Abstract:
The talk will elaborate on the importance of language resources in the digital age and the case of the less resourced languages.
A variety of technologies will be addressed and the status of these for Arabic(s) and the associated languages will be mentioned.
The importance of dialectal varieties will be discussed as well as the new use cases that involve code switching, arabizi input, etc.
The conclusion will focus on suggestions for a roadmap to compile a repository of such resources.

Abstract:
A system that can 'understand' natural language must be able to carry out inference over sentences of natural language. If someone said "I've just produced a system that can understand English", you would expect it to be able to recognise that if you said "John and Mary have just got divorced" then you would also be committed to saying that "John and Mary were once married" and "John and Mary are not now married"; and that if you said "I'd forgotten to to feed the cats" then you would be admitting that you hadn't fed the cats, whereas if you said "I'd forgotten that I fed the cats" then you would be saying that you had fed them. That's what understanding language means -- that when you hear a sentence, you can link it to your knowledge of the meanings of the words that it contains and work out its simple, obvious consequences.
Traditionally, this task has been approached by translating from natural language into some kind of formal logic, and then using a theorem prover for that logic. This has proved to be extremely difficult. The standard way of doing it involves carrying out syntactic analysis using a hand-written grammar, and then using semantic interpretation rules attached to the rules of the grammar to produce a formal paraphrase (a 'logical form'). Hand-coding grammars for natural language is very difficult; writing parsers that can apply such grammars in a reasonable amount of time is very difficult; and freely occurring text contains strange extra-grammatical constructions which will defeat even the best hand-coded grammars anyway.
Recent approaches to syntactic analysis, then, tend to be data-driven, with either no explicit grammar at all (Malt, SyntaxNet) or with VERY large sets of context-free rules that no-one could reasonably annotate with interpretation rules. But if you have no grammar rules to attach interpretations to, then how can you do compositional translation into a logical form? This talk will discuss recent work on carrying out inference directly on parse trees, and will show how pairwise matching algorithms, of the kind addressed in textual entailment and natural logic, can be embedded in a variant on a standard inference engine to carry out deep reasoning over natural language texts. The talk will be illustrated with examples from both English and Arabic.