Hello GK

Luis C. Reyes' Reply to "GK" on GK's Article on SHMERON in Luke 23:43

posted to Robert Bowman's Evangelicals and JW's discussion board, March 15, 2005

Click here for GK's Article on Luke 23:43 (as of March 15, 2005).

Click here for a dialog between Robert and "MS" on John 20:28 in which MS used a similar reliance on statistics to argue against the traditional view that Thomas calls Jesus both Lord and God. Luis's comments on GK's methodology apply equally to that of MS.

______________________________________________________

Hello GK,

I have read your post with interest, since it relates to two things that I have been personally studying: (1) The search for a theoretical framework that can explain every instance of disambiguation in natural language processing, and (2) highlighting the limitations of statistical approaches to disambiguation, as has been observed in disambiguation attempts in the AI (artificial intelligence/computational) literature with natural language processing.

Its seems that your approach takes the statistical approach to disambiguation for the interpretation of Luke 23:43. I have several comments to make regarding your post. First, let me say that I consider Luke 23:43 to be an ambiguous passage with at least two possible candidate senses:

(1a) truly I say to you today, you will be with me in paradise

(1b) truly I say to you, today you will be with me in paradise

It is obvious here that merely decoding the linguistic string is not sufficient to obtain the speaker’s intended meaning (highlighting one of the weakness of the code model of communication), and thus inferential processing is required. In attempting to ascertain the speaker’s intended meaning when dealing with ambiguity and the process of disambiguation, those interested in psycholinguistic/AI/ computational linguistics and pragmatics, usually ask:

(2a) What determines the order in which the candidate senses are tested? (What determines the accessibility of possible disambiguations?).

(2b) What pragmatic criterion is used in evaluating candidate senses and accepting or rejecting them? (What determines the acceptability of possible disambiguations?).

The psycholinguistic and AI literature has been mostly interested with (2a), and not very interested with (2b). As an interpreter, I have been particularly interested in (2b) (as those from the pragmatic spectrum are), since the point for me is to understand what the speaker wanted to communicate in the first place. From my studies I find that Relevance Theory, as advocated by Dan Sperber and Deirdre Wilson, can provide an adequate theoretical framework that explains how the disambiguation process takes place in human communication. However, it is not my purpose to show this here (I am working on doing this somewhere else), but rather my purpose here is to comment on some of the ideas that you have presented in your post.

First, you speak of this rule and provide some parameters:

“When the Greek adverb SHMERON takes the second position to a verb in a separate sentence of direct discourse it always further modifies the verb in the first position, without exception, in the corpus of the Greek Septuagint and Greek New Testament.”

I think I understand what you are saying here, but what does this mean? What is your conclusion? Is it (a), are you merely pointing out and describing what you have linguistically observed in a corpus, or (b), are you saying that by using this statistical method and that by applying this rule, that this is an adequate tool to disambiguate the interpretation of Luke 23:43? I think you are saying (b) since you further argue that interpretation (b) is mandated by the results of applying your rule.

Well, I have a problem with this conclusion (b), and the problem is that it is founded on an inadequate method of interpretation for natural language possessing when dealing with ambiguity (and many other features too). The statistical approach to disambiguation, as you are displaying, is founded on the code model of communication. While human communication can obviously use codes to communicate, the model does fall extremely short (reference resolution, ambiguity, implicit information and other things are all problems for this model). The statistical method for disambiguation uses regularities of co-occurrence frequencies in a corpus to predict the “best bet.” For example, in dealing with lexical disambiguation a computer may be able to take the word “spy” when it co-occurs with the word “bug” from a corpus, and then take the interpretation of a sentence such as, “the spy checked the room for bugs,” and provide the interpretation of “bug” as most likely to be “a hidden microphone” (not an insect). This is precisely the method that you are utilizing with this so-called rule when you attempt to deal with the structural ambiguity in Luke 23:43. You say,

This is fine if you are merely suggesting notion (a) above (as a grammar would note, since grammars are supposed to be descriptive and not prescriptive), however, if you are also suggesting (b) there is a problem. The problem is (as has often been noted in the AI literature) that there are so-called “garden path” utterances (it is called such because once you get to the most accessible interpretation, which is based on the most accessible meaning from a selected corpus, you then later discover that it was the incorrect interpretation for a particular passage after all). “Garden path” utterances show that the most accessible sense (taken form a majority of situations from a certain corpus) is not always correct for the interpretation of a particular linguistic string in every occasion (this obstacle alone destroys the statistical approach), and what is needed is a pragmatic criterion for accepting or rejecting tentative hypotheses (in order to deal with 2b). If your approach to disambiguation were correct, then the problem of ambiguity with machine translation would have been solved; if you have solved this problem I guarantee you that you will be a millionaire!

At any rate, before your conclusion (in b) can be accepted (at least by me), you first would have to provide your readers with evidence that your statistical method of disambiguation is an adequate theory for disambiguation (when applying it to all problems with disambiguation in natural language possessing). An adequate theory of disambiguation must explain on what grounds tentative disambiguations are accepted or rejected (it must adequately deal with 2b above). This of course, is something that the statistical approach has not been able to do (or perhaps not very interested in doing). To substantiate this method and approach, you would first have to show that that every case of proper disambiguation can be dealt with by rules that automatically integrate properties of the context with semantic properties of the utterance in order to obtain the proper interpretation. This simply has not been proven and this is precisely one of the reasons why machine translation has so many problems with disambiguation: it can't tell you what it means, but it can tell you what it can possibly mean. Besides, the AI literature doesn't aim to show how disambiguation actually happens, but merely to approximate its results. It may program a machine to pick the “best bet,” which must then be checked by a human. This of course, would put you back to square 1 again, and therefore still be left with the problem of (2b). Once again, an adequate theory for disambiguation should be able to explain (2b) (at least for those of us who are interested in interpretation).

That your method is not a reliable method is clearly apparent when you are forced to restrict your rule to: (1) direct discourse, and (2), situations where the adverb, which follows the verb, must not belong to a different sentence or syntactical unit.

Now, if your statistical method of interpretation for disambiguation were correct, then why are these restrictions placed? If this method of interpretation is reliable, it should be able to explain WHY some interpretations are selected over others in ALL cases.

Well, just on the face of it, I see that even from a statistical approach, your corpus has been selective for your investigation. First of all, in Luke 23:43, you have genuine case of ambiguity where two possible candidate senses are possible. I have not looked at all of your examples, but I don't take you to mean that every single instance of your example of the rule also displays ambiguity in the same manner as Luke 23:43. If you are going to exclude indirect discourse because you want the example to resemble Luke 23:43, then why not also be consistent and only consider the passages that display structural ambiguity in the same manner as Luke 23:43? This would make it an accurate investigation of the actual situation in Luke (if that is what you are aiming at). Also, it seems that you have entirely overlooked the speaker's idiolect (the linguistic system of an individual speaker), have taken and mixed everyones idiolect from the LXX and the NT, and then forced it into Jesus’ own unique idiolect in order to obtain the interpretation of Luke 23:43. Now that is certainly contaminating your search, why not restrict the corpus to only the idiolect of Jesus and only to cases where structural ambiguity is present in order to further resemble Luke 23:43? So you see, while you arbitrarily place certain restrictions on your rule you are not consistent with accurately and fairly duplicating the situation with Luke 23:43. Now even if you did all these things, it still would not be enough. You see the fact that you may find a frequent co-occurrence of some linguistic feature in a selected corpus is definitely not enough to tell you what a particular interpretation is for an utterance that was uttered in a specific and particular situation by a particular speaker, and for a particular reason. The underlying problem with such approaches to disambiguation is that it overly relies on the code model of communication, and as I have stated already, this is why it has become something that is so troublesome to those who attempt to use a statistical approach to disambiguation in the AI literature (although there has been some progress made in that area).

Whatever the case, it should be noted, that my concerns have to do with the method of interpretation that you have suggested for the disambiguation of Luke 23:43. I believe that such an approach is unwarranted for dealing with problems of structural ambiguity. Naturally, when dealing with cases of ambiguity, I will question whatever interpretative conclusion is drawn if it is based on a statistical method that has not been able to resolve problems involving the disambiguation of ambiguous utterances.

Just my comments on this, take them for what they're worth,

-Luis Reyes

____________________________________

This discussion continued with GK's post # 17270, in which GK promised to review Luis' comments carefully. Luis responded with post 17272. These and possible follow-up posts may be found here.