Polysemy in a Conceptual System
PI: Achim Stein
B5's central aim is to construct in-depth semantic analyses of subclasses of French polysemous predicates - mostly verbal phrases, but also nominal phrases - that account for the way their event inferences (entailments vs. defeasible implicatures) vary in context, and to account for the lexical, supra-lexical and sub-lexical factors that influence these inferences/ the disambiguation process. The classes of predicates under investigation all pertain to two related domains, causation and agency. In the first two funding periods, B5 concentrated on verbal ambiguities that get resolved through 'outer aspect' or through the 'theta role of the subject' (Agents vs. Causers). In collaboration with the other B projects, we ventured beyond these lexical/supra-lexical levels and addressed the word-internal syntax and semantics of the predicates under investigation.
Phase 3: 2014-2018
Researchers: Fabienne Martin, Ingrid Falk
In phase 3, we will combine further the lexical and sublexical perspectives. We choose to focus on predicates which do not fully specify the 'degree of realization of the event' they (or their infinitival/sentential complement) describe and/or the 'degree of affectedness of the Theme'. What we called `defeasible causatives' in phase 2 form a subkind of these predicates. We aim to build a full-fledged cartography of verbs displaying ambiguities of this kind in French, including accomplishments with non-culminating construals, verbs with conative construals and agentive verbs with non-agentive meaning extensions. With regard to those among these verbs that leave the theta-role of the subject underspecified (i.e. allowing the spectrum of roles between full Agents and pure Causers), the hypothesis we aim to test is that the more the external argument is associated with `agenthood' properties, the smaller is the degree of realization of the event/the degree of affectedness of the Theme entailed by the verb. Additionally, we will deepen our studies on the contribution of affixes to the syntax/semantics of verbs under consideration. During phase 2, we found that in Modern French, verbs affixed with 'a-', 'en-', 'dé-', 'é-', '-ifier' or '-iser' (i) tend to be systematically transitive whereas the corresponding non-affixed verbs regularly have intransitive construals and (ii) strongly tend to form marked (reflexive) anticausatives, while the corresponding non-affixed verbs often form unmarked anticausatives. In phase 3, we pursue research goals with regard to these tendencies, namely a) accounting for the exceptions we identified in MF, and b) testing to what extent these tendencies hold in the specific class of French psych verbs, whose reflexive readings are well-known to be particularly difficult to pin down. For the successfull fulfillment of this research program, we will provide lexical resources, including an enhanced version of the database 'Les verbes français' (LVF), labelled `NLPe-LVF'. Additionally, B5 aims to deliver a competitive parser for MF, and the first dependency parser for OF.
Phase 2: 2010-2014
Researchers: Fabienne Martin, Paul Bédaride (2010-2011), Nicolas Mazziotta (2011-2014)
In phase 2, B5 pursues goals in both theoretical and practical aspects of lexical and ontological semantics. In particular, we will elaborate a typology of the inferences and presuppositions triggered by verbs that is more fine-grained than the classical one, and investigate how the automatic treatment of language can benefit from this typology. Moreover, hyponymy relations in the verbal domain -- which have so far been treated in an undifferentiated way as troponymy -- will be investigated from a theoretical perspective and implemented in a description logic formalism. In addition, B5 will intensify ongoing work on the identification and modelling of understudied cases of systematic polysemy in the verbal domain. In particular, this work will focus on illocutionary verbs (e.g. menacer, 'to threaten'; in cooperation with B4), criterion verbs (e.g. voler, 'to steal'; in cooperation with B4), and manner and result verbs (e.g. frapper, 'to hit'; in cooperation with B1, B4 and B6). Here, the underspecification of asserted, implied or presupposed events as well as the inferences they allow wrt. the aforementioned typology are of particular interest.
With respect to application and evaluation, B5 plans to contribute to research on natural language inference in French, in order to pave the way for a shared Recognising Textual Entailment (RTE) challenge analogous to the highly competitive one for English. French has not received much attention in connection with RTE, and thus, the work of B5 in phase 2 will mainly consist of preparatory steps. These include the compilation of a data set based on the typology of inferences and hyponymy relations, which can then be used by a wider community in order to e.g. apply statistical methods to French data (see the cooperation with D4), as well as proof-of-concept implementations of inference mechanisms based on description logics and Horn logic. This system will not perform a strict bottom-up construction of semantic formulae from natural language, but instead involve both formal lexical-semantic and ontological representations as well as context features like tense and polarity. Nevertheless, it will also feature principled work on sortal disambiguation (in collaboration with B3 and B4), with a view to developing a large-scale system in a potential third phase.
Phase 1: 2006-2010
Researchers: Fabienne Martin, Dennis Spohr
The relation between the meaning of a word and its context can be considered in two ways: from the speaker's point of view, we deal with lexical selection, and from the hearer's perspective, we deal with meaning selection. In both cases, the selection process is contextually constrained, mainly by the participants to the situation, which the speaker encodes and the hearer decodes. From both perspectives, an abstract concept is specified. Independently of cognitive questions (does a choice between different concepts takes place before the lexicalisation process? Do we deal with an incremental specification process?), the definition of a conceptual system made of abstract concept types and methods of specification is relevant in several respects: firstly for the modelling of facts independent of specific languages and secondly for the interface between linguistic and extra-linguistic knowledge. Taking as its starting point the lexical-semantic approaches working with meaning classes, the project investigates possible ways to organise and represent the relevant conceptual information in a conceptual system, with the goal that as much as possible of the resulting structures will be kept independent from a specific language. Contrary to traditional lexicalist description of discrete word meanings, which are most of the time differentiated too finely for practical or computational applications, such a conceptual system allows to generalise over specific concepts, as ontological principles regulating the dependence relation between the concepts of the system are inserted. On the linguistic side, one should provide a definition of the semantic categories and relations which are relevant for such a representation (participants to the situation, aspectual and presuppositional properties), and integrate this information in a structured conceptual system, as well as a consistent description of polysemous relations and of meaning vagueness. On the ontological side, one should evaluate and modify the structuring principles of the ontology --- for instance, the identity criteria for situations should be refined --- in order to represent the concepts of lexical items as instances of entities of a conceptual system. At the level of the ontology, the accent is put on abstract domains which are until now considerably undifferentiated, like the domains of events, states, and non-material entities as referents of verbs or abstract nominals. The main result consists in a conceptual system, whose format and metalanguage fits the ontological standards and which will be evaluated through its interface with French lexical entries. In this system, the specification of meaning in context amounts to the selection of potenial entities of the conceptual system.