📄 Considering a resource-light approach to learning verb valencies

Summary

The paper asks how little linguistic infrastructure is enough to learn verb valency, meaning which arguments a verb takes and how often, for morphologically rich, under-resourced languages. The target is Quechua, and the intended consumer is L3, a rule-based dependency-grammar MT system. Quechua has no treebank to check results against, so the author tests the idea on Arabic. Running text is first passed through a morphological analyzer. Only sentences containing a single verb are kept, and case-marked nouns are counted as likely arguments using a few language-specific heuristics. The frames obtained this way are then compared with frames read off the Penn Arabic Treebank. The comparison is negative. Single-verb sentences have a different mix of verbs and leave out clausal complements entirely. On top of that, unvowelled Arabic script makes each word highly ambiguous. The author concludes that the realistic minimum is higher than planned: a POS tagger and a chunker, and for Arabic also a morphological disambiguator. This is a useful warning for anyone planning valency extraction with very few resources.

Key points

• Languages and tools: Arabic, processed with AraMorph (a Free Software Buckwalter analyzer) and checked against the Penn Arabic Treebank. Quechua, processed with Gasser's AntiMorfo on a small AVENUE project corpus.
• Arabic is highly ambiguous: about 7.5 Buckwalter analyses per word on average (σ = 8.4), and one word reaches 86 analyses. Quechua averages 1.7 (σ = 1.3, maximum 10), mainly because its orthography writes vowels.
• Restricting to single-verb sentences skews the data in two ways:
  • The most frequent verbs differ from those in the corpus as a whole.
  • Typical uses vanish. "kAn" takes a verb-phrase argument about 400 times in the treebank, and "qAl" mostly takes an SBAR clause, but neither use can appear in one-verb sentences.
• For Quechua, the analyzer often cannot tell whether a word is a verb at all. For example, waqaychu could be a verb, a noun or an infinitive.
• Two open problems remain even with better tools:
  • Negative evidence: a frame that never shows up in a medium corpus is not thereby impossible.
  • Proposed remedy: active learning, where the system generates test sentences and crowd workers judge whether they are grammatical.

Interesting ideas

• I1. Before trusting a cheap extraction method in a language with no gold data, run it on a language where a treebank exists and compare against frames from that treebank on the same text.
• I2. A data filter chosen for convenience, such as keeping only one-verb sentences, can quietly change what is being measured, so compare the kept subset with the full data.
• I3. Report ambiguity as mean, standard deviation and maximum number of analyses per token, as a quick diagnostic of how much downstream steps depend on disambiguation.
• I4. Treat valency as a frequency profile rather than a yes/no property: "eat" usually takes an object, while "put" almost always does.
• I5. To deal with missing negative evidence, generate probe sentences that use a verb in a suspected frame and ask a judge whether each one is grammatical.

For Mova

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

• I1 (learn): Any shortcut silver-data pipeline could first be scored against treebank-derived frames on identical sentences, and adopted only if the two distributions agree.
• I2 (learn): Silver-data filters, for instance keeping only sentences where all parsers agree, should be checked for skew in verb and construction distribution before the subset trains anything.
• I3 (tokens): Logging these counts per lemma or genre in the en tagger would show where morphology leaves too many candidates for the parser to resolve.
• I4 (world): Compiled verb-class knowledge could store graded frame probabilities next to VerbNet frames, so frame building can prefer likely readings and still allow rare ones.
• I5 (learn): An LLM teacher tier could act as the grammaticality judge for generated probes, filling gaps in valency data that corpus counts cannot settle.