📄 Annotating events, Time and Place Expressions in Arabic Texts

Summary

The paper tackles a practical gap in Arabic NLP: there are few free annotated corpora, lexicons or taggers. The authors want to mark events, time expressions and place expressions in raw Modern Standard Arabic text without any of these resources. Their pipeline starts with an unsupervised, knowledge-free segmentation algorithm (Aliane, 2011) that splits each word into left affix, morpheme and right affix. A small hand-written rule set, built from the discovered affixes plus two contextual cues, then assigns a partial noun/verb tagging. On top of this, any verb not on a modality list becomes an event. Time and place spans are found through marker words that either stand alone or trigger a scan of the neighbouring words. The event model follows the conceptual scheme of Saval et al. (2009): an event takes place at a location and has a date. On 30 unvowelized web articles, the system found 168 of 268 verbal events and reached F-measure 84% for time and 45% for place. This shows that cheap surface rules can produce a usable starter resource.

Key points

• Partial POS tagging uses four affix-based rules (one for nouns, three for past/present/future verbs) and two contextual rules. One rule covers the particle sawfa; the other tags a word as a noun when it follows a closed list of prepositions and exception particles. Unresolved words stay ambiguous (<L>).
• Events: every lexie tagged as a verb (<LV>) is labelled <event>, after a filter removes verb-like modality words. Only verbal events are handled; person names are not.
• Time expressions use three kinds of markers:
  • stand-alone adverbial markers (e.g., morning, tonight, summer);
  • M1 triggers such as "since" and "when", which always sit at the right edge and scan leftward (Rule1);
  • M2 triggers such as "day", month names and "minute", which scan in both directions (Rule2).
  • Helper sets D1 (before/after), D2 (middle/beginning) and D3 (numerals) widen the span.
• Place expressions use one leftward rule (Rule3) driven by directional and spatial markers such as north, behind, inside and between.
• Evaluation covered 30 MSA web articles tagged with <event>, <Timex> and <Pl>. The system found 168 of 268 events, with F 84% for time and 45% for place. The authors attribute the errors mainly to verbs and nouns left untagged by the partial tagging step.

Interesting ideas

• I1. Span boundaries are found by growing outward from a trigger word and stopping at explicit barriers: punctuation, a verb, an unresolved word outside the marker set, or a very short particle. This yields readable, deterministic extents without any learned chunker.
• I2. Triggers are split into edge-anchored ones (one scan direction) and position-free ones (both directions). Encoding where a head word can sit inside its phrase reduces overreach.
• I3. Tagging is deliberately partial: words without clear evidence stay ambiguous rather than being forced into a guess. That ambiguity then acts as a signal (a stop condition) for later rules.
• I4. The authors plan a feedback loop in which recognized time and place spans help tag the remaining nouns and verbs, after which the whole pass is rerun.
• I5. The event is modelled as a minimal schema (event → location, event → date) taken from a domain ontology, which tells the extractor exactly which slots to look for.

For Mova

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

• I1 (frames): Temporal and locative adjunct spans could be delimited this way as a cheap, explainable fallback when the UD tree for an oblique is uncertain.
• I2 (tokens): Time and place cue lexicons in crate en could carry a direction flag, so a rule-based time/place tagger extends spans only in the linguistically valid direction.
• I3 (learn): Silver data could keep an explicit "unresolved" tag where witness parsers disagree, and downstream rules could treat it as a boundary instead of trusting a forced label.
• I4 (learn): Self-induced repair rules could similarly use high-confidence frame or adjunct detections to correct tagger errors, iterating until the annotations stop changing.
• I5 (world): A small fixed where/when slot schema attached to every event node would make ordering events and answering "where/when" questions straightforward and traceable.