbed so easily precisely because they do not learn what the images are about and the sort of world to which the imaged objects and events belong.

The same holds also of the dNNs constructed for language processing purposes. A recent paper by [5] proves mathematically (which, given what was said above, we should in any case expect) that dNNs lack core computational features of traditional approaches to syntactic language analysis, for example, of the sort pioneered by [6] using probabilistic context-free grammars. As the authors show, while it is required of every valid stochastic model that it compute a valid probabilistic distribution, this condition is not in general satisfied by dNNs working from language input. But without this ability, there can be no computational representation of semantics. Thus, as shown in [9], the language constituents used by dNNs to make predictions in question-answering or textual entailment tasks often make no sense to humans at all.121212One example described in [9] rests on the input: “In 1899, John Jacob Astor IV invested $100,000 for Tesla to further develop and produce a new lighting system. Instead, Tesla used the money to fund his Colorado Springs experiments”. The described system correctly answers the question: “What did Tesla spend Astor’s money on?” with a confidence of 0.78 (where 1 is the maximum). The problem is that it provides exactly the same answer with a similar degree of confidence as its response to the nonsensical question: “did?”

This in turn means that dNNs, whatever it is that they are doing, cannot be modeling the semantics that need to be captured in order to extract information from texts, a crucial task in natural language processing for most automation purposes.

The information extraction (IE) results presented in [49] provide a poignant example of the low quality currently being achieved for tasks of this sort.131313The $F_{1}$-score of 0.52 reported by [49] seems quite high; but most of the training material is synthetic and the reported outcome only concerns information triples, which cannot be used for applied IE. The example is ‘poignant’ because the paper in question won the 2017 Prize for Information Extraction of the Association for Computational Linguistics, globally the most important meeting in the language AI field. This example reveals just how low the expectations in the field have become. The failure of dNN-based approaches to compute natural language semantics is illustrated also by the recent misclassification of the United States Declaration of Independence as hate speech by the Facebook filter algorithms.141414https://www.theguardian.com/world/2018/jul/05/facebook-declaration-of-independence-hate-speech

dNNs are also unable to perform the sorts of inferences that are required for contextual sentence interpretation. The problem is exemplified by the following simple example:

“The cat caught the mouse because it was slow” vs.

“The cat caught the mouse because it was quick.”

What is the “it” in each of these sentences? To resolve anaphora requires inference using world knowledge  –  here: about persistence of object identity, catching, speed, roles of predator and prey, and so forth. Thus far, however, little effort has been invested into discovering how one might engineer such prior knowledge into dNNs (if indeed this is possible at all).151515Currently, prior knowledge is used mainly for the selection or creation of the training data for end-to-end dNN applications. The result is that, with the exception of game-like situations in which training material can be generated synthetically, esp. in reinforcement learning, dNN models built for all current applications are still very weak, as they can only learn from the extremely narrow correlations available in just that set of annotated training material on the basis of which they were created. Even putting many dNN models together in what are called “ensembles” does not overcome the problem.161616The improvements provided by this approach are very modest and not higher than those achieved by other tweaks of dNNs such as optimised embeddings or changes in the layering architecture [27].

And worse: because the dNNs rely exclusively on just those correlations, they are also unable to distinguish correlation from causation, as they can model only input-output-relationships in ways that are agnostic to questions of, for example, evidence and causality. Thus they can detect that there is some sort of relationship between smoking and lung cancer. But they cannot determine the type of relation that is involved unless references to this very relation and to relevant types of relata themselves form part of the annotated corpus. Unfortunately, to create the needed annotated gold-standard corpora  –  one for each domain of interest  –  would be hugely expensive in terms of both time and human expertise. To make dNNs work effectively in language applications thus would require not only enormous collections of data but also, at least for many applications  –  and certainly for those examples involving the tracing of causality  –  the investment of considerable amounts of human expertise.

One final problem resulting from the properties of dNNs as very long polynomials is a lack of transparency and  –  in contrast to determinstic algorithms  –  a black box operation mode. Therefore, dNN engineers cannot tell why the network yielded its output from a given input. This poses a major challenge in areas where we need to reproduce or analyse the behaviour of the network, for example in case of disputes over liability.

Taken together, these problems rule out entirely the use of machine learning algorithms alone to drive mission-critical AI systems  –  for example with capability such as driving cars or managing nuclear power stations or intensive care units in hospitals. They are too brittle and unstable against variations in the input, can easily be fooled, lack quality and precision, and fail completely for many types of language understanding or where issues of liability can arise. Even at their very best, they remain approximative, and so any success they achieve is still, in the end, based on luck.

3 Making AI meaningful again

3.1 Adding semantics to automation solutions

To overcome these problems, ways need to be found to incorporate prior knowledge into the AI algorithms. One attempt to do this is to enhance Bayesian Networks with an explicit relationship semantics [26], which allows the model designer to build in knowledge describing entity relationships before using data to train the weights of these relationships. This reduces the learning effort on the part of the system by providing a rudimentary form of prior knowledge. But unfortunately, the expressivity of the resulting models is too low to represent the sorts of complex contexts relevant to human language understanding. Furthermore, they are not exact, secure, or robust against minor perturbations. They are also not transparent, and thus humans cannot reliably understand how they work to achieve given results. The goal of meeting this requirement is now dubbed “explainable AI”, and we will describe one promising strategy for achieving this goal that involves building applications that work in accordance with the ways humans themselves assign meaning to the reality that surrounds them. To achieve this end, we use a semantics-based representation that is able to deal with language as it is actually used by human beings. Importantly, the representation is able to incorporate prior knowledge based on low to medium amounts of input material of the sorts found in typical real-world situations. For humans in such situations find meaning not in data, but rather in the objects and events that surround them, and in the affordances that these objects and events support [15]. This implies a different sort of AI application, in the building of which not only mathematics and computer science play a role, but also philosophy.

Part of what is needed is to be found already in the early attempts to create ‘strong’ logic-based AI.171717An excellent summary can be found in [41]. For our purposes here, the most interesting example of an attempt of this sort is in the work of Patrick Hayes, a philosopher who first made his name with a paper co-authored with John McCarthy, commonly accredited with having founded the discipline of AI research. The paper is titled “Some Philosophical Problems from the Standpoint of Artificial Intelligence” and it lays forth for the first time the idea behind the calculus of situations [30]. In subsequent years Hayes set forth the idea of what he called ‘naïve physics’, by which he meant a theory consisting of various modules called ‘ontologies’, that would capture the common-sense knowledge (sets of common-sense beliefs) which give humans the capacity to act in and navigate through the physical world [19]. The theory is axiomatised using first-order logic (FOL) and Hayes proposed that something of the order of 10,000 predicates would need to be encoded in FOL axioms if the resulting theory was to have the power to simulate human reasoning about physical objects of the sorts that are encountered by humans in their everyday lives.181818Hayes’ conception of an ontology as the formalization of our knowledge of reality continues today in the work of Tom Gruber, whose Siri application, implemented by Apple in the iPhone, is built around a set of continuously evolving ontologies representing simple domains of reality such as restaurants, movies, and so forth. The problem with Hayes’ approach, as with strong AI in general, is that to mimic even simple human reasoning in real time would require a reasoning engine that is decidable, and this implies a severe restriction on the expressiveness of the logic that can be used. Standardly, one ends up with a very weak fragment of FOL such as that encapsulated nowadays in the so-called Web Ontology Language (OWL). OWL is restricted, for example, in that it can capture at most relational information involving two-place relations, and it has a similarly diminished quantifier syntax and a well-known difficulty in dealing with time-series data. For this and many other reasons, logic-based systems have rarely reached the point where they were able to drive AI-applications. They did, however, spawn the development of a huge body of mechanical theorem-proving tools [40], and they contributed to the development of modern computational ontologies, which helped to transform biology into an information-driven discipline [2]. Both of these developments are essential for the sort of AI applications combining formal logic and stochastic models that we describe below.

3.2 Inserting philosophy into AI

3.2.1 Desiderata for automated language processing

We will show in what follows how, by augmenting stochastic models (including dNNs) with philosophically driven formal logic, we can create AI applications with the ability to solve real-world problems. We present an example of such an application and describe how the machinery proposed is already in commercial production. First, however, we give details of what we take to be the minimal requirements which any real-world AI system must satisfy (Table 1). These requirements cannot be satisfied by agnostic machine-learning systems alone, as they presuppose the ability to deal with the semantics of human (natural) language. They can be satisfied, we believe, only by combining stochastic inference components with methods associated with traditional, logic-based AI in such a way as to allow incorporation of prior knowledge.

On the approach we shall describe, all the desiderata listed in Table 1 are satisfied on the basis of a formal representation of prior knowledge using a computable representation of the natural language semantics of the information the system is processing. To succeed, this representation needs two major elements: (a) a set of logical formalisms, constituted by formal ontologies that enable the storage and manipulation of language in Turing machines, and (b) a framework which enables one to define the meanings of the elements of the language.

We can describe only the rough outline of these components here, though one important feature, the methodology for development and use of ontologies to which we appeal, is described in detail in [1].

3.2.2 Representing natural language

Natural language as input is of course very hard to express in a logical framework, and a typical basic pipeline, which sacrifices a considerable part of the semantics in order to achieve computability, comprises the following elements:

morphological and syntactical error correction of an input text using dNN-models trained using large public data resources, 2. 2.

syntactical parsing with the help of a stochastic parser, e.g. a conditional random field parser as described in [10], and 3. 3.

inference applied to the parser outputs with the help of (computable) propositional logic.

However, to capture the semantics of full natural language, a much stronger, higher order intensional logic is required for its computational representation [11], and a logic of this sort cannot be used for computational purposes. To enable computation, the source text must thus be expressed by combining several computable logical dialects which together provide a representation that is adequate for a given context and purpose. For example, fundamental language constructs can be represented using FOL, while temporal relationships require temporal propositional logic. Intensional sentences can be represented using modal logic. A suite of such logical formalisms is able to achieve a good approximation to natural language semantics while still allowing computability. Computability does not, be it noted, require decidability, but robustness and completeness (in the technical logical sense) are essential. Decidability is not required because logical inference is fully possible with robustness and completeness alone. The absence of decidability certainly implies that on occasion a computation may not terminate. To take account of such cases, however, algorithms are stopped after a pre-determined maximum computation period whose length is defined by the qualitative requirements of the system. Such cases can then be handed to a human for decision.

The resulting system does not, however, strive for general AI. It always works in specific sub-domains covering segments of reality. The execution of the logics is then orchestrated with the help of domain-specific deterministic or stochastic controllers which are needed to ensure that inference steps are carried out in the order that is appropriate for the problem at hand.

3.2.3 Digression: Ambiguity and indeterminancy

A hybrid AI system along the lines described can also deal with phenomena such as ambiguity, vagueness and indeterminancy, which natural language frequently contains. We cannot give a full account of how these phenomea are dealt with here, but we provide two informative examples:

Ambiguous sentences, for example “Every man loves one woman”, which can mean that every man loves a woman, e.g. his wife (on the de dicto interpretation), or that every man loves one well-known woman, for example Marilyn Monroe (on the de re interpretation). To account for the ambiguity, an adequate logical representation creates two logical phrases from this sentence, and deduces the correct meaning from the context, for example using stochastic models. 2. 2.

Uncertainty, such as “John’s father did not return. Now John is searching for him.” Here, the transitive verb in the second sentence may or may not have an existing object. This phenomenon can be addressed using intensionality tagging for certain transitive verbs and subsequent contextual disambiguation (see paragraph 3.2.4.1).

3.2.4 Incorporating ontologies

Ontologies can be divided into two types. On the one hand are domain ontologies, which are formal representations of the kinds of entities constituting a given domain of inquiry and of the relations between such entities [43]. On the other hand are top-level ontologies, which represent the categories that are shared across a maximally broad range of domains  –  categories such as object, property, process and so forth. Each ontology is built around a taxonomic hierarchy in which the types of entities are related to each other by the relation of greater and lesser generality (an analogue of the subset relation that holds between the instances of such types). Domain ontologies have enjoyed considerable success in the formalisation of the descriptive content of scientific theories above all in many areas of biology (see especially the Gene Ontology, [2]), where they served initially as controlled, structured vocabularies for describing the many new types of entities discovered in the wake of the Human Genome Project. On the other hand are top-level ontologies such as Basic Formal Ontology (BFO, [1]), which arose to allow domain ontologies at lower levels to be created in such a way that they share a common set of domain-independent categories. As more and more such domain ontologies came to be developed and applied to the annotation and management of more and more different types of biological and biomedical data, the use of such a common top level allowed the resultant ontologically enhanced data to be more easily combined and reasoned over. BFO is now used in this way as shared top-level ontology in some 300 ontology initiatives.191919BFO is currently under review as an International Standards Organization standard under ISO/IEC: 21838-1 (Top-Level Ontologies: Requirements) and ISO/IEC: 21838-2 (BFO).

The use of a common top level also allows multiple ontologies to facilitate standardised exchange between parties communicating data about entities in different but overlapping domains. Through the incorporation of formal definitions, they also allow the application of basic inference mechanisms when interpreting data exploiting taxonomic and other relations built into the ontology. For logic-based AI applications, ontologies are needed which reflect the full spectrum of language constituents and of their logical counterparts. They must enable the expression not only of traditional taxonomical and mereological relations but also, for example, of synonymy relations at both the word and phrase level.

3.2.4.1 Resolving ambiguity

The terms in such ontologies are defined using formulae of FOL and the way these formulae represent language can be illustrated using the second uncertainty example we introduced in section 3.2.3 above.

*Natural language input: John’s father did not return. Now John is searching for him.

Formal representation: $\textrm{father}(x)\land\textrm{john}(y)\land\textrm{mod}(x,y)\land\lnot\textrm{Return}_{p}(x)\land\textrm{Searches}_{i}(x,y)$

Here unary predicates (nouns) and modifiers (mod) indicating syntactical relationships (in this case: possessive) are shown in lower case; non-unary predicates (verbs) are shown in upper case, and the subscript p indicates past tense. The subscript i indicates intensionality of the transitive verb, which is obtained from the temporal relationship of the two sentences and the fact that the presence-inducing verb of the first sentence is negated. The intensionality indicator might be used in the next step as disambiguation anchor, triggering a search in the subsequent text for a fulfilling object of the intensional predicate.

3.3 The core of the philosophy-driven machinery

In Appendix A, we give a detailed example of how our approach in building real-world AI systems combines automated transformation of text into logic with deterministic and stochastic models. In what follows, we describe the core functionality we developed to arrive automatically at a computational text representation, using logic that is semantically faithful to the input text.

3.3.1 Transforming text into logic

The process of transforming text into logic starts with stochastic error correction and syntactical tagging using dNN-Part-Of-Speach-taggers [20]. This output is used to perform a sequence of inferences, starting with:

[TABLE]

where ‘text’ is, for example, a sentence from a customer letter; $\leadsto$ means automated translation; and $\Gamma$ is the set of logic formulae202020This is a non-compact and non-complete k-order intensional logic; ‘k-order’ means that predicates of the logic can predicate over other predicates arbitrarily often. ‘Intensional’ means that the range of predication in the logic is not restricted to existing entities [11]. generated by the translation. The formulae in $\Gamma$ are generated with a proprietary AI-algorithm chain that uses world knowledge in the form of a dictionary of lexemes and their morphemes along with associated rules, relating, for example, to the transitivity and intensionality of verbs.

[TABLE]

where $\Delta$ is a collection of (first-order or propositional modal) logical formulae automatically generated ($\curvearrowright$) from $\Gamma$. This action transforms the non-computable formulae of $\Gamma$ into formulae expressed in logical dialects each of which enjoys compactness and completeness. The translation from $\Gamma$ to $\Delta$ requires world knowledge, for example about temporal succession, which is stored in the computer using ontologies.

[TABLE]

where $\vdash$ means: entailment using mechanical theorem proving, and $\phi_{i}$ is one of $n$ human-authored domain-specific formulae $\phi$ entailed by $\Delta$.

Unlike the automatically generated collections $\Gamma$ and $\Delta$, $\Omega$ is an ontology comprising human-authored domain formulae $\phi_{i}$. $\Omega$ is always related to a specific type of text (for instance, repair bills) and to a pertinent context (for instance, the regulations under which the repair occurs). The role of $\Omega$ is to express the target semantics that can be attributed to input sentences of the given type and context. $\Delta\cap\Omega\neq\varnothing$ (i.e. we can infer some $\phi_{i}$ from $\Delta$) holds only if the input text matches the type and context of the ontology.

In total, the process looks like this:

[TABLE]

where the only manual input is the creation of $\Omega$, and where this manual input itself needs to be performed only once, at system design time.

The Appendix below describes one example of how this approach is embedded already in a real-world AI production system.

4 Conclusion

As becomes clear from the example given in the Appendix below, our approach to philosophy-driven language AI is to generate a specific system for each application domain. There thus remains very little similarity to the hypothetical idea of general artificial intelligence. What we have is rather an exact, philosophy-driven context- and task-specific AI technology. Systems based on this technology are being successfully used in a range of different domains. Moreover, the method in question is generalizable to data of many different sorts, in principle  –  as the breadth of the available ontologies is extended and the sophistication of the algorithms is enhanced  –  covering more and more areas and domains of repetitive work of the sort amenable to automation. The pace of this extension to new domains will be accelerated by enhanced ontology authoring software as well as by support for semi-automated ontology generation, for example using inductive logic programming [35]. This will allow for applications such as automated encoding of medical discharge summaries, validation of the medical necessity of diagnostic and therapeutic procedures, and automation of customer correspondence.

We believe that these developments have implications beyond the merely technical (and, associated therewith, pecuniary). For they point to a new conception of the role of philosophy in human affairs which has been evolving since the end of the nineteenth century.

Beginning with the mathematician-philosopher Gottlob Frege, philosophers have been developing the methods which enable the expression in exact logical form of knowledge otherwise expressed in natural language. FOL itself was invented by Frege in 1879, and since then the FOL framework has been refined and extended to the point where it is possible to represent natural language in a formal, computable manner.212121An overview is given in [3].

Philosophers have, from the very beginning, attempted to understand how human language works and how language relates to the world. (Think of Aristotle’s Organon and Book VII of his Metaphysics.) In the 20th century, an entire branch of the discipline  –  called ‘analytical philosophy’  –  has grown up around this topic [8]. The computational discipline of ‘formal ontology’, has in recent years achieved considerable maturity in part as a result of the influence of philosophical ideas.

In spite of all this, however, there are many, especially in the twentieth century, who have proclaimed the death of philosophy, or who have seen philosophy as having a merely compensatory role in offering some sort of substitute for those traditions which, in former times, gave human beings the ability to interpret their lives as meaningful. The ways in which human lives are meaningful  –  are indeed full of meaning  –  did indeed play a role in our argument above. But we would like to draw a more far-reaching conclusion from this argument, drawing on the ways in which, beginning already with the Greeks, philosophers have helped to lay the groundwork for a series of social upheavals in the course of human history. These include, for example, the birth of democracy or of market institutions, of new artefacts such as Cartesian coordinates, and even of entire scientific disciplines. For we believe that one place where we can look for a role for philosophy in the future will lie in the way it can be used to strengthen and enable applied sciences in the digital era  –  for example, in the creation of useful and realistic artificial intelligence applications involving automatic translation of natural language texts into computer-processable logical formulae.

Appendix A Appendix: A real-world example

To represent in logical form the full meaning of complex natural language expression E as used in a given domain and for a given purpose, we will need a set of domain-specific ontologies together with algorithms which, given E, can generate a logical formula using ontology terms which are counterparts of the constituent simple expressions in E and which expresses the relations between these terms. These algorithms should then allow the representation in machine-readable form not merely of single expressions but of entire texts, even of entire corpora of texts, in which domain-specific knowledge is communicated in natural language form.

To see how philosophy is already enabling applied science-based production along these lines, let us look at a real-world example of an AI automaton used to automatically generate expert technical appraisals for insurance claims.222222This software is in production at carexpert GmbH, Walluf, Germany, a claims validation service provider processing some 70 thousand automobile glass repair bills a year with a total reimbursement sum of over **€**50 million. The bill validation process, performed by a car mechanics expert in 7-10 minutes, is completed by the system in 250 milliseconds. Today, such claims are validated by mid-level clerical staff, whose job is to compare the content of each claim  –  for example the line items in a car repair or cardiologist bill  –  with the standards legally and technically valid for the context at issue (also referred to as ‘benchmarks’). When deviations from a benchmark are detected by humans, corresponding amounts are subtracted from the indemnity amount with a written justification for the reduction. Digitalization has advanced sufficiently far in the insurance world that claims data can be made available in structured digital form (lines in the bill are stored as separate attributes in a table in a relational database). However, the relevant standards specifying benchmarks and how they are to be treated in claims processing have until recently been represented only as free-text strings. Now, however, by using technology along the lines described above, it is possible to automate both the digital representation of these standards and the results of the corresponding comparisons between standards and claims data.

To this end, we developed an application that combines stochastic models with a multi-facetted version of logic-based AI to achieve the following steps:

Compute a mathematical representation (vector) of the contents of the bill using logic for both textual and quantitative data. The text is vectorised using a procedure shown in equations (1) – (3) of section 3.3.1 above, while the quantitative content is simply inserted into the vector. 2. 2.

Recognise the exact type of bill and understand the context in which it was generated. This is done using the logical representation of the text, which is taken as input for deterministic or stochastic classification of the bill type (for example, car glass damage) and subtype (for example, rear window). 3. 3.

Identify the appropriate repair instructions (‘benchmark’) for the bill by querying the corresponding claims knowledge base for the benchmark most closely matching the bill in question. Standard sets of benchmarks are provided by the original equipment manufacturers or they are created from historic bills using unsupervised pattern identification in combination with human curation. The benchmark texts are transformed into mathematical logic. 4. 4.

Compare the bill to its benchmark by identifying matching lines using combinatorial optimisation. The matches are established by computing the logical equivalence of the matching line items using entailment in both directions: given a bill line (or line group) $p$ and its candidate match (or group) $q$, compute $p\vdash q$ and $q\vdash p$ to establish the match. 5. 5.

Subtract the value of the items on the bill that do not match the benchmark from the reimbursement sum 6. 6.

Output the justification for the subtractions using textual formulations from the appropriate standard documents.

To achieve comparable results an end-to-end dNN-based algorithm would require billions of bills with standardised appraisal results. Yet the entire German car market yields only some 2-3 million car glass damage repair bills in any given year and the appraisals are not standardised.

The technology is used for the automation of typical mid-level office tasks. It detects non-processable input, for example language resulting in a non-resolvable set of logical formulae, and passes on the cases it cannot process for human inspection. This is a core feature of our technology which may not match the expectations of an AI purist. However, applications of the sort described have the potential to automate millions of office jobs in the German-speaking countries alone.

Human beings, when properly trained, are able to perform the classification described under step 2 spontaneously. They can do this both for entire artefacts such as bills and for the single lines which are their constituents. Humans live in a world which is meaningful in precisely this respect.232323Compare footnote 6 above. The ability to classify types of entities in given contexts can be replicated in machines only if they store a machine-adequate representation of the background knowledge that humans use in guiding their actions. This is realised in the described system by means of ontologies covering both the entities to which reference is made in given textual inputs and the contexts and information artefacts associated therewith. The ontologies also incorporate formal definitions of the relevant characteristics of these objects, of the terms used in the relevant insurance rules, and so forth. The ontologies are built by hand, but involve a minimal amount of effort for those with expertise in the relevant domain (here: the contents of repair bills and insurance rules). These definitions are entailed by the bill and benchmark texts, and the latter are automatically processed into logical representations in the ontology framework without human interference.

Resulting system properties

This philosophy-driven AI application uses both stochastic models and parsers, as well as mechanical theorem provers. It meets the requirements listed in Table 1, including:

• •

Exactness  –  it has an error rate of below 0.3% (relative to the gold standard obtained by a consortium of human experts), which is below the best human error rate of 0.5%. Such low levels of error are achieved only because, unlike a stand-alone stochastic model, the system will detect if it cannot perform any of the essential inference steps and route the case to a human being.

• •

Information security  –  the system is secure because any misreactions to perturbing input by its stochastic models are detected by the logical model working in the immediately subsequent step.

• •

Robustness  –  it is robust since it will detect when it cannot interpret a given context properly, and issue a corresponding alert.

• •

Data parsimony  –  it requires very little data for training, since unlike the sorts of suboptimally separating agnostic spaces resulting from stochastic embeddings, it induces what we can call a semantic space that separates data points very effectively.

• •

Semantic fidelity  –  the system not only allows, but it is in fact based on inference and so it can easily use prior and world knowledge in both stochastic (Bayesian net) and deterministic (logical) form.

Acknowledgements

We would like to thank Prodromos Kolyvakis, Kevin Keane, James Llinas and Kirsten Gather for helpful comments.