📄 Constructing an English Valency Lexicon

Summary

The Prague English Dependency Treebank (PEDT) needs a verb valency lexicon to support its deep-syntactic (tectogrammatical) annotation, and the authors set out to build one. This treebank re-annotates the Wall Street Journal part of the Penn Treebank within the Functional Generative Description (FGD) framework. No existing resource could be converted automatically at the required quality, and building one fully by hand would be too costly, so they bootstrapped EngValLex from PropBank. XSLT templates turned PropBank material into frames shaped like the Czech PDT-VALLEX lexicon. Hand-written rules then guessed FGD functors from PropBank argument labels, and annotators corrected the output in a dedicated editor. Each frame lists functor slots and uses a compact notation for their permitted surface forms, which can describe whole small dependency subtrees. Frames also keep back links to PropBank. The work was still in progress, with roughly 80% annotated and no formal evaluation yet. The resource matters because these links allow PropBank annotation to be projected onto PEDT, which supports Czech–English machine translation research.

Key points

• The automatic conversion from PropBank produced 8,215 frames for 3,806 verb lemmas, stored in XML close to the PDT-VALLEX format. Phrasal verbs such as give up get their own entries.
• FGD distinguishes five inner participants and more than 40 free modifications. The inner participants are ACT, PAT, ADDR, ORIG and EFF, and they cannot repeat within a frame. ACT and PAT are assigned by syntactic criteria, and the remaining participants by semantic judgment.
• Slot forms combine Penn Treebank tags and lemmas, with abbreviations that can be defined recursively. For example, n = NN:NNS:NP:NPS. Bracket nesting describes subtrees, as in ACT(.n) PAT(to.TO[VB]) LOC(at.IN).
• Back links are kept at several levels:
  • frames link to PropBank rolesets,
  • slots link to PropBank arguments,
  • examples link to PropBank examples.
• The editor is derived from the PDT-VALLEX tool and includes offline PropBank and VerbNet viewers. The lexicon was split into 1,992 files by roleset name so that annotation could be distributed. Inter-annotator agreement was planned for a subset.

Interesting ideas

• I1. Instead of writing a lexicon from scratch, convert an existing resource automatically and let humans fix only the rule output, keeping links to the source at every level.
• I2. A single source sense that mixes uses is split into separate frames: the inchoative "door opened" maps the patient-like argument to ACT, while the causative keeps Arg0→ACT and Arg1→PAT.
• I3. Each slot's permitted surface form is written as a small dependency-subtree pattern (lemma.tag with nested daughters) using reusable, recursively defined abbreviations.
• I4. Role labels are assigned by a fixed ordering rule: one participant becomes ACT, two become ACT and PAT, and only further participants need semantic judgment.
• I5. Because the lexicon keeps links to PropBank, existing PropBank annotation can be projected automatically into the new formalism, which yields training data without new manual work.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): Verb-class knowledge compiled from VerbNet/FrameNet could carry per-slot provenance links, so every role in a frame can be traced back to the source entry it came from.
• I2 (frames): When mapping VerbNet roles onto UD arguments, treat causative/inchoative alternations as distinct frames so that the subject's role is not assumed to be fixed.
• I3 (frames): Frame slots could be matched against UD trees through compact subtree patterns like to[VERB], which keeps matching explainable and lets common patterns be shared.
• I4 (frames): Such a deterministic fallback could label arguments of verbs missing from VerbNet, with the choice marked as rule-based in the explanation.
• I5 (learn): Mappings between role inventories could turn existing PropBank-style annotations into silver frame data to be checked against Mova's own parsers.