Logic and Semantics: The Development Bottleneck of Artificial Intelligence

Logic and Semantics: The Development Bottleneck of Artificial Intelligence

If a person’s ultimate goal is strong AI, logic seems like an exceptionally attractive form of knowledge representation. The reason is that logic is broadly applicable. In principle, the same representation—the same logical symbolism—can be used to represent vision, learning, language, and of course any integration arising from them. In addition, it provides highly compelling theorem-proving methods for processing information.

For this reason, predicate calculus was the preferred form of knowledge representation in early AI. This logic is more expressive than propositional logic because it can “look inside” a sentence to represent its meaning. Take the sentence, “This store has a hat suitable for everyone.” Predicate calculus can clearly distinguish among three possible meanings of this sentence: “For every person, the store has a hat that fits them”; “The store has one adjustable hat that fits anyone”; and “The store has one hat (presumably folded up) large enough to fit everyone at the same time.”

For many AI researchers, predicate logic remains the first choice. The framework of CYC, for example, is based on predicate logic. The representations used in compositional semantics for natural language processing (NLP) are as well. We extend predicate logic to represent time, causation, or duty and morality. Of course, this depends on someone having already proposed those forms of modal logic, and that is by no means easy.

However, logic also has its drawbacks.

The first drawback involves combinatorial explosion. A widely used logical theorem-proving method in AI is resolution. Conclusions derived through this method may themselves be correct, but they may have nothing to do with the target conclusion. Heuristics are used to guide and limit inference, and to decide when to stop proving. But these methods are not foolproof either.

The second drawback is that resolution theorem proving assumes that not-not-X implies X. This idea is familiar enough: proof by contradiction begins by assuming that a proposition is false—that is, by negating the conclusion of the original proposition—and then deriving an obvious contradiction, from which one concludes that the assumption is false and the original proposition is therefore proven. If the domain being reasoned about is completely understood, this is logically valid. But users of built-in resolution procedures—such as those in many expert systems—often assume that if no contradiction can be found, then no contradiction exists, the so-called “negation as failure.” This is often a mistake. In real life, proving that something is false is not at all the same as failing to prove that it is true—as when one is wondering whether a partner is being unfaithful. There may still be much unknown evidence, that is, hidden assumptions.

The third drawback is that in classical logic (“monotonic reasoning”), once something has been proven true, it remains true forever. In reality, this is not always the case. We may have good reason to believe that X is true—perhaps as a default assignment, or even as the result of careful argument or persuasive evidence—but later we may discover that X is no longer true, or was never true in the first place. If so, we must revise our beliefs accordingly. For logic-based knowledge representation, this is easier said than done. Inspired by McCarthy, many researchers have tried to formulate forms of “non-monotonic reasoning” that can tolerate changing truth values. Similarly, various kinds of “fuzzy” logic have been defined, in which statements can be marked as possible/impossible or unknown rather than simply true/false. Even so, no reliable way of overcoming monotonicity has yet been found.

Some AI experts, in studying logic-based knowledge representation, have increasingly wanted to identify the basic elements of knowledge or meaning. But they were not the pioneers. McCarthy and Hayes were already doing this in their co-authored paper, “Some Philosophical Problems from the Standpoint of Artificial Intelligence.” Many of the issues discussed in that article are familiar enough: from free will to counterfactuals. These problems involve the basic ontology of the universe: states, events, properties, change, actions... what exactly?

Unless someone is deeply committed to metaphysics—which is a rather rare passion—why care about ontology at all? Why is there now growing interest in these mysterious questions? The answer is clear: if one is trying to design strong AI, one must consider what kind of ontology knowledge representation can use. We must also confront these questions when designing the Semantic Web.

The Semantic Web is different from the World Wide Web. Since the 1990s, we have had the World Wide Web. The Semantic Web is not even the current state of the art in technology; it is the future state of the art. If the Semantic Web exists—and when it truly does—machine-driven associative search will be improved and supplemented by machine understanding. In that case, applications and browsers will be able to access any information on the internet and reasonably integrate different contents in the course of solving problems through inference. This formidable task has been guided by Sir Tim Berners-Lee, and one might even say it is extremely demanding: it requires not only major engineering advances in hardware and communications infrastructure, but also that web-crawling programs develop a deeper understanding of what they are doing.

Search engines such as Google, and NLP programs more generally, can usually find associations among words or texts, but they do not possess understanding. This is not philosophical understanding, but something empirical—another obstacle to strong AI. Although some examples sound tempting, they are ultimately misleading, such as IBM’s Watson, Siri, and machine translation. Today’s computers do not know what the things they “read” or “say” actually mean. One sign of this lack of understanding is that different programs cannot really communicate with one another, or learn from one another, because they use different knowledge representations or different basic ontologies. If Semantic Web researchers could find a universal ontology, then enabling machines to understand what they receive might no longer be pure fantasy. Thus, the metaphysical questions raised in AI in the 1960s have now become highly important because of their practical relevance.