Publications

Abstract

In this work we perform multimodal detection and classification
of head movements from face to face video conversation data.
We have experimented with different models and feature sets
and provided some insight on the effect of independent features,
but also how their interaction can enhance a head movement
classifier. Used features include nose, neck and mid hip position
coordinates and their derivatives together with acoustic features,
namely, intensity and pitch of the speaker on focus. Results
show that when input features are sufficiently processed by in-
teracting with each other, a linear classifier can reach a similar
performance to a more complex non-linear neural model with
several hidden layers. Our best models achieve state-of-the-art
performance in the detection task, measured by macro-averaged
F1 score.

Abstract

Continuous vector word representations (or
word embeddings) have shown success in cap-turing semantic relations between words, as evidenced by evaluation against behavioral data of adult performance on semantic tasks (Pereira et al., 2016). Adult semantic knowl-edge is the endpoint of a language acquisition process; thus, a relevant question is whether these models can also capture emerging word
representations of young language learners. However, the data for children’s semantic knowledge across development is scarce. In this paper, we propose to bridge this gap by using Age of Acquisition norms to evaluate word embeddings learnt from child-directed input. We present two methods that evaluate word embeddings in terms of (a) the semantic neighbourhood density of learnt words, and (b) con-
vergence to adult word associations. We apply our methods to bag-of-words models, and find that (1) children acquire words with fewer semantic neighbours earlier, and (2) young learners only attend to very local context. These findings provide converging evidence for validity of our methods in understanding the prerequisite features for a distributional model of word learning.

Abstract

Most work on part-of-speech (POS) tagging is focused on high resource languages, or examines low-resource and active learning settings through simulated studies. We evaluate POS tagging techniques on an actual endangered language, Griko. We present a resource that contains 114 narratives in Griko, along with sentence-level translations in Italian, and provides gold annotations for the test set. Based on a previously collected small corpus, we investigate several traditional methods, as well as methods that take advantage of monolingual data or project cross-lingual POS tags. We show that the combination of a semi-supervised method with cross-lingual transfer is more appropriate for this extremely challenging setting, with the best tagger achieving an accuracy of 72.9%. With an applied active learning scheme, which we use to collect sentence-level annotations over the test set, we achieve improvements of more than 21 percentage points

Abstract

The term “uptalk” describes utterance-final pitch rises that carry no sentence-structural information. Uptalk is usually dialectal or sociolectal, and Australian English (AusEng) is particularly known for this attribute. We ask here whether experience with an uptalk variety affects listeners’ ability to categorise rising pitch contours on the basis of the timing and height of their onset and offset. Listeners were two groups of English-speakers (AusEng, and American English), and three groups of listeners with L2 English: one group with Mandarin as L1 and experience of listening to AusEng, one with German as L1 and experience of listening to AusEng, and one with German as L1 but no AusEng experience. They heard nouns (e.g. flower, piano) in the framework “Got a NOUN”, each ending with a pitch rise artificially manipulated on three contrasts: low vs. high rise onset, low vs. high rise offset and early vs. late rise onset. Their task was to categorise the tokens as “question” or “statement”, and we analysed the effect of the pitch contrasts on their judgements. Only the native AusEng listeners were able to use the pitch contrasts systematically in making these categorisations.

Abstract

Learning a language involves learning how to map specific forms onto their associated meanings. Such mappings can utilise arbitrariness and non-arbitrariness, yet, our understanding of how these two systems operate at different stages of vocabulary development is still not fully understood. The Sound-Symbolism Bootstrapping Hypothesis (SSBH) proposes that sound-symbolism is essential for word learning to commence, but empirical evidence of exactly how sound-symbolism influences language learning is still sparse. It may be the case that sound-symbolism supports acquisition of categories of meaning, or that it enables acquisition of individualized word meanings. In two Experiments where participants learned form-meaning mappings from either sound-symbolic or arbitrary languages, we demonstrate the changing roles of sound-symbolism and arbitrariness for different vocabulary sizes, showing that sound-symbolism provides an advantage for learning of broad categories, which may then transfer to support learning individual words, whereas an arbitrary language impedes acquisition of categories of sound to meaning.

Abstract

In order to understand early word learning we need to better understand and quantify properties of the input that young children receive. We extended the human simulation paradigm (HSP) using egocentric videos taken from infant head-mounted cameras. The videos were further annotated with gaze information indicating in-the-moment visual attention from the infant. Our new HSP prompted participants for two types of responses, thus differentiating referential from semantic ambiguity in the learning input. Consistent with findings on visual attention in word learning, we find a strongly bimodal distribution over HSP accuracy. Even in this open-ended task, most videos only lead to a small handful of common responses. What's more, referential ambiguity was the key bottleneck to performance: participants can nearly always recover the exact word that was said if they identify the correct referent. Finally, analysis shows that adult learners relied on particular, multimodal behavioral cues to infer those target referents.

Abstract

In this study, we analyze one French signing and one French speaking family’s interaction during dinner. The families composed of two parents and two children aged 3 to 11 were filmed with three cameras to capture all family members’ behaviors. The three videos per dinner were synchronized and coded on ELAN. We annotated all participants’ acting, and languaging.
Our quantitative analyses show how family members collaboratively manage multiple streams of activity through the embodied performances of dining and interacting. We uncover different profiles according to participants’ modality of expression and status (focusing on the mother and the younger child). The hearing participants’ co-activity management illustrates their monitoring of dining and conversing and how they progressively master the affordances of the visual and vocal channels to maintain the simultaneity of the two activities. The deaf mother skillfully manages to alternate smoothly between dining and interacting. The deaf younger child manifests how she is in the process of developing her skills to manage multi-activity. Our qualitative analyses focus on the ecology of visual-gestural and audio-vocal languaging in the context of co-activity according to language and participant. We open new perspectives on the management of gaze and body parts in multimodal languaging.

Abstract

Speaker diarization (answering 'who spoke when') is a widely researched subject within speech technology. Numerous experiments have been run on datasets built from broadcast news, meeting data, and call centers—the task sometimes appears close to being solved. Much less work has begun to tackle the hardest diarization task of all: spontaneous conversations in real-world settings. Such diarization would be particularly useful for studies of language acquisition, where researchers investigate the speech children produce and hear in their daily lives. In this paper, we study audio gathered with a recorder worn by small children as they went about their normal days. As a result, each child was exposed to different acoustic environments with a multitude of background noises and a varying number of adults and peers. The inconsistency of speech and noise within and across samples poses a challenging task for speaker diarization systems, which we tackled via retraining and data augmentation techniques. We further studied sources of structured variation across raw audio files, including the impact of speaker type distribution, proportion of speech from children, and child age on diarization performance. We discuss the extent to which these findings might generalize to other samples of speech in the wild.

Abstract

In two experiments, Mandarin listeners resolved potential syntactic ambiguities in spoken utterances in (a) their native language (L1) and (b) English which they had learned as a second language (L2). A new disambiguation task was used, requiring speeded responses to select the correct meaning for structurally ambiguous sentences. Importantly, the ambiguities used in the study are identical in Mandarin and in English, and production data show that prosodic disambiguation of this type of ambiguity is also realised very similarly in the two languages. The perceptual results here showed however that listeners’ response patterns differed for L1 and L2, although there was a significant increase in similarity between the two response patterns with increasing exposure to the L2. Thus identical ambiguity and comparable disambiguation patterns in L1 and L2 do not lead to immediate application of the appropriate L1 listening strategy to L2; instead, it appears that such a strategy may have to be learned anew for the L2.

Abstract

Using a phoneme detection task, the present series of
experiments examines whether listeners can entrain to
different combinations of prosodic cues to predict where focus
will fall in an utterance. The stimuli were recorded by four
female native speakers of Australian English who happened to
have used different prosodic cues to produce sentences with
prosodic focus: a combination of duration cues, mean and
maximum F0, F0 range, and longer pre-target interval before
the focused word onset, only mean F0 cues, only pre-target
interval, and only duration cues. Results revealed that listeners
can entrain in almost every condition except for where
duration was the only reliable cue. Our findings suggest that
listeners are flexible in the cues they use for focus processing.

Abstract

Listeners adapt rapidly to previously unheard talkers by
adjusting phoneme categories using lexical knowledge, in a
process termed lexically-guided perceptual learning. Although
this is firmly established for listening in the native language
(L1), perceptual flexibility in second languages (L2) is as yet
less well understood. We report two experiments examining L1
and L2 perceptual learning, the first in Mandarin-English late
bilinguals, the second in Australian learners of Mandarin. Both
studies showed stronger learning in L1; in L2, however,
learning appeared for the English-L1 group but not for the
Mandarin-L1 group. Phonological mapping differences from
the L1 to the L2 are suggested as the reason for this result.

Abstract

English and French listeners performed two tasks - click location and speeded click detection - with both English and French sentences, closely matched for syntactic and phonological structure. Clicks were located more accurately in open- than in closed-class words in both English and French; they were detected more rapidly in open- than in closed-class words in English, but not in French. The two listener groups produced the same pattern of responses, suggesting that higher-level linguistic processing was not involved in these tasks.

Abstract

Dutch listeners were slower to make judgements about the semantic relatedness between a spoken target word (e.g. atLEET, 'athlete') and a previously presented visual prime word (e.g. SPORT 'sport') when the spoken word was mis-stressed. The adverse effect of mis-stressing confirms the role of stress information in lexical recognition in Dutch. However, although the erroneous stress pattern was always initially compatible with a competing word (e.g. ATlas, 'atlas'), mis-stressed words did not produced high false alarm rates in unrelated pairs (e.g. SPORT - atLAS). This suggests that stress information did not completely rule out segmentally matching but suprasegmentally mismatching words, a finding consistent with spoken-word recognition models involving multiple activation and inter-word competition.

Abstract

The performance of automatic speech recognition systems is usually assessed in terms of error rate. Human speech recognition produces few errors, but relative difficulty of processing can be assessed via response time techniques. We report the construction of a measure analogous to response time in a machine recognition system. This measure may be compared directly with human response times. We conducted a trial comparison of this type at the phoneme level, including both tense and lax vowels and a variety of consonant classes. The results suggested similarities between human and machine processing in the case of consonants, but differences in the case of vowels.

Abstract

Simulations explored the inability of the TRACE model of spoken-word recognition to model the effects on human listening of acoustic-phonetic mismatches in word forms. The source of TRACE's failure lay not in its interactive connectivity, not in the presence of interword competition, and not in the use of phonemic representations, but in the need for continuously optimised interpretation of the input. When an analogue of TRACE was allowed to cycle to asymptote on every slice of input, an acceptable simulation of the subcategorical mismatch data was achieved. Even then, however, the simulation was not as close as that produced by the Merge model.

Abstract

Humans readily generalize, applying prior knowledge to novel situations and stimuli. Advances in machine learning have begun to approximate and even surpass human performance, but these systems struggle to generalize what they have learned to untrained situations. We present a model based on wellestablished neurocomputational principles that demonstrates human-level generalisation. This model is trained to play one video game (Breakout) and performs one-shot generalisation to a new game (Pong) with different characteristics. The model
generalizes because it learns structured representations that are functionally symbolic (viz., a role-filler binding calculus) from unstructured training data. It does so without feedback, and without requiring that structured representations are specified a priori. Specifically, the model uses neural co-activation to discover which characteristics of the input are invariant and to learn relational predicates, and oscillatory regularities in network firing to bind predicates to arguments. To our knowledge,
this is the first demonstration of human-like generalisation in a machine system that does not assume structured representa-
tions to begin with.

Abstract

Biologically inspired spiking networks are an important tool to study the nature of computation and cognition in neural systems. In this work, we investigate the representational capacity of spiking networks engaged in an identity mapping task. We compare two schemes for encoding symbolic input, one in which input is injected as a direct current and one where input is delivered as a spatio-temporal spike pattern. We test the ability of networks to discriminate their input as a function of the number of distinct input symbols. We also compare performance using either membrane potentials or filtered spike trains as state variable. Furthermore, we investigate how the circuit behavior depends on the balance between excitation and inhibition, and the degree of synchrony and regularity in its internal dynamics. Finally, we compare different linear methods of decoding population activity onto desired target labels. Overall, our results suggest that even this simple mapping task is strongly influenced by design choices on input encoding, state-variables, circuit characteristics and decoding methods, and these factors can interact in complex ways. This work highlights the importance of constraining computational network models of behavior by available neurobiological evidence.

Abstract

The study presented here focuses on two pragmatic gestures:
the hand flip (Ferré, 2011), a gesture of the Palm Up Open
Hand/PUOH family (Müller, 2004) and the closed hand which
can be considered as the opposite kind of movement to the open-
ing of the hands present in the PUOH gesture. Whereas one of
the functions of the hand flip has been described as presenting
a new point in speech (Cienki, 2021), the closed hand gesture
has not yet been described in the literature to the best of our
knowledge. It can however be conceived of as having the oppo-
site function of announcing the end of a point in discourse. The
object of the present study is therefore to determine, with the
study of prosodic features, if the two gestures are found in the
same type of speech units and what their respective scope is.
Drawing from a corpus of three TED Talks in French the
prosodic characteristics of the speech that accompanies the two
gestures will be examined. The hypothesis developed in the
present paper is that their scope should be reflected in the
prosody of accompanying speech, especially pitch key, tone,
and relative pitch range. The prediction is that hand flips and
closing hand gestures are expected to be located at the periph-
ery of Intonation Phrases (IPs), Inter-Pausal Units (IPUs) or
more conversational Turn Constructional Units (TCUs), and are
likely to be co-occurrent with pauses in speech. But because of
the natural slope of intonation in speech, the speech that accom-
pany early gestures in Intonation Phrases should reveal different
features from the speech at the end of intonational units. Tones
should be different as well, considering the prosodic structure
of spoken French.

Abstract

We analyze the problem of response suggestion in a closed domain along a real-world scenario of a digital library. We present a text-processing pipeline to generate question-answer pairs from chat transcripts. On this limited amount of training data, we compare retrieval-based, conditioned-generation, and dedicated representation learning approaches for response suggestion. Our results show that retrieval-based methods that strive to find similar, known contexts are preferable over parametric approaches from the conditioned-generation family, when the training data is limited. We, however, identify a specific representation learning approach that is competitive to the retrieval-based approaches despite the training data limitation.

Abstract

We present multi-modal adversarial autoencoders for recommendation and evaluate them on two different tasks: citation recommendation and subject label recommendation. We analyze the effects of adversarial regularization, sparsity, and different input modalities. By conducting 408 experiments, we show that adversarial regularization consistently improves the performance of autoencoders for recommendation. We demonstrate, however, that the two tasks differ in the semantics of item co-occurrence in the sense that item co-occurrence resembles relatedness in case of citations, yet implies diversity in case of subject labels. Our results reveal that supplying the partial item set as input is only helpful, when item co-occurrence resembles relatedness. When facing a new recommendation task it is therefore crucial to consider the semantics of item co-occurrence for the choice of an appropriate model.

Abstract

The last few years have revealed that several species may share the building blocks of Musicality with humans. The recognition of these building blocks (e.g., rhythm, frequency variation) was a necessary impetus for a new round of studies investigating rhythmic variation in animal vocal displays. Singing primates are a small group of primate species that produce modulated songs ranging from tens to thousands of vocal units. Previous studies showed that the indri, the only singing lemur, is currently the only known species that perform duet and choruses showing multiple rhythmic categories, as seen in human music. Rhythmic categories occur when temporal intervals between note onsets are not uniformly distributed, and rhythms with a small integer ratio between these intervals are typical of human music. Besides indris, white-handed gibbons and three crested gibbon species showed a prominent rhythmic category corresponding to a single small integer ratio, isochrony. This study reviews previous evidence on the co-occurrence of rhythmic categories in primates and focuses on the prospects for a comparative, multimodal study of rhythmicity in this clade.

Abstract

Despite the early emergence of pointing, children are generally not documented to produce iconic gestures until later in development. Although research has described this developmental trajectory and the types of iconic gestures that emerge first, there has been limited focus on iconic gestures within interactional contexts. This study identified the first 10 iconic gestures produced by five monolingual English-speaking children in a naturalistic longitudinal video corpus and analysed the interactional contexts. We found children produced their first iconic gesture between 12 and 20 months and that gestural types varied. Although 34% of gestures could have been imitated or derived from adult or child actions in the preceding context, the majority were produced independently of any observed model. In these cases, adults often led the interaction in a direction where iconic gesture was an appropriate response. Overall, we find infants can represent a referent symbolically and possess a greater capacity for innovation than previously assumed. In order to develop our understanding of how children learn to produce iconic gestures, it is important to consider the immediate interactional context. Conducting naturalistic corpus analyses could be a more ecologically valid approach to understanding how children learn to produce iconic gestures in real life contexts.

Abstract

Three groups of monolingual listeners, with Standard Chinese, Dutch and Hungarian as their native language, judged pairs of trisyllabic stimuli which differed only in their itch pattern. The segmental structure of the stimuli was made up by the experimenters and presented to subjects as being taken from a little-known language spoken on a South Pacific island. Pitch patterns consisted of a single rise-fall located on or near the second syllable. By and large, listeners selected the stimulus with the higher peak, the later eak, and the higher end rise as the one that signalled a question, regardless of language group. The result is argued to reflect innate, non-linguistic knowledge of the meaning of pitch variation, notably Ohala’s Frequency Code. A significant difference between groups is explained as due to the influence of the mother tongue.

Abstract

Face to face interaction is fundamental to human sociality but is very complex to study in a scientific fashion. This theme issue brings together cutting-edge approaches to the study of face-to-face interaction and showcases how we can make progress in this area. Researchers are now studying interaction in adult conversation, parent-child relationships, neurodiverse groups, interactions with virtual agents and various animal species. The theme issue reveals how new paradigms are leading to more ecologically grounded and comprehensive insights into what social interaction is. Scientific advances in this area can lead to improvements in education and therapy, better understanding of neurodiversity and more engaging artificial agents

Abstract

This paper investigates the time and space complexity of word order computation in the psycholinguistically motivated grammar formalism of Performance Grammar (PG). In PG, the first stage of syntax assembly yields an unordered tree ('mobile') consisting of a hierarchy of lexical frames (lexically anchored elementary trees). Associated with each lexica l frame is a linearizer—a Finite-State Automaton that locally computes the left-to-right order of the branches of the frame. Linearization takes place after the promotion component may have raised certain constituents (e.g. Wh- or focused phrases) into the domain of lexical frames higher up in the syntactic mobile. We show that the worst-case time and space complexity of analyzing input strings of length n is O(n5) and O(n4), respectively. This result compares favorably with the time complexity of word-order computations in Tree Adjoining Grammar (TAG). A comparison with Head-Driven Phrase Structure Grammar (HPSG) reveals that PG yields a more declarative linearization method, provided that the FSA is rewritten as an equivalent regular expression.

Abstract

There is evidence that the process of executing a planned utterance involves the use of both preceding-context and top-down cues. Utterance-initial words are cued only by the top-down plan. In contrast, non-initial words are cued both by top-down cues and preceding-context cues. Co-existence of both cue types raises the question of how they interact during learning. We argue that this interaction is competitive: items that tend to be preceded by predictive preceding-context cues are harder to activate from the plan without this predictive context. A novel computational model of this competition is developed. The model is tested on a corpus of repetition disfluencies and shown to account for the influences on patterns of restarts during production. In particular, this model predicts a novel Initiation Effect: following an interruption, speakers re-initiate production from words that tend to occur in utterance-initial position, even when they are not initial in the interrupted utterance.

Abstract

The representations generated by many mod-
els of language (word embeddings, recurrent
neural networks and transformers) correlate
to brain activity recorded while people read.
However, these decoding results are usually
based on the brain’s reaction to syntactically
and semantically sound language stimuli. In
this study, we asked: how does an LSTM (long
short term memory) language model, trained
(by and large) on semantically and syntac-
tically intact language, represent a language
sample with degraded semantic or syntactic
information? Does the LSTM representation
still resemble the brain’s reaction? We found
that, even for some kinds of nonsensical lan-
guage, there is a statistically significant rela-
tionship between the brain’s activity and the
representations of an LSTM. This indicates
that, at least in some instances, LSTMs and the
human brain handle nonsensical data similarly.

Abstract

This special publication aims to build a renewed enthusiasm for collecting acquisition data across many languages, including those facing endangerment and loss. It presents a guide for documenting and describing child language and child-directed language in diverse languages and cultures, as well as a collection of acquisition sketches based on this guide. The guide is intended for anyone interested in working across child language and language documentation, including, for example, field linguists and language documenters, community language workers, child language researchers or graduate students.

Abstract

What makes some words harder to learn than others in a second language? Although some robust factors have been identified based on small scale experimental studies, many relevant factors are difficult to study in such experiments due to the amount of data necessary to test them. Here, we investigate what factors affect the ease of learning of a word in a second language using a large data set of users learning English as a second language through the Duolingo mobile app. In a regression analysis, we test and confirm the well-studied effect of cognate status on word learning accuracy. Furthermore, we find significant effects for both cross-linguistic semantic alignment and English semantic density, two novel predictors derived from large scale distributional models of lexical semantics. Finally, we provide data on several other psycholinguistically plausible word level predictors. We conclude with a discussion of the limits, benefits and future research potential of using big data for investigating second language learning.

Abstract

A common assumption in the cognitive sciences is that artificial and natural language learning rely on shared mechanisms. However, attempts to bridge the two have yielded ambiguous results. We suggest that an empirical disconnect between the computations employed during learning and the methods employed at test may explain these mixed results. Further, we propose statistically-based chunking as a potential computational link between artificial and natural language learning. We compare the acquisition of non-adjacent dependencies to that of natural language structure using two types of tasks: reflection-based 2AFC measures, and processing-based recall measures, the latter being more computationally analogous to the processes used during language acquisition. Our results demonstrate that task-type significantly influences the correlations observed between artificial and natural language acquisition, with reflection-based and processing-based measures correlating within – but not across – task-type. These findings have fundamental implications for artificial-to-natural language comparisons, both methodologically and theoretically.

Abstract

Most bioacoustics software is used to analyse the already collected acoustics data in batch, i.e., after the data-collecting phase of a scientific study. However, experiments based on animal training require immediate and precise reactions from the experimenter, and thus do not easily dovetail with a typical bioacoustics workflow. Bridging this methodological gap, we have developed a custom application to live-monitor the vocal development of harbour seals in a behavioural experiment. In each trial, the application records and automatically detects an animal's call, and immediately measures duration and acoustic measures such as intensity, fundamental frequency, or formant frequencies. It then displays a spectrogram of the recording and the acoustic measurements, allowing the experimenter to instantly evaluate whether or not to reinforce the animal's vocalisation. From a technical perspective, the rapid and easy development of this custom software was made possible by combining multiple open-source software projects. Here, we integrated the acoustic analyses from Parselmouth, a Python library for Praat, together with PyAudio and Matplotlib's recording and plotting functionality, into a custom graphical user interface created with PyQt. This flexible recombination of different open-source Python libraries allows the whole program to be written in a mere couple of hundred lines of code

Abstract

n this study we investigated the durational correlates of lexical stress and pitch accent at normal and fast speech rate in Dutch. Previous literature on English shows that durations of lexically unstressed vowels are reduced more than stressed vowels when speakers increase their speech rate. We found that the same holds for Dutch, irrespective of whether the unstressed vowel is schwa or a "full" vowel. In the same line, we expected that vowels in words without a pitch accent would be shortened relatively more than vowels in words with a pitch accent. This was not the case: if anything, the accented vowels were shortened relatively more than the unaccented vowels. We conclude that duration is an important cue for lexical stress, but not for pitch accent.

Abstract

Studies on fast speech have shown that word-level timing of fast speech differs from that of normal rate speech in that unstressed syllables are shortened more than stressed syllables as speech rate increases. An earlier experiment showed that the intelligibility of time-compressed speech could not be improved by making its temporal organisation closer to natural fast speech. To test the hypothesis that segmental intelligibility is more important than prosodic timing in listening to timecompressed speech, the intelligibility of bisyllabic words was tested in three time-compression conditions: either stressed and unstressed syllable were compressed to the same degree, or the stressed syllable was compressed more than the unstressed syllable, or the reverse. As was found before, imitating wordlevel timing of fast speech did not improve intelligibility over linear compression. However, the results did not confirm the hypothesis either: there was no difference in intelligibility between the three compression conditions. We conclude that segmental intelligibility plays an important role, but further research is necessary to decide between the contributions of prosody and segmental intelligibility to the word-level intelligibility of time-compressed speech.

Abstract

Sign language is a rich form of communication, uniquely conveying meaning through a combination of gestures, facial expressions, and body movements. Existing research in sign language generation has predominantly focused on text-to-sign pose generation, while speech-to-sign pose generation remains relatively underexplored. Speech-to-sign language generation models can facilitate effective communication between the deaf and hearing communities. In this paper, we propose an architecture that utilises prosodic information from speech audio and semantic context from text to generate sign pose sequences. In our approach, we adopt a multi-tasking strategy that involves an additional task of predicting Facial Action Units (FAUs). FAUs capture the intricate facial muscle movements that play a crucial role in conveying specific facial expressions during sign language generation. We train our models on an existing Indian Sign language dataset that contains sign language videos with audio and text translations. To evaluate our models, we report Dynamic Time Warping (DTW) and Probability of Correct Keypoints (PCK) scores. We find that combining prosody and text as input, along with incorporating facial action unit prediction as an additional task, outperforms previous models in both DTW and PCK scores. We also discuss the challenges and limitations of speech-to-sign pose generation models to encourage future research in this domain. We release our models, results and code to foster reproducibility and encourage future research1.

Abstract

It has been suggested that some individuals may benefit more from social robots than do others. Using second
language (L2) as an example, the present study examined how individual differences in attitudes toward robots and personality
traits may be related to learning outcomes. Preliminary results with 24 Turkish-speaking adults suggest that negative attitudes
toward robots, more specifically thoughts and anxiety about the negative social impact that robots may have on the society,
predicted how well adults learned L2 words from a social robot. The possible implications of the findings as well as future directions are also discussed

Abstract

A central question in the psycholinguistic study of multilingualism is how syntax is shared across languages. We implement a model to investigate whether error-based implicit learning can provide an account of cross-language structural priming. The model is based on the Dual-path model of
sentence-production (Chang, 2002). We implement our model using the Bilingual version of Dual-path (Tsoukala, Frank, & Broersma, 2017). We answer two main questions: (1) Can structural priming of active and passive constructions occur between English and Spanish in a bilingual version of the Dual-
path model? (2) Does cross-language priming differ quantitatively from within-language priming in this model? Our results show that cross-language priming does occur in the model. This finding adds to the viability of implicit learning as an account of structural priming in general and cross-language
structural priming specifically. Furthermore, we find that the within-language priming effect is somewhat stronger than the cross-language effect. In the context of mixed results from
behavioral studies, we interpret the latter finding as an indication that the difference between cross-language and within-
language priming is small and difficult to detect statistically.

Abstract

A recurrently connected attractor neural network with a Hebbian learning rule is currently our best ANN analogy for a piece cortex. Functionally biological memory operates on a spectrum of time scales with regard to induction and retention, and it is modulated in complex ways by sub-cortical neuromodulatory systems. Moreover, biological memory networks are commonly believed to be highly distributed and engage many co-operating cortical areas. Here we focus on the temporal aspects of induction and retention of memory in a connectionist type attractor memory model of a piece of cortex. A continuous time, forgetful Bayesian-Hebbian learning rule is described and compared to the characteristics of LTP and LTD seen experimentally. More generally, an attractor network implementing this learning rule can operate as a long-term, intermediate-term, or short-term memory. Modulation of the print-now signal of the learning rule replicates some experimental memory phenomena, like e.g. the von Restorff effect.

Abstract

Interpreting the meaning of co-speech gesture often involves
identifying a gesture’s ‘lexical affiliate’, the word or phrase to
which it most closely relates (Schegloff 1984). Though there is
work within gesture studies that resists this simplex mapping of
meaning from speech to gesture (e.g. de Ruiter 2000; Kendon
2014; Parrill 2008), including an evolving body of literature on
recurrent gesture and gesture families (e.g. Fricke et al. 2014; Müller 2017), it is still the lexical affiliate model that is most ap-
parent in formal linguistic models of multimodal meaning(e.g.
Alahverdzhieva et al. 2017; Lascarides and Stone 2009; Puste-
jovsky and Krishnaswamy 2021; Schlenker 2020). In this work,
I argue that the lexical affiliate should be carefully reconsidered
in the further development of such models.
In place of the lexical affiliate, I suggest a further shift
toward a frame-based, action schematic approach to gestural
meaning in line with that proposed in, for example, Parrill and
Sweetser (2004) and Müller (2017). To demonstrate the utility
of this approach I present three types of compositional gesture
sequences which I call spatial contrast, spatial embedding, and
cooperative abstract deixis. All three rely on gestural context,
rather than gesture-speech alignment, to convey interactive (i.e.
pragmatic) meaning. The centrality of gestural context to ges-
ture meaning in these examples demonstrates the necessity of
developing a model of gestural meaning independent of its in-
tegration with speech.

Abstract

Citations play a crucial role in the scientific discourse, in information retrieval, and in bibliometrics. Many initiatives are currently promoting the idea of having free and open citation data. Creation of citation data, however, is not part of the cataloging workflow in libraries nowadays.
In this paper, we present our project Linked Open Citation Database, in which we design distributed processes and a system infrastructure based on linked data technology. The goal is to show that efficiently cataloging citations in libraries using a semi-automatic approach is possible. We specifically describe the current state of the workflow and its implementation. We show that we could significantly improve the automatic reference extraction that is crucial for the subsequent data curation. We further give insights on the curation and linking process and provide evaluation results that not only direct the further development of the project, but also allow us to discuss its overall feasibility.

Abstract

It is widely held that smells and flavors are impossible to put into words. In this paper we test this claim by seeking predictive patterns in wine reviews, which ostensibly aim to provide guides to perceptual content. Wine reviews have previously been critiqued as random and meaningless. We collected an English corpus of wine reviews with their structured metadata, and applied machine learning techniques to automatically predict the wine's color, grape variety, and country of origin. To train the three supervised classifiers, three different information sources were incorporated: lexical bag-of-words features, domain-specific terminology features, and semantic word embedding features. In addition, using regression analysis we investigated basic review properties, i.e., review length, average word length, and their relationship to the scalar values of price and review score. Our results show that wine experts do share a common vocabulary to describe wines and they use this in a consistent way, which makes it possible to automatically predict wine characteristics based on the review text alone. This means that odors and flavors may be more expressible in language than typically acknowledged.

Abstract

Languages differ in the degree of semantic flexibility of their syntactic roles. For example, Eng-
lish and Indonesian are considered more flexible with regard to the semantics of subjects,
whereas German and Japanese are less flexible. In Hawkins’ classification, more flexible lan-
guages are said to have a loose fit, and less flexible ones are those that have a tight fit. This
classification has been based on manual inspection of example sentences. The present paper
proposes a new, quantitative approach to deriving the measures of looseness and tightness from
corpora. We use corpora of online news from the Leipzig Corpora Collection in thirty typolog-
ically and genealogically diverse languages and parse them syntactically with the help of the
Universal Dependencies annotation software. Next, we compute Mutual Information scores for
each language using the matrices of lexical lemmas and four syntactic dependencies (intransi-
tive subjects, transitive subject, objects and obliques). The new approach allows us not only to
reproduce the results of previous investigations, but also to extend the typology to new lan-
guages. We also demonstrate that verb-final languages tend to have a tighter relationship be-
tween lexemes and syntactic roles, which helps language users to recognize thematic roles early
during comprehension.

Abstract

Large language models that exhibit instruction-following behaviour represent one of the biggest recent upheavals in conversational interfaces, a trend in large part fuelled by the release of OpenAI's ChatGPT, a proprietary large language model for text generation fine-tuned through reinforcement learning from human feedback (LLM+RLHF). We review the risks of relying on proprietary software and survey the first crop of open-source projects of comparable architecture and functionality. The main contribution of this paper is to show that openness is differentiated, and to offer scientific documentation of degrees of openness in this fast-moving field. We evaluate projects in terms of openness of code, training data, model weights, RLHF data, licensing, scientific documentation, and access methods. We find that while there is a fast-growing list of projects billing themselves as 'open source', many inherit undocumented data of dubious legality, few share the all-important instruction-tuning (a key site where human labour is involved), and careful scientific documentation is exceedingly rare. Degrees of openness are relevant to fairness and accountability at all points, from data collection and curation to model architecture, and from training and fine-tuning to release and deployment.

Abstract

Speech recognition systems are a key intermediary in voice-driven human-computer interaction. Although speech recognition works well for pristine monologic audio, real-life use cases in open-ended interactive settings still present many challenges. We argue that timing is mission-critical for dialogue systems, and evaluate 5 major commercial ASR systems for their conversational and multilingual support. We find that word error rates for natural conversational data in 6 languages remain abysmal, and that overlap remains a key challenge (study 1). This impacts especially the recognition of conversational words (study 2), and in turn has dire consequences for downstream intent recognition (study 3). Our findings help to evaluate the current state of conversational ASR, contribute towards multidimensional error analysis and evaluation, and identify phenomena that need most attention on the way to build robust interactive speech technologies.

Abstract

We present the Narrative Brain Dataset, an fMRI dataset that was collected during spoken presentation of short excerpts of three
stories in Dutch. Together with the brain imaging data, the dataset contains the written versions of the stimulation texts. The texts are
accompanied with stochastic (perplexity and entropy) and semantic computational linguistic measures. The richness and unconstrained
nature of the data allows the study of language processing in the brain in a more naturalistic setting than is common for fMRI studies.
We hope that by making NBD available we serve the double purpose of providing useful neural data to researchers interested in natural
language processing in the brain and to further stimulate data sharing in the field of neuroscience of language.

Abstract

Children learn language from the speech in their home environment. Recent work shows that more infant-directed speech
(IDS) leads to stronger lexical development. But what makes IDS a particularly useful learning signal? Here, we expand on an attention-based account first proposed by Räsänen et al. (2018): that prosodic modifications make IDS less predictable, and thus more interesting. First, we reproduce the critical finding from Räsänen et al.: that lab-recorded IDS pitch is less predictable compared to adult-directed speech (ADS). Next, we show that this result generalizes to the home language environment, finding that IDS in daylong recordings is also less predictable than ADS but that this pattern is much less robust than for IDS recorded in the lab. These results link experimental work on attention and prosodic modifications of IDS to real-world language-learning environments, highlighting some challenges of scaling up analyses of IDS to larger datasets that better capture children’s actual input.

Abstract

For (semi-)automated subject indexing systems in digital libraries, it is often more practical to use metadata such as the title of a publication instead of the full-text or the abstract. Therefore, it is desirable to have good text mining and text classification algorithms that operate well already on the title of a publication. So far, the classification performance on titles is not competitive with the performance on the full-texts if the same number of training samples is used for training. However, it is much easier to obtain title data in large quantities and to use it for training than full-text data. In this paper, we investigate the question how models obtained from training on increasing amounts of title training data compare to models from training on a constant number of full-texts. We evaluate this question on a large-scale dataset from the medical domain (PubMed) and from economics (EconBiz). In these datasets, the titles and annotations of millions of publications are available, and they outnumber the available full-texts by a factor of 20 and 15, respectively. To exploit these large amounts of data to their full potential, we develop three strong deep learning classifiers and evaluate their performance on the two datasets. The results are promising. On the EconBiz dataset, all three classifiers outperform their full-text counterparts by a large margin. The best title-based classifier outperforms the best full-text method by 9.4%. On the PubMed dataset, the best title-based method almost reaches the performance of the best full-text classifier, with a difference of only 2.9%.

Abstract

Lexical stress is realised similarly in English, German, and
Dutch. On a suprasegmental level, stressed syllables tend to be
longer and more acoustically salient than unstressed syllables;
segmentally, vowels in unstressed syllables are often reduced.
The frequency of unreduced unstressed syllables (where only
the suprasegmental cues indicate lack of stress) however,
differs across the languages. The present studies test whether
listener behaviour is affected by these vocabulary differences,
by investigating German listeners’ use of suprasegmental cues
to lexical stress in German and English word recognition. In a
forced-choice identification task, German listeners correctly
assigned single-syllable fragments (e.g., Kon-) to one of two
words differing in stress (KONto, konZEPT). Thus, German
listeners can exploit suprasegmental information for
identifying words. German listeners also performed above
chance in a similar task in English (with, e.g., DIver, diVERT),
i.e., their sensitivity to these cues also transferred to a nonnative
language. An English listener group, in contrast, failed
in the English fragment task. These findings mirror vocabulary
patterns: German has more words with unreduced unstressed
syllables than English does.

Abstract

Lexical knowledge influences how human listeners make decisions about speech sounds. Positive lexical effects (faster responses to target sounds in words than in nonwords) are robust across several laboratory tasks, while negative effects (slower responses to targets in more word-like nonwords than in less word-like nonwords) have been found in phonetic decision tasks but not phoneme monitoring tasks. The present experiments tested whether negative lexical effects are therefore a task-specific consequence of the forced choice required in phonetic decision. We compared phoneme monitoring and phonetic decision performance using the same Dutch materials in each task. In both experiments there were positive lexical effects, but no negative lexical effects. We observe that in all studies showing negative lexical effects, the materials were made by cross-splicing, which meant that they contained perceptual evidence supporting the lexically-consistent phonemes. Lexical knowledge seems to influence phonemic decision-making only when there is evidence for the lexically-consistent phoneme in the speech signal.

Abstract

We briefly describe the Merge model of phonemic decision-making, and, in the light of general arguments about the possible role of feedback in spoken-word recognition, defend Merge's feedforward structure. Merge not only accounts adequately for the data, without invoking feedback connections, but does so in a parsimonious manner.

Abstract

This paper presents lexical statistics on the pattern of occurrence of words embedded in other words. We report the results of an analysis of 25000 words, varying in length from two to six syllables, extracted from a phonetically-coded English dictionary (The Longman Dictionary of Contemporary English). Each syllable, and each string of syllables within each word was checked against the dictionary. Two analyses are presented: the first used a complete list of polysyllables, with look-up on the entire dictionary; the second used a sublist of content words, counting only embedded words which were themselves content words. The results have important implications for models of human speech recognition. The efficiency of these models depends, in different ways, on the number and location of words within words.

Abstract

The ultimate goal of our research is to improve an existing speech-based computational model of human speech recognition on the task of simulating the role of fine-grained phonetic information in human speech processing. As part of this work we are investigating articulatory feature classifiers that are able to create reliable and accurate transcriptions of the articulatory behaviour encoded in the acoustic speech signal. Articulatory feature (AF) modelling of speech has received a considerable amount of attention in automatic speech recognition research. Different approaches have been used to build AF classifiers, most notably multi-layer perceptrons. Recently, deep neural networks have been applied to the task of AF classification. This paper aims to improve AF classification by investigating two different approaches: 1) investigating the usefulness of a deep Convolutional neural network (CNN) for AF classification; 2) integrating the Mel filtering operation into the CNN architecture. The results showed a remarkable improvement in classification accuracy of the CNNs over state-of-the-art AF classification results for Dutch, most notably in the minority classes. Integrating the Mel filtering operation into the CNN architecture did not further improve classification performance.

Abstract

Analyzing EEG signals recorded while participants are listening to continuous speech with the purpose of testing linguistic hypotheses is complicated by the fact that the signals simultaneously reflect exogenous acoustic excitation and endogenous linguistic processing. This makes it difficult to trace subtle differences that occur in mid-sentence position. We apply an analysis based on multivariate temporal response functions to uncover subtle mid-sentence effects. This approach is based on a per-stimulus estimate of the response of the neural system to speech input. Analyzing EEG signals predicted on the basis of the response functions might then bring to light conditionspecific differences in the filtered signals. We validate this approach by means of an analysis of EEG signals recorded with isolated word stimuli. Then, we apply the validated method to the analysis of the responses to the same words in the middle of meaningful sentences.

Abstract

Many studies investigating the temporal alignment of co-speech
gestures to acoustic units in the speech signal find a close
coupling of the gestural landmarks and pitch accents or the
stressed syllable of pitch-accented words. In English, a pitch
accent is anchored in the lexically stressed syllable. Hence, it is
unclear whether it is the lexical phonological dimension of
stress, or the phrase-level prominence that determines the
details of speech-gesture synchronization. This paper explores
the relation between gestural phase transitions and accentual F0
peaks in Stockholm Swedish, which exhibits a lexical pitch
accent distinction. When produced with phrase-level
prominence, there are three different configurations of
lexicality of F0 peaks and the status of the syllable it is aligned
with. Through analyzing the alignment of the different F0 peaks
with gestural onsets in spontaneous dyadic conversations, we
aim to contribute to our understanding of the role of lexical
prosodic phonology in the co-production of speech and gesture.
The results, though limited by a small dataset, still suggest
differences between the three types of peaks concerning which
types of gesture phase onsets they tend to align with, and how
well these landmarks align with each other, although these
differences did not reach significance.

Abstract

Two word-spotting experiments are reported that examine whether the Possible-Word Constraint (PWC) [1] is a language-specific or language-universal strategy for the segmentation of continuous speech. The PWC disfavours parses which leave an impossible residue between the end of a candidate word and a known boundary. The experiments examined cases where the residue was either a CV syllable with a lax vowel, or a CVC syllable with a schwa. Although neither syllable context is a possible word in English, word-spotting in both contexts was easier than with a context consisting of a single consonant. The PWC appears to be language-universal rather than language-specific.

Abstract

We report two experiments in which vowels and steady-state consonants served as targets in a speeded detection task. In the first experiment, two vowels were compared with one voiced and once unvoiced fricative. Response times (RTs) to the vowels were longer than to the fricatives. The error rate was higher for the consonants. Consonants in word-final position produced the shortest RTs, For the vowels, RT correlated negatively with target duration. In the second experiment, the same two vowel targets were compared with two nasals. This time there was no significant difference in RTs, but the error rate was still significantly higher for the consonants. Error rate and length correlated negatively for the vowels only. We conclude that RT differences between phonemes are independent of vocalic or consonantal status. Instead, we argue that the process of phoneme detection reflects more finely grained differences in acoustic/articulatory structure within the phonemic repertoire.

Abstract

Simulations explored the inability of the TRACE model of spoken-word recognition to model the effects on human listening of subcategorical mismatch in word forms. The source of TRACE's failure lay not in interactive connectivity, not in the presence of inter-word competition, and not in the use of phonemic representations, but in the need for continuously optimised interpretation of the input. When an analogue of TRACE was allowed to cycle to asymptote on every slice of input, an acceptable simulation of the subcategorical mismatch data was achieved. Even then, however, the simulation was not as close as that produced by the Merge model, which has inter-word competition, phonemic representations and continuous optimisation (but no interactive connectivity).

Abstract

Phonetic convergence—i.e., adapting one’s speech
towards that of an interlocutor—has been shown
to occur in human-human conversations as well as
human-machine interactions. Here, we investigate
the hypothesis that human-to-robot convergence is
influenced by the human’s perception of the robot
and by the conversation’s topic. We conducted a
within-subjects experiment in which 33 participants
interacted with two robots differing in their eye gaze
behavior—one looked constantly at the participant;
the other produced gaze aversions, similarly to a
human’s behavior. Additionally, the robot asked
questions with increasing intimacy levels.
We observed that the speakers tended to converge
on F0 to the robots. However, this convergence
to the robots was not modulated by how the
speakers perceived them or by the topic’s intimacy.
Interestingly, speakers produced lower F0 means
when talking about more intimate topics. We
discuss these findings in terms of current theories of
conversational convergence.

Abstract

A database is presented of relative familiarity ratings for 24 sets of Japanese words, each set comprising words overlapping in the initial portions. These ratings are useful for the generation of material sets for research in the recognition of spoken words.

Abstract

Automating robot behavior in a specific situation is an active area of research. There are several approaches available in the literature of robotics to cater for the automatic behavior of a robot. However, when it comes to humanoids or human-robot interaction in general, the area has been less explored. In this paper, a pseudo-randomization approach has been introduced to automatize the gestures and facial expressions of an interactive humanoid robot called ROBIN based on its mental state. A significant number of gestures and facial expressions have been implemented to allow the robot more options to perform a relevant action or reaction based on visual stimuli. There is a display of noticeable differences in the behaviour of the robot for the same stimuli perceived from an interaction partner. This slight autonomous behavioural change in the robot clearly shows a notion of automation in behaviour. The results from experimental scenarios and human-centered evaluation of the system help validate the approach.

Abstract

Word count estimation (WCE) from audio recordings has a number of applications, including quantifying the amount of speech that language-learning infants hear in their natural environments, as captured by daylong recordings made with devices worn by infants. To be applicable in a wide range of scenarios and also low-resource domains, WCE tools should be extremely robust against varying signal conditions and require minimal access to labeled training data in the target domain. For this purpose, earlier work has used automatic syllabification of speech, followed by a least-squares-mapping of syllables to word counts. This paper compares a number of previously proposed syllabifiers in the WCE task, including a supervised bi-directional long short-term memory (BLSTM) network that is trained on a language for which high quality syllable annotations are available (a “high resource language”), and reports how the alternative methods compare on different languages and signal conditions. We also explore additive noise and varying-channel data augmentation strategies for BLSTM training, and show how they improve performance in both matching and mismatching languages. Intriguingly, we also find that even though the BLSTM works on languages beyond its training data, the unsupervised algorithms can still outperform it in challenging signal conditions on novel languages.