for Reference
The question of what authors wish to convey through their work in general is an old and difficult question,1 but we cannot doubt their specific ability to successfully convey, which individual they are writing about. Our ability to comprehend a narrative involves keeping track of which character is which. There are about 31,000 verses in the whole Bible, and (from Aaron to Zurishaddai) about 2,000 characters. The biblical narrative would make little sense if we were unable to tell whether the same character was the subject of any two of those verses, or not. There are more than 700 occurrences of the proper name “Moses” in the Old Testament, and it is crucial to our ability to comprehend the work that we understand that these are not ambiguous names for 700 different people. Chastain calls such a set of names an anaphoric chain, namely “a sequence of expressions such that if one of them refers to something then all of the others refer to it.”2 The chain does not have to consist solely of proper names, but will normally be a mixture of proper names and other singular expressions.3 The question of how we resolve anaphoric chains is remarkably difficult, but it belongs to the science of computational linguistics, rather than philosophy.
Computational linguists traditionally distinguish co-reference from anaphor. Co-reference is when two terms refer to the same entity “in the world,” anaphora is when “a term (anaphor) refers [sic] to another term (antecedent) and the interpretation of the anaphor is in some way determined by the interpretation of the antecedent.”4 I reject this distinction, for I regard co-reference and anaphora as essentially the same phenomenon. To start with, we understand complete fiction, where all proper names and pronouns are empty, because we are able to bundle up singular terms into different anaphoric chains. Hence co-reference cannot depend on reference to the same entity “in the world.” The claim that an anaphor “refers to” another term is misleading for the same reason, as though the relation between a pronoun and its proper name antecedent were intralinguistic, but the relation between two co-referring proper names were not. As for the interpretation of the anaphor being determined by the interpretation of the antecedent, the idea seems to be that pronouns are essentially ambiguous, their sense (or reference?) determined by the immediate context, whereas proper names have a fixed and context independent reference. One dictionary defines anaphor as “an expression that can refer to virtually any referent, the specific referent being defined by context.”5 But this is not true. Proper names also are essentially ambiguous and require a context. Consider the name “Moses” in a book about the Pentateuch as contrasted with a book about Moses Maimonides. Clearly the rules determining proper name co-reference will be different from the rules determining pronoun co-reference, and I shall come to that, but this does not mean that co-reference is essentially a different phenomenon from anaphora.
Pronoun Resolution
The resolution of pronoun anaphora has received much attention in the literature, although the record is dismal. Hobb’s algorithm in 1978 was an early attempt.6 The algorithm starts at the NP (noun phrase) node immediately dominating the pronoun and searches in a specified order for the first match of the correct gender and number.7 The algorithm is purely syntactic; there has been some progress since the 1970s by using semantic properties of the term. For example, it seems as though a pronoun will not have a distant antecedent, and so entities introduced recently are more salient, and thus more likely to be the antecedent of back-reference, than those introduced earlier. Lappin and Leass have proposed that salience values should be cut in half each time a new sentence is processed,8 and that entities mentioned in subject position are typically more salient than those in object position. Centering theory, developed by Barbara J. Grosz, Aravind K. Joshi, and Scott Weinstein in the 1980s, proposes that discourse has a kind of center, which remains the same for a few sentences, then shifts to a new center. It is this center that is typically pronominalized in that there is a tendency for subsequent pronouns to take it as antecedent. Modern algorithms perform better than Hobb, but state of the art accuracy for general co-reference resolution is sadly quite low, in the range of 60 percent.
This is a puzzle, for it means we do not understand something that humans do easily, namely reading a simple story. Children soon learn to do this, yet the most advanced computational techniques fail. Does this mean there are no rules? Surely not. Either co-reference (1) is enabled by a sort of telepathy between author and reader, whereby the author telepathically communicates the intended reference to the reader, (2) involves some semantic relation between language and reality that a computer could not possibly emulate, or (3) is a property of the text, in which case, there must exist some method of decoding it. I rule out the first on the assumption that telepathy is impossible, particularly between an author who probably died in the second millennium BC and a reader in the third millennium AD. I rule out the second both as implausible, and because it is the principle target of this book. That leaves the third. I summarize the reasons supporting it here.
First evidence: rules exist. Ordinary grammar books talk about the error of faulty or vague pronoun reference, and specify a rule like “A pronoun should refer back to a single unmistakable antecedent noun.” To be sure, it is difficult to give a criterion or an algorithm for “unmistakable,” but as I said earlier, human readers clearly have the ability to keep track of which individual is which, without making mistakes, also human writers have the ability to enable this by clearly expressing their meaning. When mistakes occur, this is the fault of the writer, not the reader. Humans would not make rules if humans were unable to apply them.
Second, the difficulty that computers have with some exception cases is not that there is no rule, but rather that the rule requires knowledge of human affairs and customs. For example:
Gen 4:20 Adah gave birth to Jabal; he was the father of those who live in tents and raise livestock.
Gen 4:26 Seth also had a son, and he named him Enosh.9
Gen 12:18 So Pharaoh summoned Abram. “What have you done to me?” he said.
The first example requires knowledge that “Adah” is the name of a woman, not a man, which is provided in 4:19 (“Lamech married two women, one named Adah and the other Zillah”). This knowledge could be given to the computer through a list of all proper names classified by gender, but in any case, there is a further clue given by “gave birth to.” The computer would have to understand that only women can give birth. There is another difficulty: the verse could easily have been written “Lamech had a son, Jabal; he was the father of those who live in tents and raise livestock,” with no clue given by pronoun gender. To understand why “he” still refers back to “Jabal,” we would have to understand the purpose of 4:19–22, which is to tell us which occupations the descendants Lameth followed. Jabal was the ancestor of tent dwellers, Jubal of musicians, Tubal Cain of blacksmiths. This, in Chastain’s words, is a convention “relevant to the genre.”
Gen 4:26 is an example of a double pronoun use: “he named him Enosh.” The assumption is that the verb is not reflexive, otherwise the reflexive pronoun “himself” would have been used.10 We also know that human biology generally precludes children naming their parents. Gen 12:18 requires understanding of speech conventions. If “he” refers to Pharoah, and Pharoah uttered “What have you done to me?,” then “you” refers to Abram and “me” to Pharaoh. To understand that “he” refers to Pharoah, we have to understand what Abram is doing, which is explained by 12:17. God is inflicting plagues because Pharaoh has taken Sarai, Abram’s wife.
Third, the knowledge required is likely to be timeless and universal (or at least relatively stable over time, and across languages). We can generally understand texts written in ancient languages and by authors from very different cultures. Mark wrote, “Herod feared John, knowing that he was a just man” with the intention that “he” (αὐτὸν) should refer back to “John.” His intention was realized even though he wrote in Greek, and his English translators
