📄 Automatic Identification of Arabic expressions related to future events in Lebanon's economy

Summary

The paper addresses how to automatically find statements about future economic events in Arabic web texts about Lebanon. The authors built a corpus from Google searches. The query keywords were taken from the government's "Economic and Social Reform Action Plan 2012-2015", and simple length-based heuristics pulled the main article text out of each downloaded page. They then catalogued the Arabic surface patterns that signal the future: particles, the future prefix, predictive participles, and verbs of expectation. These patterns were encoded as rules for SLCSAS, their Perl tool based on the Contextual Exploration method. Ambiguous candidates (words starting with س and words after قد) are first checked with "AlKhalil Morpho Sys" to confirm they are verbs. Every sentence containing a matched expression is labelled a future sentence. On 200 randomly chosen files that were annotated by hand, the system found all 743 true future sentences, with about 97.5% precision. The work shows that a transparent, marker-driven approach can reliably detect future-oriented statements in Modern Standard Arabic news, a genre where this topic had received little study.

Key points

• Corpus: over 1900 Arabic URLs were collected from the first two Google result pages per query (up to 100 URLs each). More than 1,600 pages yielded usable economic content. Each file stores the URL, the page title and the extracted article.
• The rules cover seven classes of future expression:
  • قد followed by a present verb
  • the prefix س on a verb
  • لن
  • سوف
  • predictive passive participles (e.g. متوقع, مرجح, مستبعد), optionally preceded by من الـ
  • past-tense predictive verbs (توقع, استبعد, ارتقب)
  • present-tense predictive verbs (يتوقع, يرجح, …)
• SLCSAS rules are ordered chains of positive and negative markers that map to a semantic category. A negative marker found in its search window cancels the match. Output is HTML with the markers highlighted, and the tool is released under the GNU GPL.
• Results: 36,816 sentences were found across the whole corpus, of which 5,535 were flagged as future. In the 200-file evaluation set (4,634 sentences), 762 were flagged and 19 of these were wrong. Recall was 100% and overall precision was 97.50%; per-class precision was 94.11 for قد and 97.19 for س.
• Two main error sources were identified:
  • قد used with epistemic, non-future verbs (e.g. "may be due to").
  • Proper nouns that look like س plus a verb (e.g. سيمون), and وسوف used as a family name.

Interesting ideas

• I1. Matching proceeds in ordered steps, and each matched marker shrinks the search area to the text after it, within a set word window. This keeps rules strict about order yet still easy to read.
• I2. Negative markers act as vetoes: if a blocking word occurs within its window, the candidate is dropped. This is a cheap way to suppress known false-positive contexts.
• I3. The cheap surface trigger is applied first, and only its candidates are sent to a morphological analyzer to confirm they really are verbs. The costly check therefore runs on just a small set of words.
• I4. Each future expression is classed as predicted, negated (لن) or ruled out (مستبعد), rather than simply tagged as future.
• I5. Recall was measured by having an expert exhaustively list every true positive in a random sample of files, rather than only checking the system's own outputs.

For Mova

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

• I1 (frames): modality and speech-act cues (e.g. "is expected to", "ruled out") could be detected as ordered marker chains over the UD tree, each with a stated scope that serves as its own explanation.
• I2 (frames): veto markers could stop "may" or "will" from being read as future or prediction in idioms like "may be due to" or in noun uses of "will".
• I3 (tokens): ambiguous forms flagged by fast patterns (e.g. "will", "may", "-'s") could be routed to deeper tagger or lemma checks only when a downstream rule depends on them.
• I4 (world): events in the text world could carry a status such as expected, excluded or asserted-future, so they are not treated as facts in the event order.
• I5 (learn): rule-based extractors could be evaluated on exhaustively annotated random samples, and false positives traced to specific rules, so that recall claims are honest.