Vers une annotation sémantique à large échelle : réconcilier lisibilité et modularité des représentations sémantiques

In semantic modeling we are interested in accurately capturing the meaning of natural language in a structure that can be processed automatically. This is interesting from a theoretical perspective, but such representations have a variety of downstream applications too, such as automated reasoning, machine translation or question answering. Numerous semantic representation formalisms have emerged over the past decades. Older logic-based ones aim to represent meaning precisely by encoding various phenomena and complex interactions such as quantifier scope, but are cumbersome to process and difficult to annotate and analyse by humans. Newer formalisms aim to resolve this latter issue through a more abstract representation, which allows for large-scale annotation and processing, but at the cost of foregoing many phenomena. The latest developments in semantic modeling show a trend towards attempting to reconcile these two views, highlighting the issue that semantic modeling faces -- finding a balance between capturing precisely and in a detailed enough manner the meaning expressed in natural language, while maintaining a representation that is simple enough to allow for large-scale annotation and processing. An important question in the design of meaning representations is what phenomena exactly to include. To address this and to evaluate the coverage of existing formalisms, we carry out a survey of semantic representation formalisms across multiple semantic phenomena and semantically-oriented considerations. We include global considerations such as scalability and universality, specific semantic phenomena such as temporality, quantification and negation, and considerations oriented towards the interfaces of semantics with syntax and pragmatics, as well as generation and analysis, which are important for applications. With the above considerations in mind, we propose a new semantic representation formalism which aims to combine the expressivity of logic-oriented representations, a wide range of semantic phenomena, and the readability desired for a representation that is easy for humans to annotate and analyse. Our proposal is for a structure that represents each semantic phenomenon as a layer, which allows not only for the clear separation between phenomena, but also for the ``switching off'' of layers that are not relevant for the user's task, making the representation more flexible and readable. Finally, practical issues in the field are the discrepancy between the amount of data available for different formalisms and the relative lack of parallel datasets, both of which limit the possibility for empirical comparison of formalisms. While automatic semantic parsing offers a way to produce meaning representations directly from text, current methods do not guarantee the precision of the resulting representation and therefore their use for tasks such as reasoning. With this in mind and a view of exploiting the relative difference in dataset availability, we carry out three rule-based graph transformation experiments between different pairs of formalisms to explore the viability of using the representations in one formalism to produce those of another. Our results show that this may be a useful approach for some pairs of formalisms, especially more closely related ones. Furthermore, the process of building such transformation systems is in itself useful in identifying empirically the differences between formalisms.