Publications

Abstract

Lifelong graph learning deals with the problem of continually adapting graph neural network (GNN) models to changes in evolving graphs. We address two critical challenges of lifelong graph learning in this work: dealing with new classes and tackling imbalanced class distributions. The combination of these two challenges is particularly relevant since newly emerging classes typically resemble only a tiny fraction of the data, adding to the already skewed class distribution. We make several contributions: First, we show that the amount of unlabeled data does not influence the results, which is an essential prerequisite for lifelong learning on a sequence of tasks. Second, we experiment with different label rates and show that our methods can perform well with only a tiny fraction of annotated nodes. Third, we propose the gDOC method to detect new classes under the constraint of having an imbalanced class distribution. The critical ingredient is a weighted binary cross-entropy loss function to account for the class imbalance. Moreover, we demonstrate combinations of gDOC with various base GNN models such as GraphSAGE, Simplified Graph Convolution, and Graph Attention Networks. Lastly, our k-neighborhood time difference measure provably normalizes the temporal changes across different graph datasets. With extensive experimentation, we find that the proposed gDOC method is consistently better than a naive adaption of DOC to graphs. Specifically, in experiments using the smallest history size, the out-of-distribution detection score of gDOC is 0.09 compared to 0.01 for DOC. Furthermore, gDOC achieves an Open-F1 score, a combined measure of in-distribution classification and out-of-distribution detection, of 0.33 compared to 0.25 of DOC (32% increase).

Abstract

Scientific convergence is a phenomenon where the distance between hitherto distinct scientific fields narrows and the fields gradually overlap over time. It is creating important potential for research, development, and innovation. Although scientific convergence is crucial for the development of radically new technology, the identification of emerging scientific convergence is particularly difficult since the underlying knowledge flows are rather fuzzy and unstable in the early convergence stage. Nevertheless, novel scientific publications emerging at the intersection of different knowledge fields may reflect convergence processes. Thus, in this article, we exploit the growing number of research and digital libraries providing bibliographic metadata to propose an automated analysis of science dynamics. We utilize and adapt machine-learning methods (DeepWalk) to automatically learn a similarity measure between scientific fields from graphs constructed on bibliographic metadata. With a time-based perspective, we apply our approach to analyze the trajectories of evolving similarities between scientific fields. We validate the learned similarity measure by evaluating it within the well-explored case of cholesterol-lowering ingredients in which scientific convergence between the distinct scientific fields of nutrition and pharmaceuticals has partially taken place. Our results confirm that the similarity trajectories learned by our approach resemble the expected behavior, indicating that our approach may allow researchers and practitioners to detect and predict scientific convergence early.

Abstract

Human communication involves the process of translating intentions into communicative actions. But how exactly do our intentions surface in the visible communicative behavior we display? Here we focus on pointing gestures, a fundamental building block of everyday communication, and investigate whether and how different types of underlying intent modulate the kinematics of the pointing hand and the brain activity preceding the gestural movement. In a dynamic virtual reality environment, participants pointed at a referent to either share attention with their addressee, inform their addressee, or get their addressee to perform an action. Behaviorally, it was observed that these different underlying intentions modulated how long participants kept their arm and finger still, both prior to starting the movement and when keeping their pointing hand in apex position. In early planning stages, a neurophysiological distinction was observed between a gesture that is used to share attitudes and knowledge with another person versus a gesture that mainly uses that person as a means to perform an action. Together, these findings suggest that our intentions influence our actions from the earliest neurophysiological planning stages to the kinematic endpoint of the movement itself.

Abstract

The structural design features of human language emerge in the process of cultural evolution, shaping languages over the course of communication, learning, and transmission. What role does this leave biological evolution? This chapter highlights the biological bases and preconditions that underlie the particular type of prosocial behaviours and cognitive inference abilities that are required for languages to emerge via cultural evolution to begin with.

Abstract

Humans are unique in their sophisticated culture and societal structures, their complex languages, and their extensive tool use. According to the human self-domestication hypothesis, this unique set of traits may be the result of an evolutionary process of self-induced domestication, in which humans evolved to be less aggressive and more cooperative. However, the only other species that has been argued to be self-domesticated besides humans so far is bonobos, resulting in a narrow scope for investigating this theory limited to the primate order. Here, we propose an animal model for studying self-domestication: the elephant. First, we support our hypothesis with an extensive cross-species comparison, which suggests that elephants indeed exhibit many of the features associated with self-domestication (e.g., reduced aggression, increased prosociality, extended juvenile period, increased playfulness, socially regulated cortisol levels, and complex vocal behavior). Next, we present genetic evidence to reinforce our proposal, showing that genes positively selected in elephants are enriched in pathways associated with domestication traits and include several candidate genes previously associated with domestication. We also discuss several explanations for what may have triggered a self-domestication process in the elephant lineage. Our findings support the idea that elephants, like humans and bonobos, may be self-domesticated. Since the most recent common ancestor of humans and elephants is likely the most recent common ancestor of all placental mammals, our findings have important implications for convergent evolution beyond the primate taxa, and constitute an important advance toward understanding how and why self-domestication shaped humans’ unique cultural niche.

Abstract

The current study presents an agent-based model which simulates the innovation and
competition among lexical items in cases of language contact. It is inspired by relatively
recent historical cases in which the linguistic ecology and sociohistorical context are highly complex. Pidgin and creole genesis offers an opportunity to obtain linguistic facts, social dynamics, and historical demography in a highly segregated society. This provides a solid ground for researching the interaction of populations with different pre-existing language systems, and how different factors contribute to the genesis of the lexicon of a newly generated mixed language. We take into consideration the population dynamics and structures, as well as a distribution of word frequencies related to language use, in order to study how social factors may affect the developmental trajectory of languages. Focusing on the case of Sranan in Suriname, our study shows that it is possible to account for the
composition of its core lexicon in relation to different social groups, contact patterns, and
large population movements.

Abstract

Large pretrained language models (PreLMs) are rev-olutionizing natural language processing across all benchmarks. However, their sheer size is prohibitive for small laboratories or for deployment on mobile devices. Approaches like pruning and distillation reduce the model size but typically retain the same model architecture. In contrast, we explore distilling PreLMs into a different, more efficient architecture, Continual Multiplication of Words (CMOW), which embeds each word as a matrix and uses matrix multiplication to encode sequences. We extend the CMOW architecture and its CMOW/CBOW-Hybrid variant with a bidirectional component for more expressive power, per-token representations for a general (task-agnostic) distillation during pretraining, and a two-sequence encoding scheme that facilitates downstream tasks on sentence pairs, such as sentence similarity and natural language inference. Our matrix-based bidirectional CMOW/CBOW-Hybrid model is competitive to DistilBERT on question similarity and recognizing textual entailment, but uses only half of the number of parameters and is three times faster in terms of inference speed. We match or exceed the scores of ELMo for all tasks of the GLUE benchmark except for the sentiment analysis task SST-2 and the linguistic acceptability task CoLA. However, compared to previous cross-architecture distillation approaches, we demonstrate a doubling of the scores on detecting linguistic acceptability. This shows that matrix-based embeddings can be used to distill large PreLM into competitive models and motivates further research in this direction.

Abstract

Learning is using past experiences to inform new behaviors and actions. Because all experiences are unique, learning always requires some generalization. An effective way of improving generalization is to expose learners to more variable (and thus often more representative) input. More variability tends to make initial learning more challenging, but eventually leads to more general and robust performance. This core principle has been repeatedly rediscovered and renamed in different domains (e.g., contextual diversity, desirable difficulties, variability of practice). Reviewing this basic result as it has been formulated in different domains allows us to identify key patterns, distinguish between different kinds of variability, discuss the roles of varying task-relevant versus irrelevant dimensions, and examine the effects of introducing variability at different points in training.

Abstract

On the surface, the fields of animal communication and human linguistics have arrived at conflicting theories and conclusions with respect to the effect of social complexity on communicative complexity. For example, an increase in group size is argued to have opposite consequences on human versus animal communication systems: although an increase in human community size leads to some types of language simplification, an increase in animal group size leads to an increase in signal complexity. But do human and animal communication systems really show such a fundamental discrepancy? Our key message is that the tension between these two adjacent fields is the result of (a) a focus on different levels of analysis (namely, signal variation or grammar-like rules) and (b) an inconsistent use of terminology (namely, the terms “simple” and “complex”). By disentangling and clarifying these terms with respect to different measures of communicative complexity, we show that although animal and human communication systems indeed show some contradictory effects with respect to signal variability, they actually display essentially the same patterns with respect to grammar-like structure. This is despite the fact that the definitions of complexity and simplicity are actually aligned for signal variability, but diverge for grammatical structure. We conclude by advocating for the use of more objective and descriptive terms instead of terms such as “complexity,” which can be applied uniformly for human and animal communication systems—leading to comparable descriptions of findings across species and promoting a more productive dialogue between fields.

Abstract

We address the problem of recommending relevant items to a user in order to "complete" a partial set of items already known. We consider the two scenarios of citation and subject label recommendation, which resemble different semantics of item co-occurrence: relatedness for co-citations and diversity for subject labels. We assess the influence of the completeness of an already known partial item set on the recommender performance. We also investigate data sparsity through a pruning parameter and the influence of using additional metadata. As recommender models, we focus on different autoencoders, which are particularly suited for reconstructing missing items in a set. We extend autoencoders to exploit a multi-modal input of text and structured data. Our experiments on six real-world datasets show that supplying the partial item set as input is helpful when item co-occurrence resembles relatedness, while metadata are effective when co-occurrence implies diversity. This outcome means that the semantics of item co-occurrence is an important factor. The simple item co-occurrence model is a strong baseline for citation recommendation. However, autoencoders have the advantage to enable exploiting additional metadata besides the partial item set as input and achieve comparable performance. For the subject label recommendation task, the title is the most important attribute. Adding more input modalities sometimes even harms the result. In conclusion, it is crucial to consider the semantics of the item co-occurrence for the choice of an appropriate recommendation model and carefully decide which metadata to exploit.

Abstract

Cross-linguistic differences in morphological complexity could have important consequences for language learning. Specifically, it is often assumed that languages with more regular, compositional, and transparent grammars are easier to learn by both children and adults. Moreover, it has been shown that such grammars are more likely to evolve in bigger communities. Together, this suggests that some languages are acquired faster than others, and that this advantage can be traced back to community size and to the degree of systematicity in the language. However, the causal relationship between systematic linguistic structure and language learnability has not been formally tested, despite its potential importance for theories on language evolution, second language learning, and the origin of linguistic diversity. In this pre-registered study, we experimentally tested the effects of community size and systematic structure on adult language learning. We compared the acquisition of different yet comparable artificial languages that were created by big or small groups in a previous communication experiment, which varied in their degree of systematic linguistic structure. We asked (a) whether more structured languages were easier to learn; and (b) whether languages created by the bigger groups were easier to learn. We found that highly systematic languages were learned faster and more accurately by adults, but that the relationship between language learnability and linguistic structure was typically non-linear: high systematicity was advantageous for learning, but learners did not benefit from partly or semi-structured languages. Community size did not affect learnability: languages that evolved in big and small groups were equally learnable, and there was no additional advantage for languages created by bigger groups beyond their degree of systematic structure. Furthermore, our results suggested that predictability is an important advantage of systematic structure: participants who learned more structured languages were better at generalizing these languages to new, unfamiliar meanings, and different participants who learned the same more structured languages were more likely to produce similar labels. That is, systematic structure may allow speakers to converge effortlessly, such that strangers can immediately understand each other.

Abstract

Social network structure has been argued to shape the structure of languages, as well as affect the spread of innovations and the formation of conventions in the community. Specifically, theoretical and computational models of language change predict that sparsely connected communities develop more systematic languages, while tightly knit communities can maintain high levels of linguistic complexity and variability. However, the role of social network structure in the cultural evolution of languages has never been tested experimentally. Here, we present results from a behavioral group communication study, in which we examined the formation of new languages created in the lab by micro‐societies that varied in their network structure. We contrasted three types of social networks: fully connected, small‐world, and scale‐free. We examined the artificial languages created by these different networks with respect to their linguistic structure, communicative success, stability, and convergence. Results did not reveal any effect of network structure for any measure, with all languages becoming similarly more systematic, more accurate, more stable, and more shared over time. At the same time, small‐world networks showed the greatest variation in their convergence, stabilization, and emerging structure patterns, indicating that network structure can influence the community's susceptibility to random linguistic changes (i.e., drift).

Abstract

Understanding worldwide patterns of language diversity has long been a goal for evolutionary scientists, linguists and philosophers. Research over the past decade has suggested that linguistic diversity may result from differences in the social environments in which languages evolve. Specifically, recent work found that languages spoken in larger communities typically have more systematic grammatical structures. However, in the real world, community size is confounded with other social factors such as network structure and the number of second languages learners in the community, and it is often assumed that linguistic simplification is driven by these factors instead. Here, we show that in contrast to previous assumptions, community size has a unique and important influence on linguistic structure. We experimentally examine the live formation of new languages created in the laboratory by small and larger groups, and find that larger groups of interacting participants develop more systematic languages over time, and do so faster and more consistently than small groups. Small groups also vary more in their linguistic behaviours, suggesting that small communities are more vulnerable to drift. These results show that community size predicts patterns of language diversity, and suggest that an increase in community size might have contributed to language evolution.

Abstract

Experimental work in the field of language evolution has shown that novel signal systems become more structured over time. In a recent paper, Kirby, Tamariz, Cornish, and Smith (2015) argued that compositional languages can emerge only when languages are transmitted across multiple generations. In the current paper, we show that compositional languages can emerge in a closed community within a single generation. We conducted a communication experiment in which we tested the emergence of linguistic structure in different micro-societies of four participants, who interacted in alternating dyads using an artificial language to refer to novel meanings. Importantly, the communication included two real-world aspects of language acquisition and use, which introduce compressibility pressures: (a) multiple interaction partners and (b) an expanding meaning space. Our results show that languages become significantly more structured over time, with participants converging on shared, stable, and compositional lexicons. These findings indicate that new learners are not necessary for the formation of linguistic structure within a community, and have implications for related fields such as developing sign languages and creoles.

Abstract

Literacy affects many aspects of cognitive and linguistic processing. Among them, it increases the salience of words as units of linguistic processing. Here, we explored the impact of literacy acquisition on children’s learning of an artifical language. Recent accounts of L1–L2 differences relate adults’ greater difficulty with language learning to their smaller reliance on multiword units. In particular, multiword units are claimed to be beneficial for learning opaque grammatical relations like grammatical gender. Since literacy impacts the reliance on words as units of processing, we ask if and how acquiring literacy may change children’s language-learning results. We looked at children’s success in learning novel noun labels relative to their success in learning article-noun gender agreement, before and after learning to read. We found that preliterate first graders were better at learning agreement (larger units) than at learning nouns (smaller units), and that the difference between the two trial types significantly decreased after these children acquired literacy. In contrast, literate third graders were as good in both trial types. These findings suggest that literacy affects not only language processing, but also leads to important differences in language learning. They support the idea that some of children’s advantage in language learning comes from their previous knowledge and experience with language—and specifically, their lack of experience with written texts.

Abstract

Recent work suggests that cultural transmission can lead to the emergence of linguistic structure as speakers’ weak individual biases become amplified through iterated learning. However, to date no published study has demonstrated a similar emergence of linguistic structure in children. The lack of evidence from child learners constitutes a problematic
2
gap in the literature: if such learning biases impact the emergence of linguistic structure, they should also be found in children, who are the primary learners in real-life language transmission. However, children may differ from adults in their biases given age-related differences in general cognitive skills. Moreover, adults’ performance on iterated learning tasks may reflect existing (and explicit) linguistic biases, partially undermining the generality of the results. Examining children’s performance can also help evaluate contrasting predictions about their role in emerging languages: do children play a larger or smaller role than adults in the creation of structure? Here, we report a series of four iterated artificial language learning studies (based on Kirby, Cornish & Smith, 2008) with both children and adults, using a novel child-friendly paradigm. Our results show that linguistic structure does not emerge more readily in children compared to adults, and that adults are overall better in both language learning and in creating linguistic structure. When languages could become underspecified (by allowing homonyms), children and adults were similar in developing consistent mappings between meanings and signals in the form of structured ambiguities. However, when homonimity was not allowed, only adults created compositional structure. This study is a first step in using iterated language learning paradigms to explore child-adult differences. It provides the first demonstration that cultural transmission has a different effect on the languages produced by children and adults: While children were able to develop systematicity, their languages did not show compositionality. We focus on the relation between learning and structure creation as a possible explanation for our findings and discuss implications for children’s role in the emergence of linguistic structure.

Abstract

Infants, children and adults are capable of extracting recurring patterns from their environment through statistical learning (SL), an implicit learning mechanism that is considered to have an important role in language acquisition. Research over the past 20 years has shown that SL is present from very early infancy and found in a variety of tasks and across modalities (e.g., auditory, visual), raising questions on the domain generality of SL. However, while SL is well established for infants and adults, only little is known about its developmental trajectory during childhood, leaving two important questions unanswered: (1) Is SL an early-maturing capacity that is fully developed in infancy, or does it improve with age like other cognitive capacities (e.g., memory)? and (2) Will SL have similar developmental trajectories across modalities? Only few studies have looked at SL across development, with conflicting results: some find age-related improvements while others do not. Importantly, no study to date has examined auditory SL across childhood, nor compared it to visual SL to see if there are modality-based differences in the developmental trajectory of SL abilities. We addressed these issues by conducting a large-scale study of children's performance on matching auditory and visual SL tasks across a wide age range (5–12y). Results show modality-based differences in the development of SL abilities: while children's learning in the visual domain improved with age, learning in the auditory domain did not change in the tested age range. We examine these findings in light of previous studies and discuss their implications for modality-based differences in SL and for the role of auditory SL in language acquisition. A video abstract of this article can be viewed at: https://www.youtube.com/watch?v=3kg35hoF0pw.

Abstract

Infants, children and adults are capable of implicitly extracting regularities from their environment through statistical learning (SL). SL is present from early infancy and found across tasks and modalities, raising questions about the domain generality of SL. However, little is known about its’ developmental trajectory: Is SL fully developed capacity in infancy, or does it improve with age, like other cognitive skills? While SL is well established in infants and adults, only few studies have looked at SL across development with conflicting results: some find age-related improvements while others do not. Importantly, despite its postulated role in language learning, no study has examined the developmental trajectory of auditory SL throughout childhood. Here, we conduct a large-scale study of children's auditory SL across a wide age-range (5-12y, N=115). Results show that auditory SL does not change much across development. We discuss implications for modality-based differences in SL and for its role in language acquisition.

Abstract

Recent work suggests that cultural transmission can lead to the emergence of linguistic structure as speakers’ weak individual biases become amplified through iterated learning. However, to date, no published study has demonstrated a similar emergence of linguistic structure in children. This gap is problematic given that languages are mainly learned by children and that adults may bring existing linguistic biases to the task. Here, we conduct a large-scale study of iterated language learning in both children and adults, using a novel, child-friendly paradigm. The results show that while children make more mistakes overall, their languages become more learnable and show learnability biases similar to those of adults. Child languages did not show a significant increase in linguistic structure over time, but consistent mappings between meanings and signals did emerge on many occasions, as found with adults. This provides the first demonstration that cultural transmission affects the languages children and adults produce similarly.