Language Processing Lab, Kanpur

Individual differences research program (Project Indra)

A considerable number of studies in the last few decades (Estes 1956; Underwood 1975; Maddox 1999; Fific 2014; Kidd et al. 2018; Yadav et al. 2022) have highlighted that individual differences cannot be ignored in theory development. The inferences drawn from the individual-level behavior can reveal important theoretical insights about the underlying process, which may otherwise be masked by the average behavior of the sample (Fific 2014; Tanner 2019). A comprehensive investigation of individual differences in sentence processing could be instrumental in developing a complete theory that can account for a range of empirical phenomena observed across languages. We use self-paced reading, eye-tracking, and other behavioral tests to identify the sources of individual differences in processing. For example, an individual will be measured on a battery of tasks to obtain scores for certain parameters, such as reading speed, predictive strength, and cue weighting; these individual-level parameter estimates will then be plugged into a hypothesized model to generate predictions for each individual on a sentence processing task. The predicted effect for each individual will be compared against the observed effect to evaluate an individual difference model. In principle, this approach can be used to build a theory of the sentence comprehension process exhibited by a particular individual. Individual differences have implications not only in sentence processing but also in perceptual decision-making and memory research (Houpt et al. 2016). A natural extension of this program will be modeling individual differences in cognition. Here is a list of ongoing projects as a part of this research program.

Independent measurements benchmark for individual differences research (IMB project). We are developing a benchmark dataset for studying individual differences in sentence processing. We take multiple independent measurements from each participant including working memory span, syntactic predictability, cue weighting, lexical accessibility, reading speed, locality effect, anti-locality effect, interference, and agreement attraction. The measurements on sentence processing tasks are taken from Indian languages including Hindi, Telugu, and Malayalam.
Individual differences in second language processing. We plan to do experimental work to investigate how native Telugu/Malayalam speakers differ in their processing of English sentences. We first test the prediction-strategy-transfer hypothesis: The native Telugu speakers who are better at predicting the upcoming verb should show weaker locality effects on English sentences compared to the ones who are poorer in predictive processing in their native language.
Assessment of reliability of individual differences in sentence processing. We record the same measurements over multiple sessions in the same set of individuals to assess the reliability of individual differences on certain sentence-processing tasks. The goal is to diagnose the reliability problem and propose some solutions to mitigate this serious challenge.
A theory of sentence processing at the individual level. We use the IMB dataset to develop a theory of an individual's behavior on a sentence-processing task as the parametric variation in an underlying cognitive process. Such models could be instrumental in understanding why individuals from different language populations differ on a sentence processing task, why individuals with impaired comprehension show qualitatively different processing behavior from the controls, etc.
How do prediction and memory constraints interact at the individual level? Working memory constraints are argued be to central to sentence comprehension, but there is little evidence for certain working memory effects in the verb-final languages. Instead, a predictive processing strategy is considered as the key factor in these languages. The experimental data from these verb-medial and verb-final languages indicate the role of both memory constraints and predictions. However, little is known about how these processes interact. We take an individual differences approach to investigate how working memory constraints and prediction interact to generate the observed processing behavior (locality effect) in an individual (Hindi speaker).
The role of cue weighting in sentence comprehension. Due to their varying language exposure and learning strategies at an early age, individuals differ in which linguistic cues they most rely on during real-time sentence comprehension. For instance, some individuals may weigh syntactic cues higher over the semantic and word order cues, while others weigh them equally. The cue-based retrieval theory and the cue competition theory predict that an individual's cue weighting strategy can considerably affect their behavior on a sentence processing task. This study provides a more conservative test for the hypothesis that the comprehender distributes a fixed, limited resource to different types of linguistic cues to optimize the comprehension process. We test whether differences in cue weighting can explain individual differences in semantic/syntactic interference in Hindi/Telugu comprehenders.

Newell theory development program (Project Newt)

The existing theoretical research in sentence processing faces three main challenges: (i) the lack of computationally implemented models that can generate well-constrained quantitative predictions (Nicenboim & Vasishth 2018), (ii) the lack of empirical coverage (e.g., most theories are designed to explain only one or two empirical phenomena) (Vasishth et al. 2019), (iii) the lack of quantitative evaluation of competing theories using data from multiple studies simultaneously; the focus is often on the question of whether a particular study supports the qualitative prediction of a theory or not (Vasishth & Gelman 2021). Our goal is to develop a sufficient theory of sentence comprehension behavior (Newell 1973): a theory that can explain multiple empirical phenomena (quantitatively) using a single set of well-specified assumptions and can also predict individual-level behavior. This theory development program will have two steps: (i) Implementing a set of competing theories as computational models; these models should generate falsifiable quantitative predictions for multiple phenomena of interest, and (ii) A quantitative evaluation of competing theories using benchmark data from multiple studies from typologically different languages. We use Bayesian methods for implementing computational models, for estimating model parameters, and for model comparison. However, many sentence processing models are too complex that it is difficult to express their likelihood function analytically, and therefore, one cannot use conventional Bayesian methods, unless the models are drastically simplified (Smith et al. 2022; Engbert et al. 2022). To deal with this challenge, we use a likelihood-free inference approach, the Approximate Bayesian Computation (ABC), for parameter estimation and model evaluation. This research program will emphasize on building theories that find independent support for their assumptions across different cognitive tasks including decision-making, perception, reading, recall, and recognition. Such theoretical research can yield insights into a unified theory of cognition that can account for a wide range of phenomena across different tasks (Newell 1973). The following projects are underway:

An integrated model of working memory constraints and prediction. We are building a model of how working memory constraints and predictive processes interact during real-time language comprehension. This model is critical in resolving an outstanding empirical puzzle: why do verb-final languages show the anti-locality effect and verb-medial languages show the locality effect on the same type of constructions? The modeling work will be an important step in the direction of explaining cross-linguistic differences in sentence processing and building models based on invariant aspects of the underlying comprehension process.
The distortion-preactivation-retrieval theory of sentence comprehension. We are working on a new theory of sentence processing that makes three simple assumptions: (i) When linguistic input is temporarily stored in memory, it distorts to a non-veridical representation probabilistically, (ii) the (probabilistically) distorted input in memory preactivates certain linguistic items based on their history of co-occurrences, and (iii) dependency completion is driven by a content-addressable search in memory, which operates on potentially distorted input and might be facilitated by preactivation. The assumptions are well-motivated and have found independent empirical support in sentence processing data. Our model attempts to unify these assumptions in a single set of architectural constraints (based on ACT-R).
A comprehensive evaluation of information-theoretic models of sentence processing. We compare all information-theoretic models of sentence processing proposed in the literature on a set of robustly observed sentence processing phenomena. We implement the surprisal models, noisy-channel model, lossy-context surprisal model, and the rate-distortion model to predict locality, interference, agreement attraction, and garden-path effects. The models will be evaluated using k-fold cross-validation on the data from published studies. This study will also address the problem of difficulty-to-reading-time linking function in these models.

An evolution theory of language as a complex adaptive system

Language can be viewed as a complex adaptive system that evolves as a result of adaptation to its environment, the communication channel between the encoder and the decoder. This communication channel involves cognitive constraints of the users, communicative needs, social and situational constraints, transmission medium noise, etc. A formal model of the communication channel and how it interacts with language can help in discovering the principles of language evolution. Here, we see language as a complex system characterized by a quantitative distribution of structural preferences. This distribution of structural preferences changes over time in a language and can go in different directions in different languages, but is still constrained by the properties of its environment, the communication channel. The study of why all human languages fall within a class of quantitative distribution of structural preferences and why they differ in their structural preferences can contribute to understanding how cognitive constraints and communicative needs shape a human language over time. While a generative engine allows a large set of structures to occur in human languages and a really large set does indeed occur. But it seems only a very specific set of structural preferences survive the environmental constraints. Our lab is dedicated to building a theory of how cognitive constraints and communicative needs of the users shape the quantitative distributions of structural preferences across languages. Here is a list of ongoing projects in this direction.

The role of predictability maximization pressure in shaping word-order choices across languages. We test whether the consideration to maximize the predictability of the upcoming input influences word-order choices across languages. We compare human language against artificially generated sentences under the same generative (grammar) constraints. If a human language prefers structures that maximize predictability, we should see a higher rate of predictable structures in human language sentences compared to artificially generated sentences.
Working memory constraints and predictability maximization as opposite pressures to explain the variability in structural preferences across languages. Under some theoretical assumptions, the constraints on working memory and predictability considerations predict the opposite about the placement of constituents in a phrase/sentence. We pursue a theory of language evolution and language diversity that assumes that a language fluctuates between the working memory pressure and predictability maximization pressure over time. All human languages lie on a continuum between absolute working memory minimization and absolute predictability maximization.
Dependency length minimization in humans vs machines. Do sentences produced by large language models (LLMs) minimize dependency lengths in the same way as humans do? Is the extent of dependency length minimization the same in human-produced sentences and LLM-produced sentences?

Diagnostic tools for language deficits

Sentence processing research can potentially contribute to the domain of treating Aphasia and other language impairments. We plan to start the following projects in these directions.

Experimental research for developing diagnostic tools for Aphasia in the Indian language population: We plan to develop culturally and linguistically suitable batteries for the diagnosis of language impairments in Indian languages. This would require large-scale empirical and experimental research on language production and comprehension in Individuals with Aphasia and controls.
A unified theory of sentence processing in individuals with Aphasia and in controls. We plan to study the individuals with Aphasia (IWAs) on sentence processing tasks in an Indian language. The goal is to implement and test a model that can potentially explain why IWAs differ from controls and from each other in their processing behavior.
Dependency relations knowledge developed by LLMs. Do the LLMs learn to use a hierarchical dependency structure for making predictions about the upcoming word? What kind of dependency relations are learned and how they are weighed by the LLMs?

Language behavior analytics

Language behavior involves the type of linguistic constructions we produce more often than others, the type of words we utter, the distribution of topics, discourse in our conversations, structural complexity, sentiment behind our text or speech, and fluency of our utterances, etc. Modeling such language behavior has potential applications in marketing, investment, product development, and mental health diagnosis, among others. We are working on these topics in this direction.

Consumer sentiment analysis based on text from Amazon product reviews
Investor sentiment analysis
Mental health diagnosis tools

Statistical modeling for sports

Chess analytics
Cricket analytics

Language comprehension-ease-based design for UI/UX

Vision

The ultimate goal of our research is to develop a unified theory of cognition in humans and machines: A theory that can explain multiple empirical phenomena observed in complex cognitive functions that humans and machines can perform. The criteria for the theory are as follows:

Should describe what computations, representations, and algorithms are exactly computed and performed by the brain and the machines (e.g., transformer-based models) when they generate language processing, perception, and decision-making behavior. It should provide a reasonably approximate generalization of the distributed neural activities at an algorithmic level using a set of equations.
Should be able to characterize how the machines are different from humans in the underlying, emerged cognitive architectures and processing strategies?
Should explain empirical phenomena across multiple cognitive tasks including language processing, perception, and decision-making.
Should generate predictions at the neural level as well as the behavior level.

Page updated

Google Sites

Report abuse