Homeric Meter

[jeidsath]

With some time free at Christmas, I spent the last two days writing a computer program to analyze Homeric meter. My main interest is to create a text with marked vowel length, correption, and synizesis, for reading aloud.

My program successfully analyzes all lines of the Iliad except for three (using van Thiel's text). Here are the three exceptions:

Π857 ον ποτμον γοοωϲα λιπουϲ ανδροτητα και ηβην
Χ363 ον ποτμον γοοωϲα λιπουϲ ανδροτητα και ηβην
Ω6 Πατροκλου ποθεων ανδροτητα τε και μενοϲ ηυ

ανδροτητα clearly resolves as long - long - short. So, following Monro, I change it to αδροτητα, and presto, the program is able to resolve it (it inserts an inverted breve beneath the mute-liquid/nasal pair that is read short, δρ):

ον ποτμον γοοωϲα λιπουϲ αδ̮ρ̮οτητα και‿ηβην
πατροκλου ποθεων αδ̮ρ̮οτητα τε και μενοϲ ηϋ

About 1000 lines have multiple results that I can't distinguish between easily. In fact, many more lines have multiple possible resolutions through synizesis, etc., but the program can usually score each result and choose the best. That approach breaks down on things like this:

Original:
Α276 αλλ εα ωϲ οι πρωτα δοϲαν γεραϲ υιεϲ Αχαιων
Multiple Results:
Α276 αλλ εα ωϲ οι πρωτα δοϲαν γεραϲ υιεϲ αχαιων
Α276 αλλ εᾱ‿ωϲ οι πρωτα δοϲαν γεραϲ υιεϲ αχαιων

Original:
Α363 εξαυδα μη κευθε νοωι ινα ειδομεν αμφω
Multiple Results:
Α363 εξαυδᾱ μη κευθε νοωι ινα ειδομεν αμφω
Α363 εξαυδᾱ μη κευθε νοωι ινᾱ‿ειδομεν αμφω

Cross-referencing (or some grammar awareness) will help for most of these, and I can mark the rest manually.

One decision that I made -- perhaps I was correct, perhaps not -- was to ignore diaeresis as marked by editors and to calculate it manually. By diaeresis I refer to the mark over a vowel, not the line break. My program will insert it if necessary, but generally prefers not to break up a diphthong. So here is the most highly scored resolution to the first line of the Iliad:

Original:
Α1 μηνιν αειδε θεα Πηληιαδεω Αχιληοϲ
Result:
Α1 μηνιν αειδε θεᾱ πηληιαδεω αχιληοϲ

The above scores -1 points because of the missing correption. Note that it is reading πηληιαδεω as πηλῃαδεω.

Now here is the textbook resolution. Note that I mark synizesis with a ring below the vowel that is not pronounced:

Points: -4.5
μηνιν αειδε θεᾱ πηληϊαδε̥ω αχιληοϲ

It scores -4.5 because in addition to the missing correption, there is a synizesis and a broken diphthong.

However Π686 argues for the diaeresis:

νηπιοϲ ει δε εποϲ Πηληιαδαο φυλαξεν

Also λ557 from the Odyssey:

ισον Αχιλληοσ κεφαληι Πηληιαδαο

The Πηλειδ- form does not need the diaeresis, but sometimes synizesis occurs:

Original:
Ο64 Πηλειδεω Αχιληοϲ ο δ ανϲτηϲει ον εταιρον
Result:
Ο64 πηλειδε̥ω αχιληοϲ ο δ ανϲτηϲει ον εταιρον

Also the Πηλεωϲ form:

Original:
Α489 διογενηϲ Πηλεωϲ υιοϲ ποδαϲ ωκυϲ Αχιλλευϲ
Result:
Α489 δῑογενηϲ πηλε̥ωϲ υιοϲ ποδαϲ ωκυϲ αχιλλευϲ

I will post on the Odyssey shortly.

[Paul Derouda]

Very nice! The best I ever achieved was a program I made in the 90's to drill irregular verbs in English, German and Swedish. (It worked! I still remember them: sit, sat, sat, sitzen, sitzt, sass, hat gesessen, sitta, sitter, satt, suttit...)

I'm slightly skeptical though that your program has got them ALL right except for these very few exceptions... There are so many complications, and not all lines are regular.

My program successfully analyzes all lines of the Iliad except for three (using van Thiel's text). Here are the three exceptions:

Π857 ον ποτμον γοοωϲα λιπουϲ ανδροτητα και ηβην
Χ363 ον ποτμον γοοωϲα λιπουϲ ανδροτητα και ηβην
Ω6 Πατροκλου ποθεων ανδροτητα τε και μενοϲ ηυ

ανδροτητα clearly resolves as long - long - short. So, following Monro, I change it to αδροτητα, and presto, the program is able to resolve it (it inserts an inverted breve beneath the mute-liquid/nasal pair that is read short, δρ):

The (or "a") modern interpretation is that these lines, or specifically ανδροτητα, are a hexameter fossil from time immemorial, from the time when Greek had syllabic consonants. According to this theory, the word used to be pronounced as a-nṛ-tā-ta, the r with a small dot below (ṛ) representing a syllabic consonant, i.e. pronounced as if it were a vowel. (compare "l" in English gentleman ˈdʒentlmən). But this doesn't tell us how "Homer", who knew nothing of syllabic consonants, actually pronounced the word.

About 1000 lines have multiple results that I can't distinguish between easily. In fact, many more lines have multiple possible resolutions through synizesis, etc., but the program can usually score each result and choose the best. That approach breaks down on things like this:

Original:
Α276 αλλ εα ωϲ οι πρωτα δοϲαν γεραϲ υιεϲ Αχαιων
Multiple Results:
Α276 αλλ εα ωϲ οι πρωτα δοϲαν γεραϲ υιεϲ αχαιων
Α276 αλλ εᾱ‿ωϲ οι πρωτα δοϲαν γεραϲ υιεϲ αχαιων

Original:
Α363 εξαυδα μη κευθε νοωι ινα ειδομεν αμφω
Multiple Results:
Α363 εξαυδᾱ μη κευθε νοωι ινα ειδομεν αμφω
Α363 εξαυδᾱ μη κευθε νοωι ινᾱ‿ειδομεν αμφω

Cross-referencing (or some grammar awareness) will help for most of these, and I can mark the rest manually.

So if I understand you correctly, the program doesn't know whether α ι υ are long or short, but deduces that from the context? It doesn't know that εᾱ imperative has a long α and ινα a short one?

It scores -4.5 because in addition to the missing correption, there is a synizesis and a broken diphthong.

Perhaps the rule should be the other way round – at least for some words, synizesis of εω is the rule, not the exception (I'm not sure about this, didn't look up the numbers). But this is the problem: in real language, things are not always so regular and creating scoring rules that take everything into account is difficult. For example, I think synizesis is more frequent with some vowels combinations and some words than with others.

None of this is means that I don't think that your program must be great and very helpful!

[jeidsath]

I'm slightly skeptical though that your program has got them ALL right except for these very few exceptions... There are so many complications, and not all lines are regular.

Oh, I didn't mean to say that it resolves them all correctly. The correct resolution is probably always there, but is not always the highest rated result. The program needs more awareness of grammar and a word database to improve accuracy. Rather, I should have said that the program had multiple possible resolutions for all lines of the Iliad except for those three (for which it had none).

Here's the output for the first 20 lines (only showing lines with the highest point totals):

μηνιν αειδε θεᾱ πηληιαδεω αχιληοϲ
ουλομενην η μῡρι αχαιοιϲ αλγε εθηκε
πολλαϲ δ ιφθῑμουϲ ψῡχᾱϲ αιδι προιαψεν
ηρωων αυτουϲ δε ελωρια τευχε κυνεϲϲιν
οιωνοιϲι τε πᾱϲι διοϲ δ ετελειετο βουλη
εξ ου δη τα πρωτα διαϲτητην εριϲαντε
ατρειδηϲ τε αναξ ανδρων και δῑοϲ αχιλλευϲ
τιϲ τ αρ ϲφωε θεων εριδι ξυνεηκε μαχεϲθαι
λητουϲ και διοϲ υιοϲ ο γαρ βαϲιληϊ χολωθειϲ
νουϲον ανα ϲτρατον ωρϲε κακην ολεκοντο δε λᾱοι
ουνεκα τον χρῡϲην ητῑμαϲεν ᾱρητηρα
ατρειδηϲ ο γαρ ηλθε θοᾱϲ επι νηαϲ αχαιων
λῡϲομενοϲ τε θυγατρα φερων τ απερειϲι αποινα
ϲτεμματ εχων εν χερϲιν εκηβολου‿ᾱπολλωνοϲ
χρῡϲε̥ωι‿ανα ϲκηπτρωι και‿ελιϲϲετο πανταϲ αχαιουϲ
ατρειδᾱ δε μαλιϲτα δυω κοϲμητορε λᾱων
ατρειδαι τε και‿αλλοι ευκνημῑδεϲ αχαιοι
ῡμιν μεν θε̥οι δοιεν ολυμπια δωματ εχοντεϲ
εκπερϲαι πριαμοιο πολῑν ευ δ οικαδ ικεϲθαι
παιδα δ εμοι λῡϲαιτε φιλην τα δ αποινα δεχεϲθαι

The first line has already been discussed. In line 3, πολλαϲ can't be marked without non-metrical information.

The following two lines, I think are a vindication of the possibilities of this program:

χρῡϲε̥ωι‿ανα ϲκηπτρωι και‿ελιϲϲετο πανταϲ αχαιουϲ
ῡμιν μεν θε̥οι δοιεν ολυμπια δωματ εχοντεϲ

So if I understand you correctly, the program doesn't know whether α ι υ are long or short, but deduces that from the context? It doesn't know that εᾱ imperative has a long α and ινα a short one?

Correct.

Perhaps the rule should be the other way round – at least for some words, synizesis of εω is the rule, not the exception (I'm not sure about this, didn't look up the numbers). But this is the problem: in real language, things are not always so regular and creating scoring rules that take everything into account is difficult. For example, I think synizesis is more frequent with some vowels combinations and some words than with others.

That’s very possible. With εω, synizesis seems to occur frequently (perhaps less than 30-40% of the time, I can’t say exactly until I’ve completed grammar/word-database features).

[Paul Derouda]

The following two lines, I think are a vindication of the possibilities of this program:
χρῡϲε̥ωι‿ανα ϲκηπτρωι και‿ελιϲϲετο πανταϲ αχαιουϲ
ῡμιν μεν θε̥οι δοιεν ολυμπια δωματ εχοντεϲ

Indeed!

But my guess is still that your program has interpreted these difficult lines incorrectly:

IL.3.172 αἰδοῖός τέ μοί ἐσσι φίλε ἑκυρὲ δεινός τε
IL.5.343 ἣ δὲ μέγα ἰάχουσα ἀπὸ ἕο κάββαλεν υἱόν
IL.5.359 φίλε κασίγνητε κόμισαί τέ με δός δέ μοι ἵππους

How about the first syllable in
IL.23.2 ἐπεὶ δὴ νῆάς τε καὶ Ἑλλήσποντον ἵκοντο

[jeidsath]

3.172:

αιδοιοϲ τε μοι εϲϲι φιλε εκυρε δεινοϲ τε
αιδοιοϲ τε μοι‿εϲϲι φιλε̱ εκυρε̱ δεινοϲ τε

Notice the macron underneath the short vowels. That's how the program marks them as long. It's easier to see what's going on if I give it the following input:

αιδοιοϲ τε μοι εϲϲι φιλε ϲϝεκυρε δϝεινοϲ τε
αιδοιοϲ τε μοι‿εϲϲι φιλε ϲϝεκυρε δϝεινοϲ τε

The highest rated result for the next is clearly wrong:

5.343:

η δε μεγα ιαχουϲα απο εο καββαλεν υιον
η δε μεγᾱ ιαχουϲᾱ ᾱπο ε̥ο καββαλεν υιον

However, here are the next results in terms of point rating:

η δε μεγᾱ ιαχουϲᾱ‿απο̱ εο καββαλεν υιον
η δε μεγᾱ ιαχουϲα απο̱ εο καββαλεν υιον

5.359:

φιλε καϲιγνητε κομιϲαι τε με δοϲ δε μοι ιππουϲ
φῑλε καϲιγνητε̱ κομιϲαι τε με δοϲ δε μοι‿ιππουϲ

23.2

επει δη νηαϲ τε και Ελληϲποντον ικοντο
ε̱πει δη νηαϲ τε και‿ελληϲποντον ικοντο

[Paul Derouda]

So basically, when your program scans a line, it tries all sorts of weird stuff like making ε long, but assigns that possibility a very low score, so that it is considered a likely possibility only when nothing better is found?

Did your program give αιδοιοϲ τε μοι‿εϲϲι φιλε̱ εκυρε̱ δεινοϲ τε as the most likely result even before you added the digammas?

[jeidsath]

That's the basic idea. It turned out to be more complicated in implementation. And yes, that was the result before adding digamma.