Publications

Abstract

Proficient readers typically fixate near the center of a word, with a slight bias towards word onset. We explore a novel account of this phenomenon based on combining information-theory with visual perceptual constraints in a connectionist model of visual word recognition. This account posits that the amount of information-content available for word identification varies across fixation locations and across languages, thereby explaining the overall fixation location bias in different languages, making the novel prediction that certain words are more readily identified when fixating at an atypical fixation location, and predicting specific cross-linguistic differences. We tested these predictions across several simulations in English and Hebrew, and in a pilot behavioral experiment. Results confirmed that the bias to fixate closer to word onset aligns with maximizing information in the visual signal, that some words are more readily identified at atypical fixation locations, and that these effects vary to some degree across languages.

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

A production experiment investigated the tonal shape of Finnish finite verbs in transitive sentences without narrow focus. Traditional descriptions of Finnish stating that non-focused finite verbs do not receive accents were only partly supported. Verbs were found to have a consistently smaller pitch range than words in other word classes, but their pitch contours were neither flat nor explainable by pure interpolation.

Abstract

In the paper we describe a new attempt to come to automatic detectors processing real scene audio-video streams that can be used by researchers world-wide to speed up their annotation and analysis work. Typically these recordings are taken in field and experimental situations mostly with bad quality and only little corpora preventing to use standard stochastic pattern recognition techniques. Audio/video processing components are taken out of the expert lab and are integrated in easy-to-use interactive frameworks so that the researcher can easily start them with modified parameters and can check the usefulness of the created annotations. Finally a variety of detectors may have been used yielding a lattice of annotations. A flexible search engine allows finding combinations of patterns opening completely new analysis and theorization possibilities for the researchers who until were required to do all annotations manually and who did not have any help in pre-segmenting lengthy media recordings.

Abstract

Annotation of digital recordings in humanities research still is, to a largeextend, a process that is performed manually. This paper describes the firstpattern recognition based software components developed in the AVATecH projectand their integration in the annotation tool ELAN. AVATecH (AdvancingVideo/Audio Technology in Humanities Research) is a project that involves twoMax Planck Institutes (Max Planck Institute for Psycholinguistics, Nijmegen,Max Planck Institute for Social Anthropology, Halle) and two FraunhoferInstitutes (Fraunhofer-Institut für Intelligente Analyse- undInformationssysteme IAIS, Sankt Augustin, Fraunhofer Heinrich-Hertz-Institute,Berlin) and that aims to develop and implement audio and video technology forsemi-automatic annotation of heterogeneous media collections as they occur inmultimedia based research. The highly diverse nature of the digital recordingsstored in the archives of both Max Planck Institutes, poses a huge challenge tomost of the existing pattern recognition solutions and is a motivation to makesuch technology available to researchers in the humanities.

Abstract

Data augmentation techniques have been widely used in visual recognition tasks as it is easy to generate new
data by simple and straight forward image transformations. However, when it comes to text data augmen-
tations, it is difficult to find appropriate transformation techniques which also preserve the contextual and
grammatical structure of language texts. In this paper, we explore various text data augmentation techniques
in text space and word embedding space. We study the effect of various augmented datasets on the efficiency
of different deep learning models for relation classification in text.

Abstract

The event-related potential (ERP) component named phonological mismatch negativity (PMN) arises when listeners hear an unexpected word form in a spoken sentence [1]. The PMN is thought to reflect the mismatch between expected and perceived auditory speech input. In this paper, we use the PMN to test a central premise in the predictive coding framework [2], namely that the mismatch between prior expectations and sensory input is an important mechanism of perception. We test this with natural speech materials containing approximately 50,000 word tokens. The corresponding EEG-signal was recorded while participants (n = 48) listened to these materials. Following [3], we quantify the mismatch with two word probability distributions (WPD): a WPD based on preceding context, and a WPD that is additionally updated based on the incoming audio of the current word. We use the between-WPD cross entropy for each word in the utterances and show that a higher cross entropy correlates with a more negative PMN. Our results show that listeners anticipate auditory input while processing each word in naturalistic speech. Moreover, complementing previous research, we show that predictive language processing occurs across the whole probability spectrum.

Abstract

The mismatch between top-down predicted and bottom-up perceptual input is an important mechanism of perception according to the predictive coding framework (Friston, [1]). In this paper we develop and validate a new information-theoretic measure that quantifies the mismatch between expected and observed auditory input during speech processing. We argue that such a mismatch measure is useful for the study of speech processing. To compute the mismatch measure, we use naturalistic speech materials containing approximately 50,000 word tokens. For each word token we first estimate the prior word probability distribution with the aid of statistical language modelling, and next use automatic speech recognition to update this word probability distribution based on the unfolding speech signal. We validate the mismatch measure with multiple analyses, and show that the auditory-based update improves the probability of the correct word and lowers the uncertainty of the word probability distribution. Based on these results, we argue that it is possible to explicitly estimate the mismatch between predicted and perceived speech input with the cross entropy between word expectations computed before and after an auditory update.

Abstract

1. Introduction External input is necessary to acquire language. Consequently, the comprehension of various constituents of language, such as lexical items or syntactic and semantic structures should emerge at the same time as or even precede their production. However, in the case of pronouns this general assumption does not seem to hold. On the contrary, while children at the age of four use pronouns and reflexives appropriately during production (de Villiers, et al. 2006), a number of comprehension studies across different languages found chance performance in pronoun trials up to the age of seven, which co-occurs with a high level of accuracy in reflexive trials (for an overview see e.g. Conroy, et al. 2009; Elbourne 2005).

Abstract

In the present paper we show that a general-purpose word learning model can simulate several important findings from recent experiments in language acquisition. Both the addition of background noise and varying the speaker have been found to influence infants’ performance during word recognition experiments. We were able to replicate this behaviour in our artificial word learning agent. We use the results to discuss both advantages and limitations of computational models of language acquisition.

Abstract

Why do people accommodate to each other’s linguistic behavior? Studies of natural interactions (Giles, Taylor & Bourhis, 1973) suggest that speakers accommodate to achieve interactional goals, influencing what their interlocutor thinks or feels about them. But is this the only reason speakers accommodate? In real-world conversations, interactional motivations are ubiquitous, making it difficult to assess the extent to which they drive accommodation. Do speakers still accommodate even when interactional goals cannot be achieved, for instance, when their interlocutor cannot interpret their accommodation behavior? To find out, we asked participants to enter an immersive virtual reality (VR) environment and to converse with a virtual interlocutor. Participants accommodated to the speech rate of their virtual interlocutor even though he could not interpret their linguistic behavior, and thus accommodation could not possibly help them to achieve interactional goals. Results show that accommodation does not require explicit interactional goals, and suggest other social motivations for accommodation.

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

We describe our computer-supported framework to overcome the rule of metadata schism. It combines the use of controlled vocabularies, managed by a data category registry, with a component-based approach, where the categories can be combined to yield complex metadata structures. A metadata scheme devised in this way will thus be grounded in its use of categories. Schema designers will profit from existing prefabricated larger building blocks, motivating re-use at a larger scale. The common base of any two metadata schemes within this framework will solve, at least to a good extent, the semantic interoperability problem, and consequently, further promote systematic use of metadata for existing resources and tools to be shared.

Abstract

Unlike most of the world's languages, Korean distinguishes three types of voiceless stops, namely lenis, fortis, and aspirated stops. All occur at three places of articulation. In previous work, acoustic measurements are mostly collapsed over the three places of articulation. This study therefore provides acoustic measurements of Korean lenis, fortis, and aspirated stops at all three places of articulation separately. Clear differences are found among the acoustic characteristics of the stops at the different places of articulation

Abstract

We investigated the extent to which emotionally valenced words automatically cue spatio-motor representations. Participants made speeded button presses, moving their hand upward or downward while viewing words with positive or negative valence. Only the color of the words was relevant to the response; on target trials, there was no requirement to read the words or process their meaning. In Experiment 1, upward responses were faster for positive words, and downward for negative words. This effect was extinguished, however, when words were repeated. In Experiment 2, participants performed the same primary task with the addition of distractor trials. Distractors either oriented attention toward the words’ meaning or toward their color. Congruity effects were increased with orientation to meaning, but eliminated with orientation to color. When people read words with emotional valence, vertical spatio-motor representations are activated highly automatically, but this automaticity is modulated by repetition and by attentional orientation to the words’ form or meaning.

Abstract

This paper investigates whether surprisal theory can account for differential processing difficulty in the NP-/S-coordination ambiguity in Dutch. Surprisal is estimated using a Probabilistic Context-Free Grammar (PCFG), which is induced from an automatically annotated corpus. We find that our lexicalized surprisal model can account for the reading time data from a classic experiment on this ambiguity by Frazier (1987). We argue that syntactic and lexical probabilities, as specified in a PCFG, are sufficient to account for what is commonly referred to as an NP-coordination preference.

Abstract

Speech input causes listeners to activate multiple
candidate words which then compete with one
another. These include onset competitors, that share a
beginning (bumper, butter), but also, counterintuitively,
rhyme competitors, sharing an ending
(bumper, jumper). In L1, competition is typically
stronger for onset than for rhyme. In L2, onset
competition has been attested but rhyme competition
has heretofore remained largely unexamined. We
assessed L1 (Dutch) and L2 (English) word
recognition by the same late-bilingual individuals. In
each language, eye gaze was recorded as listeners
heard sentences and viewed sets of drawings: three
unrelated, one depicting an onset or rhyme competitor
of a word in the input. Activation patterns revealed
substantial onset competition but no significant
rhyme competition in either L1 or L2. Rhyme
competition may thus be a “luxury” feature of
maximally efficient listening, to be abandoned when
resources are scarcer, as in listening by late
bilinguals, in either language.

Abstract

Across cultures, people conceptualize time as if it flows along a horizontal timeline, but the direction of this implicit timeline is culture-specific: in cultures with left-to-right orthography (e.g., English-speaking cultures) time appears to flow rightward, but in cultures with right-to-left orthography (e.g., Arabic-speaking cultures) time flows leftward. Can orthography influence implicit time representations independent of other cultural and linguistic factors? Native Dutch speakers performed a space-time congruity task with the instructions and stimuli written in either standard Dutch or mirror-reversed Dutch. Participants in the Standard Dutch condition were fastest to judge past-oriented phrases by pressing the left button and future-oriented phrases by pressing the right button. Participants in the Mirror-Reversed Dutch condition showed the opposite pattern of reaction times, consistent with results found previously in native Arabic and Hebrew speakers. These results demonstrate a causal role for writing direction in shaping implicit mental representations of time.

Abstract

Previous studies on focus marking in French have shown that post-focus deaccentuation, phrasing and phonetic cues like peak height and duration are employed to encode narrow focus but tonal patterns appear to be irrelevant. These studies either examined Standard French or did not control for the regional varieties spoken by the speakers. The present study investigated the use of all these cues in expressing narrow focus in naturally spoken declarative sentences in Toulousian French. It was found that similar to Standard French, Toulousian French uses post-focus deaccentuation and phrasing to mark focus. Different from Standard French, Toulousian French does not use the phonetic cues but use tonal patterns to encode focus. Tonal patterns ending with H\% occur more frequently in the VPs when the subject is in focus but tonal patterns ending with L\% occur more frequently in the VPs when the object is in focus. Our study thus provides a first insight into the similarities and differences in focus marking between Toulousian French and Standard French.

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

Part of the remarkable efficiency of listening is
accommodation to unfamiliar talkers’ specific
pronunciations by retuning of phonemic intercategory
boundaries. Such retuning occurs in second
(L2) as well as first language (L1); however, recent
research with emigrés revealed successful adaptation
in the environmental L2 but, unprecedentedly, not in
L1 despite continuing L1 use. A possible explanation
involving relative exposure to novel talkers is here
tested in heritage language users with Mandarin as
family L1 and English as environmental language. In
English, exposure to an ambiguous sound in
disambiguating word contexts prompted the expected
adjustment of phonemic boundaries in subsequent
categorisation. However, no adjustment occurred in
Mandarin, again despite regular use. Participants
reported highly asymmetric interlocutor counts in the
two languages. We conclude that successful retuning
ability requires regular exposure to novel talkers in
the language in question, a criterion not met for the
emigrés’ or for these heritage users’ L1.

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

One of a listener's major tasks in understanding continuous speech is segmenting the speech signal into separate words. When listening conditions are difficult, speakers can help listeners by deliberately speaking more clearly. In three experiments, we examined how word boundaries are produced in deliberately clear speech. We found that speakers do indeed attempt to mark word boundaries; moreover, they differentiate between word boundaries in a way which suggests they are sensitive to listener needs. Application of heuristic segmentation strategies makes word boundaries before strong syllables easiest for listeners to perceive; but under difficult listening conditions speakers pay more attention to marking word boundaries before weak syllables, i.e. they mark those boundaries which are otherwise particularly hard to perceive.

Abstract

Two lexical decision studies examined the effects of single-phoneme mismatches on lexical activation in spoken-word recognition. One study was carried out in English, and involved spoken primes and visually presented lexical decision targets. The other study was carried out in Dutch, and primes and targets were both presented auditorily. Facilitation was found only for spoken targets preceded immediately by spoken primes; no facilitation occurred when targets were presented visually, or when intervening input occurred between prime and target. The effects of vowel mismatches and consonant mismatches were equivalent.

Abstract

Recognising continuous speech in a second language is often unexpectedly difficult, as the operation of segmenting speech is so attuned to native-language structure. We report the initial steps in development of a novel training method for second-language listening, focusing on speech segmentation and employing a task designed for studying this: word-spotting. Listeners detect real words in sequences consisting of a word plus a minimal context. The present validation study shows that learners from varying non-English backgrounds successfully perform a version of this task in English, and display appropriate sensitivity to structural factors that also affect segmentation by native English listeners.

Abstract

Do interlocutors adjust their conversational strategies to the specific contextual demands of a given situation? Prior studies have yielded conflicting results, making it unclear how strategies vary with demands. We combine insights from qualitative and quantitative approaches in a within-participant experimental design involving two different contexts: spontaneously occurring conversations (SOC) and task-oriented conversations (TOC). We systematically assess backchanneling, other-repair and linguistic alignment. We find that SOC exhibit a higher number of backchannels, a reduced and more generic repair format and higher rates of lexical and syntactic alignment. TOC are characterized by a high number of specific repairs and a lower rate of lexical and syntactic alignment. However, when alignment occurs, more linguistic forms are aligned. The findings show that conversational strategies adapt to specific contextual demands.

Abstract

Speakers often use spatial metaphors to talk about musical pitch (e.g., a low note, a high soprano). Previous experiments suggest that English speakers also think about pitches as high or low in space, even when theyʼre not using language or musical notation (Casasanto, 2010). Do metaphors in language merely reflect pre-existing associations between space and pitch, or might language also shape these non-linguistic metaphorical mappings? To investigate the role of language in pitch tepresentation, we conducted a pair of non-linguistic spacepitch interference experiments in speakers of two languages that use different spatial metaphors. Dutch speakers usually describe pitches as ʻhighʼ (hoog) and ʻlowʼ (laag). Farsi speakers, however, often describe high-frequency pitches as ʻthinʼ (naazok) and low-frequency pitches as ʻthickʼ (koloft). Do Dutch and Farsi speakers mentally represent pitch differently? To find out, we asked participants to reproduce musical pitches that they heard in the presence of irrelevant spatial information (i.e., lines that varied either in height or in thickness). For the Height Interference experiment, horizontal lines bisected a vertical reference line at one of nine different locations. For the Thickness Interference experiment, a vertical line appeared in the middle of the screen in one of nine thicknesses. In each experiment, the nine different lines were crossed with nine different pitches ranging from C4 to G#4 in semitone increments, to produce 81 distinct trials. If Dutch and Farsi speakers mentally represent pitch the way they talk about it, using different kinds of spatial representations, they should show contrasting patterns of cross-dimensional interference: Dutch speakersʼ pitch estimates should be more strongly affected by irrelevant height information, and Farsi speakersʼ by irrelevant thickness information. As predicted, Dutch speakersʼ pitch estimates were significantly modulated by spatial height but not by thickness. Conversely, Farsi speakersʼ pitch estimates were modulated by spatial thickness but not by height (2x2 ANOVA on normalized slopes of the effect of space on pitch: F(1,71)=17,15 p<.001). To determine whether language plays a causal role in shaping pitch representations, we conducted a training experiment. Native Dutch speakers learned to use Farsi-like metaphors, describing pitch relationships in terms of thickness (e.g., a cello sounds ʻthickerʼ than a flute). After training, Dutch speakers showed a significant effect of Thickness interference in the non-linguistic pitch reproduction task, similar to native Farsi speakers: on average, pitches accompanied by thicker lines were reproduced as lower in pitch (effect of thickness on pitch: r=-.22, p=.002). By conducting psychophysical tasks, we tested the ʻWhorfianʼ question without using words. Yet, results also inform theories of metaphorical language processing. According to psycholinguistic theories (e.g., Bowdle & Gentner, 2005), highly conventional metaphors are processed without any active mapping from the source to the target domain (e.g., from space to pitch). Our data, however, suggest that when people use verbal metaphors they activate a corresponding non-linguistic mapping from either height or thickness to pitch, strengthening this association at the expense of competing associations. As a result, people who use different metaphors in their native languages form correspondingly different representations of musical pitch. Casasanto, D. (2010). Space for Thinking. In Language, Cognition and Space: State of the art and new directions. V. Evans & P. Chilton (Eds.), 453-478, London: Equinox Publishing. Bowdle, B. & Gentner, D. (2005). The career of metaphor. Psychological Review, 112, 193-216.

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

Previous studies have shown that speakers align their speech to each other at multiple linguistic levels. This study investigates whether alignment is mostly the result of priming from the immediately preceding
speech materials, focussing on pitch and articulation rate (AR). Native Dutch speakers completed sentences, first by themselves (pre-test), then in alternation with Confederate 1 (Round 1), with Confederate 2 (Round 2), with Confederate 1 again
(Round 3), and lastly by themselves again (post-test). Results indicate that participants aligned to the confederates and that this alignment lasted during the post-test. The confederates’ directly preceding sentences were not good predictors for the participants’ pitch and AR. Overall, the results indicate that alignment is more of a global effect than a local priming effect.

Abstract

Pre-lexical representations of speech sounds have been to shown to change dynamically through a mechanism of lexically driven learning. [Norris et al. (2003).] Here we investigated whether this type of learning occurs in native British English (BE) listeners for a word-final stop contrast which is commonly de-voiced in Dutch-accented English. Specifically, this study asked whether the change in pre-lexical representation also encodes information about the position of the critical sound within a word. After exposure to a native Dutch speaker's productions of de-voiced stops in word-final position (but not in any other positions), BE listeners showed evidence of perceptual learning in a subsequent cross-modal priming task, where auditory primes with voiceless final stops (e.g., [si:t], “seat”) facilitated recognition of visual targets with voiced final stops (e.g., “seed”). This learning generalized to test pairs where the critical contrast was in word-initial position, e.g., auditory primes such as [taun] (“town”), facilitated recognition of visual targets like “down”. Control listeners, who had not heard any stops by the speaker during exposure, showed no learning effects. The results suggest that under these exposure conditions, word position is not encoded in the pre-lexical adjustment to the accented phoneme contras

Abstract

This paper shows that the dictation task, a well-
known testing instrument in language education, has
untapped potential as a research tool for studying
speech perception. We describe how transcriptions
can be scored on measures of lexical, orthographic,
phonological, and semantic similarity to target
phrases to provide comprehensive information about
accuracy at different processing levels. The former
three measures are automatically extractable,
increasing objectivity, and the middle two are
gradient, providing finer-grained information than
traditionally used. We evaluate the measures in an
English dictation task featuring phonetically reduced
continuous speech. Whereas the lexical and
orthographic measures emphasize listeners’ word
identification difficulties, the phonological measure
demonstrates that listeners can often still recover
phonological features, and the semantic measure
captures their ability to get the gist of the utterances.
Correlational analyses and a discussion of practical
and theoretical considerations show that combining
multiple measures improves the dictation task’s
utility as a research tool.

Abstract

Lexically guided perceptual learning has traditionally been studied with ambiguous consonant sounds to which native listeners are exposed in a purely receptive listening context. To extend previous research, we investigate whether lexically guided learning applies to a vowel shift encountered by non-native listeners in an interactive dialogue. Dutch participants played a two-player game in English in either a control condition, which contained no evidence for a vowel shift, or a lexically constraining condition, in which onscreen lexical information required them to re-interpret their interlocutor’s /ɪ/ pronunciations as representing /ε/. A phonetic categorization pre-test and post-test were used to assess whether the game shifted listeners’ phonemic boundaries such that more of the /ε/-/ɪ/ continuum came to be perceived as /ε/. Both listener groups showed an overall post-test shift toward /ɪ/, suggesting that vowel perception may be sensitive to directional biases related to properties of the speaker’s vowel space. Importantly, listeners in the lexically constraining condition made relatively more post-test /ε/ responses than the control group, thereby exhibiting an effect of lexically guided adaptation. The results thus demonstrate that non-native listeners can adjust their phonemic boundaries on the basis of lexical information to accommodate a vowel shift learned in interactive conversation.

Abstract

The problem of auxiliary fronting in complex polar questions occupies a prominent position within the nature versus nurture controversy in language acquisition. We employ a model of statistical learning which uses sequential and semantic information to produce utterances from a bag of words. This linear learner is capable of generating grammatical questions without exposure to these structures in its training environment. We also demonstrate that the model performs superior to n-gram learners on this task. Implications for nativist theories of language acquisition are discussed.

Abstract

Achieving linguistic proficiency requires identifying words from speech, and discovering the constraints that govern the way those words are used. In a recent study of non-adjacent dependency learning, Frost and Monaghan (2016) demonstrated that learners may perform these tasks together, using similar statistical processes - contrary to prior suggestions. However, in their study, non-adjacent dependencies were marked by phonological cues (plosive-continuant-plosive structure), which may have influenced learning. Here, we test the necessity of these cues by comparing learning across three conditions; fixed phonology, which contains these cues, varied phonology, which omits them, and shapes, which uses visual shape sequences to assess the generality of statistical processing for these tasks. Participants segmented the sequences and generalized the structure in both auditory conditions, but learning was best when phonological cues were present. Learning was around chance on both tasks for the visual shapes group, indicating statistical processing may critically differ across domains.

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

Large-scale graph data in real-world applications is often not static but dynamic,
i. e., new nodes and edges appear over time. Current graph convolution approaches
are promising, especially, when all the graph’s nodes and edges are available dur-
ing training. When unseen nodes and edges are inserted after training, it is not
yet evaluated whether up-training or re-training from scratch is preferable. We
construct an experimental setup, in which we insert previously unseen nodes and
edges after training and conduct a limited amount of inference epochs. In this
setup, we compare adapting pretrained graph neural networks against retraining
from scratch. Our results show that pretrained models yield high accuracy scores
on the unseen nodes and that pretraining is preferable over retraining from scratch.
Our experiments represent a first step to evaluate and develop truly online variants
of graph neural networks.

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

This paper presents the Demo / Kemo corpus of Dutch and Korean emotional speech. The corpus has been specifically developed for the purpose of cross-linguistic comparison, and is more balanced than any similar corpus available so far: a) it contains expressions by both Dutch and Korean actors as well as judgments by both Dutch and Korean listeners; b) the same elicitation technique and recording procedure was used for recordings of both languages; c) the same nonsense sentence, which was constructed to be permissible in both languages, was used for recordings of both languages; and d) the emotions present in the corpus are balanced in terms of valence, arousal, and dominance. The corpus contains a comparatively large number of emotions (eight) uttered by a large number of speakers (eight Dutch and eight Korean). The counterbalanced nature of the corpus will enable a stricter investigation of language-specific versus universal aspects of emotional expression than was possible so far. Furthermore, given the carefully controlled phonetic content of the expressions, it allows for analysis of the role of specific phonetic features in emotional expression in Dutch and Korean.

Abstract

Researchers in human language processing and acquisition are making an increasing use of computational models. Computer simulations provide a valuable platform to reproduce hypothesised learning mechanisms that are otherwise very difficult, if not impossible, to verify on human subjects. However, computational models come with problems and risks. It is difficult to (automatically) extract essential information about the developing internal representations from a set of simulation runs, and often researchers limit themselves to analysing learning curves based on empirical recognition accuracy through time. The associated risk is to erroneously deem a specific learning behaviour as generalisable to human learners, while it could also be a mere consequence (artifact) of the implementation of the artificial learner or of the input coding scheme. In this paper a set of simulation runs taken from the ACORNS project is investigated. First a look `inside the box' of the learner is provided by employing novel quantitative methods for analysing changing structures in large data sets. Then, the obtained findings are discussed in the perspective of their ecological validity in the field of child language acquisition.

Abstract

Child songs are a great source for linguistic learning. Here we explore whether children can acquire novel words in a second language by playing a game featuring child songs in a playhouse. We present data from three studies that serve as scientific proof for the functionality of one game of the playhouse: the Energy Center. For this game, three hand-bikes were mounted on a panel. When children start moving the hand-bikes, child songs start playing simultaneously. Once the children produce enough energy with the hand-bikes, the songs are additionally accompanied with the sounds of musical instruments. In our studies, children executed a picture-selection task to evaluate whether they acquired new vocabulary from the songs presented during the game. Two of our studies were run in the field, one at a Dutch and one at an Indian pre-school. The third study features data from a more controlled laboratory setting. Our results partly confirm that the Energy Center is a successful means to support vocabulary acquisition in a second language. More research with larger sample sizes and longer access to the Energy Center is needed to evaluate the overall functionality of the game. Based on informal observations at our test sites, however, we are certain that children do pick up linguistic content from the songs during play, as many of the children repeat words and phrases from songs they heard. We will pick up upon these promising observations during future studies

Abstract

This corpus-based study shows that the presence and duration of schwa in Dutch word-initial syllables are affected by a word’s predictability and its morphological structure. Schwa is less reduced in words that are more predictable given the following word. In addition, schwa may be longer if the syllable forms a prefix, and in prefixes the duration of schwa is positively correlated with the frequency of the word relative to its stem. Our results suggest that the conditions which favor reduced realizations are more complex than one would expect on the basis of the current literature.

Abstract

Non-native (L2) speakers of English often experience difficulties in producing English interdental fricatives (e.g. the voiceless [θ]), and this leads to frequent substitutions of these fricatives (e.g. with [t], [s], and [f]). Differences in the choice of [θ]-substitutions across L2 speakers with different native (L1) language backgrounds have been extensively explored. However, even within one foreign accent, more than one substitution choice occurs, but this has been less systematically studied. Furthermore, little is known about whether the substitutions of voiceless [θ] are phonetically clear instances of [t], [s], and [f], as they are often labelled. In this study, we attempted a phonetic approach to examine language-specific preferences for [θ]-substitutions by carrying out acoustic measurements of L1 and L2 realizations of these sounds. To this end, we collected a corpus of spoken English with L1 speakers (UK-English), and Dutch and German L2 speakers. We show a) that the distribution of differential substitutions using identical materials differs between Dutch and German L2 speakers, b) that [t,s,f]-substitutes differ acoustically from intended [t,s,f], and c) that L2 productions of [θ] are acoustically comparable to L1 productions.

Abstract

Prediction in language has traditionally been studied using
simple designs in which neural responses to expected
and unexpected words are compared in a categorical
fashion. However, these designs have been contested
as being ‘prediction encouraging’, potentially exaggerating
the importance of prediction in language understanding.
A few recent studies have begun to address
these worries by using model-based approaches to probe
the effects of linguistic predictability in naturalistic stimuli
(e.g. continuous narrative). However, these studies
so far only looked at very local forms of prediction, using
models that take no more than the prior two words into
account when computing a word’s predictability. Here,
we extend this approach using a state-of-the-art neural
language model that can take roughly 500 times longer
linguistic contexts into account. Predictability estimates
fromthe neural network offer amuch better fit to EEG data
from subjects listening to naturalistic narrative than simpler
models, and reveal strong surprise responses akin to
the P200 and N400. These results show that predictability
effects in language are not a side-effect of simple designs,
and demonstrate the practical use of recent advances
in AI for the cognitive neuroscience of language.

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

Speakers express word meaning through systematic but non-canonical acoustic variation of tone of voice (ToV), i.e., variation of speaking rate, pitch, vocal effort, or loudness. Words are, for example, pronounced at a higher pitch when referring to small than to big referents. In the present study, we examined whether listeners can use ToV to learn the meaning of novel adjectives (e.g., “blicket”). During training, participants heard sentences such as “Can you find the blicket one?” spoken with ToV representing hot-cold, strong-weak, and big-small. Participants’ eye movements to two simultaneously shown objects with properties representing the relevant two endpoints (e.g., an elephant and an ant for big-small) were monitored. Assignment of novel adjectives to endpoints was counterbalanced across participants. During test, participants heard the sentences spoken with a neutral ToV, while seeing old or novel picture pairs varying along the same dimensions (e.g., a truck and a car for big-small). Participants had to click on the adjective’s referent. As evident from eye movements, participants did not infer the intended meaning during first exposure, but learned the meaning with the help of ToV during training. At test listeners applied this knowledge to old and novel items even in the absence of informative ToV.

Abstract

This study investigates effects of prosodic factors (prominence, boundary) on coarticulatory Vnasalization in Australian English (AusE) in CVN and NVC in comparison to those in American English
(AmE). As in AmE, prominence was found to
lengthen N, but to reduce V-nasalization, enhancing N’s nasality and V’s orality, respectively (paradigmatic contrast enhancement). But the prominence effect in CVN was more robust than that in AmE. Again similar to findings in AmE, boundary
induced a reduction of N-duration and V-nasalization phrase-initially (syntagmatic contrast enhancement), and increased the nasality of both C and V phrasefinally.
But AusE showed some differences in terms
of the magnitude of V nasalization and N duration. The results suggest that the linguistic contrast enhancements underlie prosodic-structure modulation of coarticulatory V-nasalization in
comparable ways across dialects, while the fine phonetic detail indicates that the phonetics-prosody interplay is internalized in the individual dialect’s phonetic grammar.

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

We present a software module, the LAT Bridge, which enables bidirectionalcommunication between the annotation and exploration tools developed at the MaxPlanck Institute for Psycholinguistics as part of our Language ArchivingTechnology (LAT) tool suite. These existing annotation and exploration toolsenable the annotation, enrichment, exploration and archive management oflinguistic resources. The user community has expressed the desire to usedifferent combinations of LAT tools in conjunction with each other. The LATBridge is designed to cater for a number of basic data interaction scenariosbetween the LAT annotation and exploration tools. These interaction scenarios(e.g. bootstrapping a wordlist, searching for annotation examples or lexicalentries) have been identified in collaboration with researchers at ourinstitute.We had to take into account that the LAT tools for annotation and explorationrepresent a heterogeneous application scenario with desktop-installed andweb-based tools. Additionally, the LAT Bridge has to work in situations wherethe Internet is not available or only in an unreliable manner (i.e. with a slowconnection or with frequent interruptions). As a result, the LAT Bridge’sarchitecture supports both online and offline communication between the LATannotation and exploration tools.

Abstract

Many theories hold that semantic variation in the world’s languages can be explained in terms of a universal conceptual space that is partitioned differently by different languages. Recent work has supported this view in the semantic domain of containers (Malt et al., 1999), and assumed it in the domain of spatial relations (Khetarpal et al., 2009), based in both cases on similarity judgments derived from pile-sorting of stimuli. Here, we reanalyze data from these two studies and find a more complex picture than these earlier studies suggested. In both cases we find that sorting is similar across speakers of different languages (in line with the earlier studies), but nonetheless reflects the sorter’s native language (in contrast with the earlier studies). We conclude that there are cross-culturally shared conceptual tendencies that can be revealed by pile-sorting, but that these tendencies may be modulated to some extent by language. We discuss the implications of these findings for accounts of semantic variation.

Abstract

Previous studies on Cantonese Chinese showed that rising
question intonation contours on low-toned words lead to
frequent misperceptions of the tones. Here we explored the
processing consequences of this interaction between tone and
intonation by comparing the processing and identification of
monosyllabic critical words at the end of questions and
statements, using a tone identification task, and ERPs as an
online measure of speech comprehension. Experiment 1
yielded higher error rates for the identification of low tones at
the end of questions and a larger N400-P600 pattern, reflecting
processing difficulty and reanalysis, compared to other
conditions. In Experiment 2, we investigated the effect of
immediate lexical context on the tone by intonation interaction.
Increasing contextual constraints led to a reduction in errors
and the disappearance of the P600 effect. These results
indicate that there is an immediate interaction between tone,
intonation, and context in online speech comprehension. The
difference in performance and activation patterns between the
two experiments highlights the significance of context in
understanding a tone language, like Cantonese-Chinese.

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

Verb learning is important for young children. While most
previous research has focused on linguistic and conceptual
challenges in early verb learning (e.g. Gentner, 1982, 2006),
the present paper examined early verb learning at the
attentional level and quantified the input for early verb learning
by measuring verb-action co-occurrence statistics in parent-
child interaction from the learner’s perspective. To do so, we
used head-mounted eye tracking to record fine-grained
multimodal behaviors during parent-infant joint play, and
analyzed parent speech, parent and infant action, and infant
attention at the moments when parents produced verb labels.
Our results show great variability across different action verbs,
in terms of frequency of verb utterances, frequency of
corresponding actions related to verb meanings, and infants’
attention to verbs and actions, which provide new insights on
why some verbs are harder to learn than others.

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

Continuous Bag of Words (CBOW) is a powerful text embedding method. Due to its strong capabilities to encode word content, CBOW embeddings perform well on a wide range of downstream tasks while being efficient to compute. However, CBOW is not capable of capturing the word order. The reason is that the computation of CBOW's word embeddings is commutative, i.e., embeddings of XYZ and ZYX are the same. In order to address this shortcoming, we propose a
learning algorithm for the Continuous Matrix Space Model, which we call Continual Multiplication of Words (CMOW). Our algorithm is an adaptation of word2vec, so that it can be trained on large quantities of unlabeled text. We empirically show that CMOW better captures linguistic properties, but it is inferior to CBOW in memorizing word content. Motivated by these findings, we propose a hybrid model that combines the strengths of CBOW and CMOW. Our results show that the hybrid CBOW-CMOW-model retains CBOW's strong ability to memorize word content while at the same time substantially improving its ability to encode other linguistic information by 8%. As a result, the hybrid also performs better on 8 out of 11 supervised downstream tasks with an average improvement of 1.2%.

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

Studies have claimed that blind people’s spatial representations are different from sighted people, and blind people display superior auditory processing. Due to the nature of auditory and haptic information, it has been proposed that blind people have spatial representations that are more sequential than sighted people. Even the temporary loss of sight—such as through blindfolding—can affect spatial representations, but not much research has been done on this topic. We compared blindfolded and sighted people’s linguistic spatial expressions and non-linguistic localization accuracy to test how blindfolding affects the representation of path in auditory motion events. We found that blindfolded people were as good as sighted people when localizing simple sounds, but they outperformed sighted people when localizing auditory motion events. Blindfolded people’s path related speech also included more sequential, and less holistic elements. Our results indicate that even temporary loss of sight influences spatial representations of auditory motion events

Abstract

Lombard speech, speech produced in noise, is acoustically different from speech produced in quiet (plain speech) in several ways, including having a higher and wider F0 range (pitch). Extensive research on native Lombard speech does not consider that non-natives experience a higher cognitive load while producing
speech and that the native language may influence the non-native speech. We investigated pitch range in plain and Lombard speech in native and non-natives.
Dutch and American-English speakers read contrastive question-answer pairs in quiet and in noise in English, while the Dutch also read Dutch sentence pairs. We found that Lombard speech is characterized by a wider pitch range than plain speech, for all speakers (native English, non-native English, and native Dutch).
This shows that non-natives also widen their pitch range in Lombard speech. In sentences with early-focus, we see the same increase in pitch range when going from plain to Lombard speech in native and non-native English, but a smaller increase in native Dutch. In sentences with late-focus, we see the biggest increase for the native English, followed by non-native English and then native Dutch. Together these results indicate an effect of the native language on non-native Lombard speech.

Abstract

Lombard speech, speech produced in noise, is
typically produced with a higher fundamental
frequency (F0, pitch) compared to speech in quiet. This paper examined the potential differences in native and non-native Lombard speech by analyzing median pitch in sentences with early- or late-focus produced in quiet and noise. We found an increase in pitch in late-focus sentences in noise for Dutch speakers in both English and Dutch, and for American-English speakers in English. These results
show that non-native speakers produce Lombard speech, despite their higher cognitive load. For the early-focus sentences, we found a difference between the Dutch and the American-English speakers. Whereas the Dutch showed an increased F0 in noise
in English and Dutch, the American-English speakers did not in English. Together, these results suggest that some acoustic characteristics of Lombard speech, such as pitch, may be language-specific, potentially
resulting in the native language influencing the non-native Lombard speech.

Abstract

Grice in pragmatics and Levelt in psycholinguistics have proposed models of human communication where the starting point of communicative action is an individual intention. This assumption, though, has to face serious objections with regard to the alleged existence of explicit representations of the communicative goals to be pursued. Here evidence is surveyed which shows that in fact speaking may ordinarily be a quite automatic activity prompted by contextual cues and driven by behavioural schemata abstracted away from social regularities. On the one hand, this means that there could exist no intentions in the sense of explicit representations of communicative goals, following from deliberate reasoning and triggering the communicative action. On the other hand, however, there are reasons to allow for a weaker notion of intention than this, according to which communication is an intentional affair, after all. Communicative action is said to be intentional in this weaker sense to the extent that it is subject to a double mechanism of control, with respect both to present-directed and future-directed intentions.

Abstract

Humans learn language by interaction with their environment and listening to other humans. It should also be possible for computational models to learn language directly from speech but so far most approaches require text. We improve on existing neural network approaches to create visually grounded embeddings for spoken utterances. Using a combination of a multi-layer GRU, importance sampling, cyclic learning rates, ensembling and vectorial self-attention our results show a remarkable increase in image-caption retrieval performance over previous work. Furthermore, we investigate which layers in the model learn to recognise words in the input. We find that deeper network layers are better at encoding word presence, although the final layer has slightly lower performance. This shows that our visually grounded sentence encoder learns to recognise words from the input even though it is not explicitly trained for word recognition.

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.