A question I've been interested in through my career is the extent to which humans fully recover the meaning of expressions: i.e., is everything that we hear or read fully interpreted, or are we in some cases recovering only part of its meaning?
My initial work on this question (Poesio, 1991; Poesio, 1993a; Poesio, 1994a; Poesio, 1996) was motivated by doubts about the extent to which scope disambiguation takes place-- i.e., whether and how humans decide the respective scope of scope-taking operators (negation, modality, quantifiers) in sentences such as
1. In most democratic countries, most politicians can fool most of the people on almost every issue most of the time.
Humans do not seem to be very good at identifying the alternative interpretations of such sentences; are they really assigning to them the interpretations proposed in formal semantics?
The starting point for my work was the idea of underspecified logical form introduced in NLP research such as (Schubert and Pelletier, 1982) as a way of characterizing the space of possible semantic interpretations of a sentence without actually computing them. I hypothesized that, in some cases at least, the final result of semantic interpretation might be an underspecified logical form (I was among the first to make a proposal along these lines); and concentrated on trying to identify what semantics these representations might have - i.e., what inferences we would expect humans to be able to draw from them ( Poesio, 1991; Poesio, in preparation ).
In recent years, I focused on finding empirical evidence concerning the role of underspecification in various aspects of language interpretation, using a variety of methods ranging from the analysis of disagreements in annotation to behavioral experiments to neuroimaging, focusing first on anaphora, and more recently, on logical polysemy.
A problem that has particularly interested me over the years is that of disagreements in interpretation unearthed during our work on anaphoric annotation (Poesio and Reyle, 2001; Poesio et al, 2006). There are many anaphoric expressions on whose interpretation people disagree without finding them ambiguous. An example are pronouns like IT in the following example,
2. Can you kindly hook up engine E3 to the boxcar at Elmira and send IT to Corning as soon as possible please?
Over the years we identified several cases of such disagreements in interpretation, first in the TRAINS dialogues, then during the annotation of the Vieira/Poesio corpus (Poesio and Vieira, 1998 and of the GNOME corpus. I studied this phenomenon for several years in a number of projects.
In work with Uwe Reyle, Frank Keller, Rosemary Stevenson, and Patrick Sturt funded by the Royal Society, I first used corpus analysis in the attempt to identify cases in which interpretations are truly left underspecified ( Poesio and Reyle,1999; Poesio and Reyle, 2001). We then used magnitude estimation experiments to test whether in fact one could argue that such anaphoric expressions were left uninterpreted ( Poesio, Keller, Reyle, Stevenson and Sturt, 2001 ) and formulated the Justified Sloppiness Hypothesis, specifying the conditions according to which ambiguous anaphoric reference was not considered unfelicitous (Poesio, Reyle, and Stevenson, 2007)
This work led to the ARRAU project, funded by EPSRC - a three year project investigating the extent to which certain types of anaphoric expressions are actually resolved (Poesio and Artstein, 2005a; Poesio and Artstein, 2005b; Poesio et al, 2006). The desire to carry out a more systematic investigation of the extent of ambiguity in anaphora was one of the key motivations for the subsequent development of the Phrase Detectives Game-With-A-Purpose for anaphoric annotation (Poesio et al, 2013) (see also the page on my interests in anaphora).
This work is currently being pursued in the DALI project, funded by the European Research Council (ERC). DALI involves five different work packages. In WP1 we are developing a number of Games-With-A-Purpose to collect multiple judgments about anaphora and mentions. In WP2 we are developing methods to analyse and aggregate these judgments, using, in particular, probabilistic aggregation methods. In WP3 we are using these methods to deliver datasets to study anaphora and disagreement in anaphora. In WP4 we are using these data to study the phenomenon of disagreement in anaphora from a linguistic and psycholinguistic perspective. In WP5 we are using the data to develop models for anaphoric interpretation taking disagreements into account.
The best-known example of ambiguity that does not result in multiple interpretations being generated are the cases of logical polysemy studied, among others, by Pustejovsky (1995) in linguistics and by Frazier and Rayner (1990) and Frisson (2014) among psychologists. The term logical polysemy refers to the fact that words like book, record, painting, film, etc. have multiple, but systematically related interpretations: in sentence 3., the concrete sense of the book is referred to, whereas in 4., the abstract (informational) sense of is. Yet there is plenty of evidence from linguistics and psychology that book is not ambiguous: rather, this word-form is associated with a single underspecified interpretation that is further specified in context.
3. I burned the book
4. I memorized the book
In work with Yuan Tao from the University of Trento Center for Mind Brain Sciences (CIMEC) and Andrew Anderson (formerly of CIMEC, now at the University of Rochester) we collected brain evidence suggesting that indeed the interpretation of such words works differently from the interpretation of words like bank.
After developing my first proposals concerning the underspecified interpretation of scopally ambiguous sentences (Poesio, 1991). I quickly came to realize that a cognitively plausible theory of disambiguation had to take into account the results of psychological work on ambiguity processing, which emphasized the role of incrementality (Poesio, 1994a; Poesio, 1996c). Unfortunately, very little work had been done to study the respective roles of incrementality and underspecification in scope disambiguation. So, for my dissertation, I ended up relying quite a lot on corpus data (my own analysis of the preferred scopal assignments in the TRAINS corpus; see below); and on evidence about underspecified interpretations coming from other areas of semantic processing where more psychological evidence was available, such as lexical disambiguation and anaphoric processing (Poesio, 1996b; Poesio, in preparation ). I also arrived at the conclusion that the `disjunctive' semantics proposed in my earlier work (and equivalent to the semantics later proposed by Reyle) was not a good characterization of such partial interpretations.
In later work ( Poesio, 1996b; Poesio, 1996a; ) and in particular in my forthcoming CSLI monograph (Poesio, 2003 ) I propose a theory of utterance processing and semantic underspecification that, from the semantic point of view, is based on the idea - closely related to ideas developed in work on syntactic underspecification - that the interpretations produced by the language processor are descriptions of the properties of utterances and of how utterances fit together with other utterances in hierarchical structures. The idea that disambiguation is a form of defeasible reasoning also plays a central role in the theory; defeasible logics are used to define the notion of precisification (disambiguated interpretation) that in other theories of underspecification is incorporated in the semantics of the language.