Publications

Abstract

Affective behaviors enable social robots to not only establish better connections with humans but also serve as a tool for the robots to express their internal states. It has been well established that emotions are important to signal understanding in Human-Robot Interaction (HRI). This work aims to harness the power of Large Language Models (LLM) and proposes an approach to control the affective behavior of robots. By interpreting emotion appraisal as an Emotion Recognition in Conversation (ERC) tasks, we used GPT-3.5 to predict the emotion of a robot’s turn in real-time, using the dialogue history of the ongoing conversation. The robot signaled the predicted emotion using facial expressions. The model was evaluated in a within-subjects user study (N = 47) where the model-driven emotion generation was compared against conditions where the robot did not display any emotions and where it displayed incongruent emotions. The participants interacted with the robot by playing a card sorting game that was specifically designed to evoke emotions. The results indicated that the emotions were reliably generated by the LLM and the participants were able to perceive the robot’s emotions. It was found that the robot expressing congruent model-driven facial emotion expressions were perceived to be significantly more human-like, emotionally appropriate, and elicit a more positive impression. Participants also scored significantly better in the card sorting game when the robot displayed congruent facial expressions. From a technical perspective, the study shows that LLMs can be used to control the affective behavior of robots reliably in real-time. Additionally, our results could be used in devising novel human-robot interactions, making robots more effective in roles where emotional interaction is important, such as therapy, companionship, or customer service.

Abstract

The current study explored the role of sentential inference in connecting lexical/grammatical knowledge and overall text comprehension in foreign language learning. Using structural equation modeling (SEM), causal relationships were examined between four latent variables: lexical knowledge, grammatical knowledge, sentential inference, and text comprehension. The study analyzed 281 Chinese university students learning Japanese as a second language and compared two causal models: (1) the partially-mediated model, which suggests that lexical knowledge, grammatical knowledge, and sentential inference concurrently influence text comprehension, and (2) the wholly-mediated model, which posits that both lexical and grammatical knowledge impact sentential inference, which then further affects text comprehension. The SEM comparison analysis supported the wholly-mediated model, showing sequential causal relationships from lexical knowledge to sentential inference and then to text comprehension, without significant contribution from grammatical knowledge. The results indicate that sentential inference serves as a crucial bridge between lexical knowledge and text comprehension.

Abstract

This study is the first to report chronometric evidence on Tongan language production. It has been speculated that the mora plays an important role during Tongan phonological encoding. A mora follows the (C)V form, so /a/ and /ka/ (but not /k/) denote a mora in Tongan. Using a picture-word naming paradigm, Tongan native speakers named pictures containing superimposed non-word distractors. This task has been used before in Japanese, Korean, and Vietnamese to investigate the initially selected unit during phonological encoding (IPU). Compared to control distractors, both onset and mora overlapping distractors resulted in faster naming latencies. Several alternative explanations for the pattern of results - proficiency in English, knowledge of Latin script, and downstream effects - are discussed. However, we conclude that Tongan phonological encoding likely natively uses the phoneme, and not the mora, as the IPU..

Abstract

Does the way a word is written influence its spoken production? Previous studies suggest that orthography is involved only when the orthographic representation is highly relevant during speaking (e.g., in reading-aloud tasks). To address this issue, we carried out two experiments using the blocked cyclic picture-naming paradigm. In both experiments, participants were asked to name pictures repeatedly in orthographically homogeneous or heterogeneous blocks. In the naming task, the written form was not shown; however, the radical of the first character overlapped between the four pictures in this block type. A facilitative orthographic effect was found when picture names shared part of their written forms, compared with the heterogeneous condition. This facilitative effect was independent of the position of orthographic overlap (i.e., the left, the lower, or the outer part of the character). These findings strongly suggest that orthography can influence speaking even when it is not highly relevant (i.e., during picture naming) and the orthographic effect is less likely to be attributed to strategic preparation.

Abstract

We argue, based on a study of brain responses to speech sound differences in Japanese, that memory encoding of functional speech sounds-phonemes-are highly abstract. As an example, we provide evidence for a theory where the consonants/p t k b d g/ are not only made up of symbolic features but are underspecified with respect to voicing or laryngeal features, and that languages differ with respect to which feature value is underspecified. In a previous study we showed that voiced stops are underspecified in English [Hestvik, A., & Durvasula, K. (2016). Neurobiological evidence for voicing underspecification in English. Brain and Language], as shown by asymmetries in Mismatch Negativity responses to /t/ and /d/. In the current study, we test the prediction that the opposite asymmetry should be observed in Japanese, if voiceless stops are underspecified in that language. Our results confirm this prediction. This matches a linguistic architecture where phonemes are highly abstract and do not encode actual physical characteristics of the corresponding speech sounds, but rather different subsets of abstract distinctive features.

Abstract

Objectives: A new computer-aided screening system for osteoporosis using panoramic radiographs was developed. The conventional system could detect porotic changes within the lower border of the mandible, but its severity could not be evaluated. Our aim was to enable the system to measure severity by implementing a linear bone resorption severity index (BRSI) based on the cortical bone shape.
Methods: The participants were 68 females (>50 years) who underwent panoramic radiography and lumbar spine bone density measurements. The new system was designed to extract the lower border of the mandible as region of interests and convert them into morphological skeleton line images. The total perimeter length of the skeleton lines was defined as the BRSI. 40 images were visually evaluated for the presence of cortical bone porosity. The correlation between visual evaluation and BRSI of the participants, and the optimal threshold value of BRSI for new system were investigated through a receiver operator characteristic analysis. The diagnostic performance of the new system was evaluated by comparing the results from new system and lumbar bone density tests using 28 participants.
Results: BRSI and lumbar bone density showed a strong negative correlation (p < 0.01). BRSI showed a strong correlation with visual evaluation. The new system showed high diagnostic efficacy with sensitivity of 90.9%, specificity of 64.7%, and accuracy of 75.0%.
Conclusions: The new screening system is able to quantitatively evaluate mandibular cortical porosity. This allows for preventive screening for osteoporosis thereby enhancing clinical prospects.

Abstract

An ongoing debate in the speech production literature suggests that the initial building block to build up speech sounds differs between languages. That is, Germanic languages are suggested to use the phoneme, but Japanese and Chinese are proposed to use the mora or syllable, respectively. Several studies investigated this matter from a chronometric perspective (i.e., RTs and accuracy). However, a less attention has been paid to the actual acoustic utterances. The current study investigated the verbal responses of two Japanese-English bilingual groups of different proficiency levels (i.e., high and low) when naming English words and found that the presence or absence of vowel epenthesis depended on proficiency. The results indicate that: (1) English word pronunciation by low-proficient Japanese English bilinguals is likely based on their L1 (Japanese) building block and (2) that future studies would benefit from analyzing the acoustic data as well when making inferences from chronometric data.

Abstract

Previous studies suggest that bilinguals' lexical access is language non-selective, especially for orthographically identical translation equivalents across languages (i.e., identical cognates). The present study investigated how such words (e.g., meaning "school" in both Chinese and Japanese) are processed in the (late) Chinese - Japanese bilingual brain. Using an L2-Japanese lexical decision task, both behavioral and electrophysiological data were collected. Reaction times (RTs), as well as the N400 component, showed that cognates are more easily recognized than non-cognates. Additionally, an early component (i.e., the N250), potentially reflecting activation at the word-form level, was also found. Cognates elicited a more positive N250 than non-cognates in the frontal region, indicating that the cognate facilitation effect occurred at an early stage of word formation for languages with logographic scripts.

Abstract

In a masked priming word-naming task, a facilitation due to the initial-segmental sound overlap for 2-character kanji prime-target pairs was affected by certain orthographic properties (Yoshihara, Nakayama, Verdonschot, & Hino, 2017). That is, the facilitation that was due to the initial mora overlap occurred only when the mora was the whole pronunciation of their initial kanji characters (i.e., match pairs; e.g., /ka-se.ki/-/ka-rjo.ku/). When the shared initial mora was only a part of the kanji characters' readings, however, there was no facilitation (i.e., mismatch pairs; e.g., /ha.tu-a.N/-/ha.ku-bu.tu/). In the present study, we used a masked priming picture-naming task to investigate whether the previous results were relevant only when the orthography of targets is visually presented. In Experiment 1. the main findings of our word-naming task were fully replicated in a picture-naming task. In Experiments 2 and 3. the absence of facilitation for the mismatch pairs were confirmed with a new set of stimuli. On the other hand, a significant facilitation was observed for the match pairs that shared the 2 initial morae (in Experiment 4), which was again consistent with the results of our word-naming study. These results suggest that the orthographic properties constrain the phonological expression of masked priming for kanji words across 2 tasks that are likely to differ in how phonology is retrieved. Specifically, we propose that orthography of a word is activated online and constrains the phonological encoding processes in these tasks.

Abstract

This chapter describes how speakers access words from the mental lexicon. Lexical access is a crucial
component in the process of transforming thoughts into speech. Some theories consider lexical access to be
strictly serial and discrete, while others view this process as being cascading or even interactive, i.e. the different
sub-levels influence each other. We discuss some of the evidence in favour and against these viewpoints, and
also present arguments regarding the ongoing debate on how words are selected for production. Another important
issue concerns the access to morphologically complex words such as derived and inflected words, as well as
compounds. Are these accessed as whole entities from the mental lexicon or are the parts assembled online? This
chapter tries to provide an answer to that question as well.

Abstract

Recent research has put forward the idea that Chinese speech production is governed by the syllable as the fundamental phonological unit. However, it may be that onset priming might be more difficult to obtain in Mandarin Chinese. Therefore, in this study, the degree of overlap between prime and target was increased from C to CV (i.e., extending beyond the phoneme) as well as whether primes and targets had an overlapping structure (CV vs. CVN). Subsyllabic priming effects were found (i.e., onset + vowel overlap but not purely onset overlap), contrasting with the claim that the syllable is the compulsory building block in the initial construction of Mandarin Chinese phonology.

Abstract

The Microsoft KINECT is a 3D sensing device originally developed for the XBOX. The Microsoft KINECT opens up many exciting new opportunities for conducting experimental research on human behavior. We investigated some of these possibilities within the field of psycholinguistics (specifically: language production) by creating software, using C#, allowing for the KINECT to be used in a typical psycholinguistic experimental setting. The results of a naming experiment using this software confirmed that the KINECT was able to measure the effects of a robust psycholinguistic variable (word frequency) on naming latencies. However, although the current version of the software is able to measure psycholinguistic variables of interest, we also discuss several points where the software can still stand to be improved. The main aim of this paper is to make the software freely available for assessment and use by the psycholinguistic community and to illustrate the KINECT as a potentially valuable tool for investigating human behavior, especially in the field of psycholinguistics.