Research activity in 1998
During the 1998, the research on Natural Language Processing has concerned the following topics:
- Formal syntax and cognitive architectures for Natural Language Processing:
This research aims at developing formal tools for describing dependency grammar, and to define cognitive models of human language processing based on the operational constraints posed by the formal system.
In 1998, we have extended the formal system developed in 1997 in order to cope with difficult linguistic phenomena, in particular the ones concerned with coordinate constructions. We have shown that with suitable extensions, the existing formal model can be adapted to cope with unit coordination (Mario ha incontrato Luigi e la sua ragazza – Mario met Luigi and his girlfriend) and gapping (Carla ha mangiato pollo e Lucia pizza – Carla ate chicken and Lucy pizza). We have also shown how the parsing algorithm based on Earley's schema (which was defined in 1997) can be modified in order to parse grammar definitions including coordinated constructs. Finally, we have evaluated the expressive power of the resulting formal system..
Currently, we are investigating the possibility of translating a grammar of English from a Lexicalized Tree Adjoining Grammar framework (the XTAG system) into a Dependency grammar, in order to test the parser on a wide-coverage grammar.
We have also proceeded on the integration between the formal defintition of the grammar and the hierarchical representation of subcategorization frames, showing how the latter can be used to improve the parsing algorithm. In particular, the study carried out in 1997 has provided us with a first nucleus of a hierarchy, and the inheritance mechanism has been applied to represent information in a compact way.
As far as cognitive architectures are concerned, we are evaluating different approaches to the problem of selecting the most promising syntactic analysis in presence of ambiguity. In particular, we have extended the environment developed in 1997, enabling it to exploit left-corner information, in order to test the performances of race-based parsing.
In the area of the cognitive modelling, the collaboration with the Centre for Cognitive Science of the University of Edinburgh (Dr. Patrick Sturt), partially funded by a CNR-BC joint research programme, has proceeded, as well as the collaboration involving the CNR Psychology Institute in Rome (Dr. Marica De Vincenzi).
- Lexical semantics
In this area, we have studied the problem of representing locative expressions. In fact, they involve interesting phenomena, that have received much attention in the recent literature. Our approach was based on a network based representation, exploiting the inheritance mechanism of hierarchies in order to follow the ideas arisen in the recent studies on underdetermination. In other word, we have assumed that, in the interpretation of a sentence including a locative expression, it is not necessary to obtained a fully disambiguated reading. On the contrary, each locative marker (e.g. preposition) has a single interpretation, which mat be furtherly “specified” on the basis of contextual knowledge. More precisely, each marker has a “set” of top level interpretations (representing true ambiguity); under each top-level meaning, a tree of “more specific” senses is defined. In this way, we are able to keep apart homonymy and polisemy.
- Recognition of a speaker's plans in flexible user interfaces
The reseach in the area of plan recognition has followed the direction set by the group’s activity in the prevous years, trying to extend it in different ways.
In particular, new research lines are based on the notion of BDI (Belief, Desire, Intention) agent previously defined to model NL dialogue in cooperative environments, the main idea being that speakers are agents endowed with domain and communicative goals that direct their behaviour. Moreover, since agents use the same procedures to select and execute speech acts and domain actions, the linguistic behaviour has been modelled in a uniform way with respect to domain behaviour. The two-level agent architecture is characterized by a declarative plan-based representation of the knowledge about acting, both at the object level (linguistic / domain actions representing the actions that the agent has at disposal to reach his goals) and at the metalevel, (agent modelling plans describing the knowledge about how to build and execute plans). The metalevel plans describe the behaviour of a reactive planner and take object actions as their objects. Thanks to the presence of a declarative representation of actions and plans, it is possible to exploit this representation both to interpret the behaviour of an observed agent, and to generate the agent behaviour.
In particular, by representing each interactant as an instance of an agent, interactional behaviour has been interpreted in terms of the (metalevel and object level) plans they are carrying on. In cooperative interactions, in fact, the turns of each interactant contribute to the satisfaction of some contextual goals previously established by the partner, or to the continuation of the agent’s own activity towards his communicative or domain goals. Based on this idea, communication and domain goals shared by the interactants have been used to model the coherence of a dialogue. More specifically, beside the Plan Continuation relation widespread in the plan recognition literature for modelling the stepwise execution of a plan, two main coherence relations were used, that derive from the Goal Adoption notion introduced by Castelfranchi and Parisi. Within this framework of dialogic cooperation, several dialogic phenomena were explained, like adjacency pairs, insertion sequences, presequences.
At each turn, from an interactant’s point of view, the coherence is fulfilled if one of the above mentioned relations can be recognised between the new turn and the interactional context, composed of the previous turns of the interaction. Since the system maintains two separate representations of the two participants – each with his own representation of the context (context model) – it is possible to model the disparities in knowledge and interpretation that can cause misunderstanding. In fact, these representations remain normally aligned during the interaction, to a reasonable degree, unless a misunderstanding occurs.
When a coherence link cannot be identifie, the participant who has detected the lack of coherence, due to his cooperativity, is obliged both to signal the potential occurrence of a misunderstanding and to work in order to solve it. Based on the fact that misunderstandings normally last for at least one turn – since, by definition, they are normally undetected when they appear - an algorithm for the identifying misunderstandings has been proposed, that exploits the notion of coherence to make hypotheses about the wrong coherence relations between misaligned interpretations of previous turns. Moreover, this algorithm can be exploited to detect and solve both the misunderstandings that are triggered by ambiguities and by interpretation mistakes due to disparities between the interactants’ beliefs. This work relies on the study of a corpus of examples of misunderstanding, that allowed to highlight the interactional development of misunderstandings in natural dialogues and the resources that human interactants exploit to detect and solve them.
Other apparently unrelated linguistic phenomena, like overanswering, grounding phenomena, acknowledgments, clarification subdialogues and repair turns fit in the framework depicted above, since they all contribute, in different ways, to promote the alignment between the participants during the interaction, and to solve the disparities that may have occurred between them.
The research on dialogue has been also exploited within the project “Dialoghi cooperativi ad iniziativa mista” with the CSELT (Centro Studi e Laboratori Telecomunicazioni) company. The project resulted in the proposal of a new architecture for a dialogue system for telephone information services based on speech recognition and speech generation technologies. With respect to the previous architecture, the proposed one constitues an extension toward modularity, that provides more flexibility to the dialogue system, in different ways. First, thanks to a more accurate, tree-like representation of the interactional context and of the underlying information task, it allows the user to take the initiative to a higher degree than he was allowed before. Then, the system is now able to react to user initiative, that results in a more natural interaction, in line with what emerged from the analysis of natural language information dialogues between human interactans.
1998 Publications
Book Chapters
L.Ardissono and G.Boella: “An agent model for Natural Language Dialog Interfaces”. In Lecture Notes in Artificial Intelligence, n. 1480: Artificial Intelligence: Methodology, Systems and Applications, 14-27, Springer Verlag, 1998.
V. Di Tomaso, V. Lombardo, L.Lesmo: “A computational Model for the Interpretation of Static Locative Expressions”. In Olivier, Gapp (eds.) Representation and Processing of Spatial Expressions, 73-90, Lawrence Erlbaum Ass., 1998.
International Journals
L. Ardissono, G. Boella, and R. Damiano. A plan-based model of misunderstandings in cooperative dialogue, Int. J. of Human-Computer Studies, 649-679, 1998.
International Conferences and Workshops
C. Barbero, V.Lombardo, L.Lesmo, P.Merlo: "Integration of Syntactic and Lexical Information in a Hierarchical Dependency Grammar", Proc. Workshop on Processing of Dependency-Based Grammar, 58-67, Montreal, 1998.
G. Boella and L.Lesmo: “Automatic Refinement of Linguistic Rules for Tagging”. Proc. of 1st. International Conference of Language Resources and Evaluation, 923-930, Granada, 1998.
V. Lombardo, L.Lesmo: "Formal Aspects and Parsing Issues in Dependency Theory", Proc. COLING-ACL’98, 787-793, Montreal, 1998.
V. Lombardo, L.Lesmo: "Unit Coordination and Gapping in Dependency Theory", Proc. Workshop on Processing of Dependency-Based Grammar, 11-20, Montreal, 1998.
Other publications
L. Ardissono, G. Boella, and L. Lesmo: “An agent architecture for NL dialog modeling”. Proc. Second Workshop on Human-Computer Conversation, Bellagio, Italy, 1998.
L. Ardissono, G. Boella, and L. Lesmo. Plan based agent architecture for interpreting natural language dialogue. Technical Report, 1998. (PostScript)
Leonardo Lesmo is the chairperson of the interest group on Natural Language Processing of AI*IA the Italian Association for Artificial Intelligence.
The NL group has actively contributed to organize, in cooperation with the University of Bari, the workshop on “Dialogo e Modelli di Utente nelle Interfacce Intelligenti” (Dialogue and User Modeling in Intelligent Interfaces), held in Padova on 24 September 1998.
