• Digital Humanities and Cultural Heritage
• Publicationsclose
• Research dataclose
• Research softwareclose
• Other research productsclose
• Open Access close
• Article close
• 0501 psychology and cognitive sciences close
• European Commission close
• EC|H2020 close
• EU close
• Social Science and Humanities close

Textbook Question Answering is a complex task in the intersection of Machine Comprehension and Visual Question Answering that requires reasoning with multimodal information from text and diagrams. For the first time, this paper taps on the potential of transformer language models and bottom-up and top-down attention to tackle the language and visual understanding challenges this task entails. Rather than training a language-visual transformer from scratch we rely on pre-trained transformers, fine-tuning and ensembling. We add bottom-up and top-down attention to identify regions of interest corresponding to diagram constituents and their relationships, improving the selection of relevant visual information for each question and answer options. Our system ISAAQ reports unprecedented success in all TQA question types, with accuracies of 81.36%, 71.11% and 55.12% on true/false, text-only and diagram multiple choice questions. ISAAQ also demonstrates its broad applicability, obtaining state-of-the-art results in other demanding datasets. Accepted for publication as a long paper in EMNLP2020

In visual word identification, readers automatically access word internal information: they recognize orthographically embedded words (e.g., HAT in THAT) and are sensitive to morphological structure (DEAL-ER, BASKET-BALL). The exact mechanisms that govern these processes, however, are not well established yet - how is this information used? What is the role of affixes in this process? To address these questions, we tested the activation of meaning of embedded word stems in the presence or absence of a morphological structure using two semantic categorization tasks in Italian. Participants made category decisions on words (e.g., is CARROT a type of food?). Some no-answers (is CORNER a type of food?) contained category-congruent embedded word stems (i.e., CORN-). Moreover, the embedded stems could be accompanied by a pseudo-suffix (-er in CORNER) or a non-morphological ending (-ce in PEACE) - this allowed gauging the role of pseudo-suffixes in stem activation. The analyses of accuracy and response times revealed that words were harder to reject as members of a category when they contained an embedded word stem that was indeed category-congruent. Critically, this was the case regardless of the presence or absence of a pseudo-suffix. These findings provide evidence that the lexical identification system activates the meaning of embedded word stems when the task requires semantic information. This study brings together research on orthographic neighbors and morphological processing, yielding results that have important implications for models of visual word processing.

Abstract Four language production experiments examine how English speakers plan compound words during phonological encoding. The experiments tested production latencies in both delayed and online tasks for English noun-noun compounds (e.g., daytime), adjective-noun phrases (e.g., dark time), and monomorphemic words (e.g., denim). In delayed production, speech onset latencies reflect the total number of prosodic units in the target sentence. In online production, speech latencies reflect the size of the first prosodic unit. Compounds are metrically similar to adjective-noun phrases as they contain two lexical and two prosodic words. However, in Experiments 1 and 2, native English speakers treated the compounds as single prosodic units, indistinguishable from simple words, with RT data statistically different than that of the adjective-noun phrases. Experiments 3 and 4 demonstrate that compounds are also treated as single prosodic units in utterances containing clitics (e.g., dishcloths are clean) as they incorporate the verb into a single phonological word (i.e. dishcloths-are). Taken together, these results suggest that English compounds are planned as single recursive prosodic units. Our data require an adaptation of the classic model of phonological encoding to incorporate a distinction between lexical and postlexical prosodic processes, such that lexical boundaries have consequences for post-lexical phonological encoding.

Sound correspondence patterns play a crucial role for linguistic reconstruction. Linguists use them to prove language relationship, to reconstruct proto-forms, and for classical phylogenetic reconstruction based on shared innovations. Cognate words which fail to conform with expected patterns can further point to various kinds of exceptions in sound change, such as analogy or assimilation of frequent words. Here we present an automatic method for the inference of sound correspondence patterns across multiple languages based on a network approach. The core idea is to represent all columns in aligned cognate sets as nodes in a network with edges representing the degree of compatibility between the nodes. The task of inferring all compatible correspondence sets can then be handled as the well-known minimum clique cover problem in graph theory, which essentially seeks to split the graph into the smallest number of cliques in which each node is represented by exactly one clique. The resulting partitions represent all correspondence patterns which can be inferred for a given dataset. By excluding those patterns which occur in only a few cognate sets, the core of regularly recurring sound correspondences can be inferred. Based on this idea, the paper presents a method for automatic correspondence pattern recognition, which is implemented as part of a Python library which supplements the paper. To illustrate the usefulness of the method, we present how the inferred patterns can be used to predict words that have not been observed before.

Natural language is compositional; the meaning of a sentence is a function of the meaning of its parts. This property allows humans to create and interpret novel sentences, generalizing robustly outside their prior experience. Neural networks have been shown to struggle with this kind of generalization, in particular performing poorly on tasks designed to assess compositional generalization (i.e. where training and testing distributions differ in ways that would be trivial for a compositional strategy to resolve). Their poor performance on these tasks may in part be due to the nature of supervised learning which assumes training and testing data to be drawn from the same distribution. We implement a meta-learning augmented version of supervised learning whose objective directly optimizes for out-of-distribution generalization. We construct pairs of tasks for meta-learning by sub-sampling existing training data. Each pair of tasks is constructed to contain relevant examples, as determined by a similarity metric, in an effort to inhibit models from memorizing their input. Experimental results on the COGS and SCAN datasets show that our similarity-driven meta-learning can improve generalization performance. ACL2021 Camera Ready; fix a small typo

Structured information is easier to remember and recall than random one. In real life, information exhibits multi-level hierarchical organization, such as clauses, sentences, episodes and narratives in language. Here we show that multi-level grouping emerges even when participants perform memory recall experiments with random sets of words. To quantitatively probe brain mechanisms involved in memory structuring, we consider an experimental protocol where participants perform ‘final free recall’ (FFR) of several random lists of words each of which was first presented and recalled individually. We observe a hierarchy of grouping organizations of FFR, most notably many participants sequentially recalled relatively long chunks of words from each list before recalling words from another list. More-over, participants who exhibited strongest organization during FFR achieved highest levels of performance. Based on these results, we develop a hierarchical model of memory recall that is broadly compatible with our findings. Our study shows how highly controlled memory experiments with random and meaningless material, when combined with simple models, can be used to quantitatively probe the way meaningful information can efficiently be organized and processed in the brain, so to be easily retrieved.Significance StatementInformation that people communicate to each other is highly structured. For example, a story contains meaningful elements of various degrees of complexity (clauses, sentences, episodes etc). Recalling a story, we are chiefly concerned with these meaningful elements and not its exact wording. Here we show that people introduce structure even when recalling random lists of words, by grouping the words into ‘chunks’ of various sizes. Doing so improves their performance. The so formed chunks closely correspond in size to story elements described above. This suggests that our memory is trained to create a structure that resembles the one it typically deals with in real life, and that using random material like word lists can be used to quantitatively probe these memory mechanisms.

This study applies relative entropy in naturalistic large-scale corpus to calculate the difference among L2 (second language) learners at different levels. We chose lemma, token, POStrigram, conjunction to represent lexicon and grammar to detect the patterns of language proficiency development among different L2 groups using relative entropy. The results show that information distribution discrimination regarding lexical and grammatical differences continues to increase from L2 learners at a lower level to those at a higher level. This result is consistent with the assumption that in the course of second language acquisition, L2 learners develop towards a more complex and diverse use of language. Meanwhile, this study uses the statistics method of time series to process the data on L2 differences yielded by traditional frequency-based methods processing the same L2 corpus to compare with the results of relative entropy. However, the results from the traditional methods rarely show regularity. As compared to the algorithms in traditional approaches, relative entropy performs much better in detecting L2 proficiency development. In this sense, we have developed an effective and practical algorithm for stably detecting and predicting the developments in L2 learners’ language proficiency. H2020 European Research Council

There is widespread interest in the relationship between the neurobiological systems supporting human cognition and emerging computational systems capable of emulating these capacities. Human speech comprehension, poorly understood as a neurobiological process, is an important case in point. Automatic Speech Recognition (ASR) systems with near-human levels of performance are now available, which provide a computationally explicit solution for the recognition of words in continuous speech. This research aims to bridge the gap between speech recognition processes in humans and machines, using novel multivariate techniques to compare incremental ‘machine states’, generated as the ASR analysis progresses over time, to the incremen- tal ‘brain states’, measured using combined electro- and magneto-encephalography (EMEG), generated as the same inputs are heard by human listeners. This direct comparison of dynamic human and machine internal states, as they respond to the same incrementally delivered sensory input, revealed a significant correspondence between neural response patterns in human superior temporal cortex and the structural properties of ASR-derived phonetic models. Spatially coherent patches in human temporal cortex responded selectively to individual phonetic features defined on the basis of machine-extracted regularities in the speech to lexicon mapping process. These results demonstrate the feasibility of relating human and ASR solutions to the problem of speech recognition, and suggest the potential for further studies relating complex neural computations in human speech comprehension to the rapidly evolving ASR systems that address the same problem domain. This research was supported financially by an Advanced Investigator grant to WMW from the European Research Council (AdG 230570 NEUROLEX), by MRC Cognition and Brain Sciences Unit (CBSU) funding to WMW (U.1055.04.002.00001.01), and by a European Research Council Advanced Investigator grant under the European Community’s Horizon 2020 Research and Innovation Programme (2014-2020 ERC Grant agreement no 669820) to Lorraine K. Tyler. LS was partly supported by the NIHR Biomedical Research Centre and Biomedical Unit in Dementia based at Cambridge University Hospital NHS Foundation Trust.

We describe the development of a computational model predicting listener-perceived expressions of music in branding contexts. Representative ground truth from multi-national online listening experiments was combined with machine learning of music branding expert knowledge, and audio signal analysis toolbox outputs. A mixture of random forest and traditional regression models is able to predict average ratings of perceived brand image on four dimensions. Resulting cross-validated prediction accuracy (R²) was Arousal: 61%, Valence: 44%, Authenticity: 55%, and Timeliness: 74%. Audio descriptors for rhythm, instrumentation, and musical style contributed most. Adaptive sub-models for different marketing target groups further increase prediction accuracy.

International audience; Auditory display can complement visual representations in order to better interpret scientific data. A previous article showed that the free categorization of “audified seismic signals” operated by listeners can be explained by various geophysical parameters. The present article confirms this result and shows that cognitive representations of listeners can be used as heuristics for the characterization of seismic signals. Free sorting tests are conducted with audified seismic signals, with the earthquake/seismometer relative location, playback audification speed, and earthquake magnitude as controlled variables. The analysis is built on partitions (categories) and verbal comments (categorization criteria). Participants from different backgrounds (acousticians or geoscientists) are contrasted in order to investigate the role of the participants' expertise. Sounds resulting from different earthquake/station distances or azimuths, crustal structure and topography along the path of the seismic wave, earthquake magnitude, are found to (a) be sorted into different categories, (b) elicit different verbal descriptions mainly focused on the perceived number of events, frequency content, and background noise level. Building on these perceptual results, acoustic descriptors are computed and geophysical interpretations are proposed in order to match the verbal descriptions. Another result is the robustness of the categories with respect to the audification speed factor.