📄 Transitive and Intransitive Verbs in Akkadian: A Case Study of the Epic of Gilgamesh Using Natural Language Processing

Summary

The paper examines how Akkadian distinguishes verbs that take direct objects from those that do not. Its corpus is the Standard Babylonian Epic of Gilgamesh, with Tablet XI (the flood narrative) as the main focus. The author combines close philological reading of hand-annotated examples with computational processing. The Classical Language Toolkit (CLTK) and ORACC data support lemmatization, morphological tagging, dependency parsing and frequency statistics. Hopper and Thompson's (1980) scalar model of transitivity and the notion of verbal valency frame the analysis.

The main claim is that Akkadian transitivity is a gradient rather than a yes/no property. Clauses fall along a continuum shaped by the interplay of stem morphology (the G, D, Š, N and Gt forms plus further stems), aspect and mood with lexical meaning and the surrounding discourse. Quantitative tallies tie the degree of transitivity to stem, semantic class, subject type (gods, humans, natural forces) and narrative mode.

The study matters for Semitic grammar. It also shows that digital tools can supplement, though not replace, philological judgment on fragmentary ancient texts.

Key points

• Stem shares in Tablet XI (after George, 2003) are:
  • G 49.0%
  • D 18.4%
  • Š 12.2%
  • Gt 10.2%
  • N 6.1%
  • other stems 4.1%
• Each stem has a typical effect on valency:
  • Š (causative) introduces a new Agent, as when "rain" becomes "cause to rain".
  • D intensifies a verb or makes it transitive.
  • N and Gt demote the Agent or make the action reflexive or middle.
• Verb forms split sharply between direct speech and narrative:
  • Imperatives are 15% of verb forms in direct speech but only 2% in narrative.
  • Precatives are 12% in speech and 3% in narrative.
  • Preterites are 30% in speech and 45% in narrative.
  • Perfects are 8% in speech and 15% in narrative.
• Semantic classes differ in how often they appear in high-transitivity uses:
  • Creation/destruction verbs: about 85% high-transitivity.
  • Perception verbs (e.g. amāru): about 75% high-transitivity.
  • Stative verbs: about 15% high and 60% low.
  • Motion verbs: a balanced spread.
• Divine subjects rank above human ones on agency, affectedness and individuation. The gap is largest for how fully the object is affected and smallest for affirmation/realis. The available text reports no parser accuracy figures such as LAS.

Interesting ideas

• I1. Transitivity is scored as a bundle of separate parameters rather than read off the presence of an object. The parameters are participant count, telicity, volitionality, object affectedness and realis mode, so clauses can be ranked even when they share a syntactic frame. — For Mova (frames): attach a Hopper–Thompson style feature vector to each extracted frame, so downstream reasoning can tell how strongly an agent acted on a patient, not just whether an object slot is filled.
• I2. Morphological stems are treated as operators on argument structure. A causative adds an Agent, while a passive or reflexive removes or merges one, which gives a compact and predictable account of valency change. — For Mova (frames): model English causative, passive and middle alternations as explicit, explainable transformations of VerbNet role lists instead of storing each variant separately.
• I3. Verb-form distributions are cross-tabulated by discourse mode, showing that directive forms concentrate in quoted speech while past narration carries the storyline. — For Mova (world): use reported-speech boundaries and form type to decide which clauses advance the event timeline and which are requests or commands that should not be asserted as world facts.
• I4. Subject type (divine, human, natural force) is related to the typical degree of agency, so entity category becomes a predictor of how transitive an action will be. — For Mova (world): use ontology classes of entities as priors on agent plausibility, flagging role assignments in which an unlikely entity is cast as a high-agency actor.
• I5. Corpus counts are used to test traditional grammatical claims instead of simply assuming them, while the paper acknowledges the limits of noisy, fragmentary input. — For Mova (learn): routinely check hand-written grammar or repair rules against silver-data statistics, keeping only those whose predicted distributions actually show up.