Jon Stevens

The Pursuit of Word Meanings.

We evaluate here the performance of four models of cross-situational word learning: two global models, which extract and retain multiple referential alternatives from each word occurrence; and two local models, which extract just a single referent from each occurrence. One of these local models, dubbed Pursuit, uses an associative learning mechanism to estimate word-referent probability but pursues and tests the best referent-meaning at any given time. Pursuit is found to perform as well as global models under many conditions extracted from naturalistic corpora of parent-child interactions, even though the model maintains far less information than global models. Moreover, Pursuit is found to best capture human experimental findings from several relevant cross-situational word-learning experiments, including those of Yu and Smith (), the paradigm example of a finding believed to support fully global cross-situational models. Implications and limitations of these results are discussed, most notably that the model characterizes only the earliest stages of word learning, when reliance on the co-occurring referent world is at its greatest.

A Strategic Reasoning Model for Generating Alternative Answers

We characterize a class of indirect answers to yes/no questions, alternative answers, where information is given that is not directly asked about, but which might nonetheless address the underlying motivation for the question. We develop a model rooted in game theory that generates these answers via strategic reasoning about possible unobserved domain-level user requirements. We implement the model within an interactive question answering system simulating real estate dialogue. The system learns a prior probability distribution over possible user requirements by analyzing training dialogues, which it uses to make strategic decisions about answer selection. The system generates pragmatically natural and interpretable answers which make for more efficient interactions compared to a baseline.

Pragmatic question answering: A game-theoretic approach

Abstract We present results of quantitative evaluations of a content selection scheme for answer generation in sales dialogue which is based on an interactive game-theoretic model of the dialogue scenario. The model involves representing a probability distribution over possible customer requirements, i.e., needs that must be met before a customer will agree to buy an object. Through game-theoretic analysis we derive a content selection procedure which constitutes an optimal strategy in the dialogue game. This procedure is capable of producing pragmatically appropriate indirect answers to yes/no questions, and is implemented in an online question answering system. Evaluation results show that these answers are pragmatically natural and contribute to dialogue efficiency. The model allows for systems that learn probabilities of customer requirements, both online and from previous data.

The emergence of focus in an experimental communication game

We report an experiment investigating the emergence of focus, the prosodic or morphosyntactic marking of critical elements (Schmitz, 2008) in a sentence. Stevens (2016) argued for a theory of focus based in information theory (Shannon & Weaver, 1949; Schmitz, 2008; Bergen & Goodman, 2015). Language users must deal with noise – the random deletion or alteration of parts of a signal. A solution is to compensate by adding redundancy (e.g., greater prosodic or morphosyntactic emphasis). However, redundancy costs both effort and time, so we should expect speakers to restrict redundancy to critical elements, particularly when effort and time pressures are high. (Redundancy on critical elements will be referred to here as critical redundancy, as compared with non-critical redundancy on other elements.) These factors should be expected to operate over multiple timescales. In a single interaction, speakers respond dynamically to perceived noise, time and effort pressures (Krauss & Weinheimer, 1964; Clark, 1996; Brennan & Clark, 1996). Developmentally, language learners acquire strategies for adding redundancy (Romaine, 1984). Over generations, such strategies can be expected to become grammaticalized as focus systems (Tamariz & Kirby, 2016). We thus expect focus-like behavior to emerge and evolve in any communication system that involves sending messages under similar constraints and make the following predictions: (1) Overall message length should vary according to time and effort costs; (2) longer messages should differ from shorter messages not only with respect to length – shorter messages should also have lower proportions of non-critical redundancy; (3) critical redundancy should be higher when noise is higher, both in an absolute sense and in a relative sense (more critical than noncritical redundancy); (4) unless noise and time pressures actually prevent accurate communication, communicative accuracy should remain relatively constant, because focus is designed to help maintain accuracy under different conditions. We tested these predictions experimentally by having participants play a simple communication game. Players sat separately and took turns to be “Sender” or “Receiver”. The Sender would see three grids, two with line figures (Figure 1; in 422

Pragmatic Query Answering: Results from a Quantitative Evaluation

This paper reports on an implementation of methods for generating indirect responses in question-answering dialogue based on domain-level strategic reasoning. User’s questions are interpreted as reflexes of underlying user requirements which are potentially satisfied by information beyond what is directly asked about. We find that the algorithms that reason about user requirements yield significantly shorter dialogues than a simpler baseline, and that users are able to interact with these systems in a pragmatically natural way.