Research in the Human Language Processing Lab
T. Florian Jaeger, Brain and Cognitive Sciences, University of Rochester — May 9th, 2015
The stunning ability to communicate abstract messages is a quintessential human trait that uniquely defines us in the animal kingdom. At the same time, human language is a complex behavior that presumably draws in large parts on evolutionarily younger neural/cognitive systems. Research in the Human Language Processing (HLP) Lab seeks to understand the computational cognitive systems that allow the human brain to communicate information at a rate and complexity that far exceeds that of other animals.
To illustrate the complexity of speech perception, for example, imagine reading a hand-written text without spaces between words through a somewhat blurry lense showing only 1-2 letters a time moving under variably rapid speeds of 3-7 syllables/second that are not under your control … What then allows us to typically understand each other (though certainly not without the occasional problem)? Understanding the perceptual and motor inference processes that underlie our ability to robustly communicate, is the primary goal of research in HLP Lab. To this end, we study language comprehension and production, and the adaptive learning processes that facilitate message transfer despite noise in the environment and in our biological systems. We employ mathematical and computational models and evaluate them against behavioral data (and, more recently, also brain imaging data). This includes research at the phonetic, phonological, morphological, syntactic, and discourse level in written and spoken comprehension as well as spoken production. Together with research from other labs, our work suggests that the language production and comprehension systems are organized to exploit statistical structure that is present the world (e.g., the speech signal) to de- and encode language robustly. For example, both speakers and listeners take advantage of redundancy in the signal as well as expectations about likely messages based on previous experience.
The role of prediction and uncertainty in language understanding.
To understand language, comprehenders must map an acoustic (or visual) signal onto sequences of phonemes, words, and phrases, with the ultimate goal of inferring the intended message. This signal is perturbed by noise –for example, noise in the environment, but also perceptual noise. How then do we understand each other? It is now broadly accepted that top-down expectations play a crucial role in achieving robust and perhaps even efficient language understanding. These expectations are based on beliefs about the distributions of phonetic features, sound classes, words, and syntactic structures.
Research in HLP Lab seeks to understand how these beliefs and uncertainty about them affect language understanding via prediction (e.g., Bicknell et al., 2014; Fine et al., 2010, 2013; Linzen and Jaeger, 2014). For example, we recently have investigated the role of word-by-word entropy in parsing written input (Linzen and Jaeger, 2014, in prep). In other recent research, we have investigated how much uncertainty listeners can maintain about the speech signal as it unfolds over time (Bicknell et al., 2014, in prep). This research suggests that the human brain compresses the high-dimensional speech signal in a way that even many syllables downstream, subphonetic details are readily recoverable. Furthermore, this work suggests that the combination of information from the bottom-up acoustic signal and top-down constraints follows the principles of an ideal observer (Bicknell et al., in prep).
The lack of invariance problem and how we can understand each other at all.
This ability to integrate information from different sources into predictions about upcoming sounds, words, etc. is even more astonishing once we take into consideration one of the oldest puzzle of the speech sciences, the lack of invariance: not only is the speech signal perturbed by noise, different talkers also differ in how they realize the same sound. This means that one speaker’s “s” can be physically more similar to another speaker’s “sh”, than that speaker’s “s”. Technically, this means that the statistics of linguistic sound categories are non-stationary (they change, for example, based on the talker).
An increasing focus in our research is dedicated to how the human mind overcomes this problem. We have proposed that the key in understanding robust speech perception is to view it as a problem of inference under uncertainty at multiple levels (Kleinschmidt and Jaeger, 2012, 2015; see also Qian et al., 2012, under revision for an extension beyond language processing). Listeners need to be able to recognize the statistics of previously encountered (familiar) talkers, adapt to the statistics of novel talkers, and generalize based on previous experience with similar talkers. This draws a picture of language competence that differs starkly from the idea that speaking and understanding a language involves ‘one grammar’. Rather, we continuously adapt both previous beliefs about the structure of the linguistic input we have received and beliefs about the statistical structure of the current talker (Kleinschmidt and Jaeger, 2015). We have demonstrated that this view can account for a variety of otherwise puzzling properties of speech perception, including adaptation behaviors previously considered to be due to separate mechanisms (Kleinschmidt and Jaeger, 2011, 2012), explaining-away effects in perception, our ability to generalize to novel speakers of the same language background, etc.
We have also extended this view to second language acquisition and limitations in our ability to learn new languages (Pajak et al, in press; Toscano et al., under revision). Further tests of this account are currently under way, employing artificial language learning (Chu et al., in prep) as well as the statistical modeling of big data from over 50,000 second language learners of Dutch (Schepens et al., in prep).
This ability to adjust to talker-specific preferences is not limited to pronunciation (e.g., Fine et al., 2013; Jaeger and Snider, 2013; Yildirim et al., 2013, under review). In self-paced reading (Fine et al., 2010, 2013; Fine and Jaeger, 2013a, under review; Fraundorf and Jaeger, 2014; Linzen and Jaeger, in prep), eye-tracking reading studies (Farmer et al., 2014), and spoken visual world studies (Fraundorf et al., in prep), we have found that comprehenders can rapidly adjust their expectations about the relative frequency of different syntactic structures in the current context. We have found that the magnitude of expectation violation at any point in time is a predictor of the change in expectations at subsequent processing (prediction error-based learning, Fine and Jaeger, 2013b; Jaeger and Snider, 2013). Furthermore, the same type of computational model that provides excellent qualitative and quantitative fits against data from phonetic adaptation also provides a good fit against our experiments on expectation adaptation during sentence processing (Fine et al., 2010; Kleinschmidt et al., 2012). More broadly, we are also interested in understanding how prediction and uncertainty affect language processing (Bicknell et al., to appear; Kuperberg and Jaeger, under review; Linzen & Jaeger, 2015).
To what extent and how do speakers contribute to successful communication?
Another focus of research in HLP Lab lies on the speaker. Languages often provide several near-meaning equivalent ways of encoding the same message. We have investigated to what extent speakers’ preferences between different forms for the same meaning are affected by a bias to trade off production ease or effort against the goal to be understood (e.g., Jaeger, 2006, 2010; Levy and Jaeger, 2007; Jaeger and Ferreira, 2013). For example, we have investigated speakers’ preference to produce more or less reduced words or structures based on the contextual inferability of the meaning they encode (e.g., Frank and Jaeger, 2008; Jaeger, 2011; Kurumada and Jaeger, 2013, 2015; Norcliffe and Jaeger, 2015; Wasow et al., 2011). At any level of linguistic representations –ranging from phonetics, via morphology, syntactic production, to supra-clausal structure—we have found that more predictable elements tend to be reduced (see Jaeger, 2013 for an overview). Together with a growing body of research from other lab –both from corpora and experiments—this provides evidence that language production is organized in a way that on average dedicates less effort to the production of linguistic form that can be inferred from the context.
We have observed similar preferences in language learning. Specifically, we have employed artificial language learning to investigate whether learners have abstract biases for languages that balance effort and robust message transfer (Fedzechkina et al., 2011, 2012, 2013, under revision). For example, we have exposed monolingual speakers of English (a language without case) to a miniature (artificial) language with optional case-marking (a phenomenon that exists in Japanese and Korean, for example). Learners restructure the language to resemble the case-marking patterns observed in actual languages –specifically, the increased the rate of case-marking in sentences that would otherwise have been ambiguous but omitted case-marking where it wasn’t necessary.
More recently we have begun to study exactly how speaker come to trade off production effort/ease against robust message transfer. Specifically, we have proposed that speakers integrate feedback both from the perception of their own speech and from interlocutors with the goal to increase the communicative success of subsequent utterances (in ways similar to implicit learning in motor control, cf. Jaeger, 2013; Jaeger and Ferreira, 2013). To study these questions, we study how speakers change their pronunciation (Buz et al., 2014, under review; Seyfarth et al., 2015) and syntactic production preferences (Roche et al., in prep). In related lines of work, we have investigated how speakers adapt to what they perceive to be their interlocutors’ expectations (Jaeger and Snider, 2013) or in response to social relations or status of interlocutors (Weatherholtz et al., 2014).
This includes introductions and reviews of advanced statistical methods (e.g., Generalized Linear Mixed Models, Jaeger, 2008) and their application to typology, which typically provides hierarchically organized data (Jaeger et al., 2011, 2012) or corpus-based research, which typically involves heterogeneous and high-dimensional data (Frank and Jaeger, 2008; Jaeger, 2006, 2010, 2011). Researchers in HLP Lab have also been strongly involved in the development, testing, and application of new crowdsourcing paradigms for language research (e.g., web-based paradigms for acceptability judgments, Jaeger, 2004; self-paced reading, Fine et al., 2010; artificial language learning, Tily et al., 2011; iterative language learning, Gutman, 2012; speech perception and adaptation, Kleinschmidt and Jaeger, 2012; spoken sentence recall, Jaeger and Grimshaw, 2013; the role of feedback in articulation, Buz et al., 2014; socially-mediated alignment in spoken scene descriptions, Weatherholtz et al., 2014). These paradigms allow the collection of data from speakers and listeners of different socio-economic and language backgrounds. Going beyond the study of college populations also motivates our study of language production via speech corpora (Degen and Jaeger, in prep; Gallo et al., 2008, 2009; Jaeger, 2010, 2011; Jaeger and Wasow, 2006; Jaeger and Snider, 2013; see also Jaeger et al., 2012a,b on phonological production in unscripted speech) and a dedication to cross-linguistic validation and extension of psycholinguistic theories (e.g., Jaeger and Norcliffe, 2009; Qian, 2009; Qian and Jaeger, 2012), including research on Yucatec Maya (Butler et al., 2011, 2013, 2014, in prep; Norcliffe and Jaeger, 2015), Mexican Spanish (e.g., Gallo et al., 2009; Butler et al., in prep), and Japanese (Kurumada and Jaeger, 2013, 2015).