Recent advancements in NLP, largely driven by Large Language Models (LLMs), have significantly improved performance on an array of tasks.However, Dialogue Act (DA) classification remains challenging, particularly in the fine-grained 50-class, multiparty setting.This paper investigates the root causes of LLMs' poor performance in DA classification through a linguistically motivated analysis.We identify three key pre-tasks essential for accurate DA prediction: Turn Management, Communicative Function Identification, and Dialogue Structure Prediction.Our experiments reveal that LLMs struggle with these fundamental tasks, often failing to outperform simple rulebased baselines.Additionally, we establish a strong empirical correlation between errors in these pre-tasks and DA classification failures.A human study further highlights the significant gap between LLM and human-level dialogue understanding.These findings indicate that LLMs' shortcomings in dialogue comprehension hinder their ability to accurately predict DAs, highlighting the need for improved dialogue-aware training approaches. InstructionYou are an intelligent annotator capable of classifying the intention behind each speaker's utterance.You will be provided with a list of possible Dialogue Acts and their definitions.You will be given an utterance surrounded by '#'.Your task is to predict the correct label for that utterance.You will also be given a snapshot of the conversation to provide context for your prediction.Return the answer in the format: 'label:predicted label'.The Dialogue Acts and their definitions are as follows: Statement: General statements.Accept: A short utterance indicating acceptance of a previous speaker's statement.Disruption: Indecipherable or disrupted speech.Defending/Explanation: The speaker defends their opinion or provides an explanation.

Yuanyuan Lei, Md Messal Monem Miah, Ayesha Qamar, Sai Ramana Reddy, Jonathan Tong, Haotian Xu, Ruihong Huang. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Stereoscopic AR and VR headsets have displays and lenses that are either fixed or adjustable to match a limited range of user inter-pupillary distances (IPDs). Projective geometry predicts a misperception of depth when either the displays or virtual cameras used to render images are misaligned with the eyes. However, misalignment between the eyes and lenses might also affect binocular convergence, which could further distort perceived depth. This possibility has been largely ignored in previous studies. Here, we evaluated this phenomenon in a VR headset in which the inter-lens and inter-axial camera separations are coupled and adjustable. In a baseline condition, both were matched to observers' IPDs. In two other conditions, the inter-lens and inter-axial camera separations were set to the maximum and minimum allowed by the headset. In each condition, observers were instructed to adjust a fold created by two intersecting, textured surfaces until it appeared to have an angle of 90°. The task was performed at three randomly interleaved viewing distances, monocularly and binocularly. In the monocular condition, observers underestimated the fold angle and there was no effect of viewing distance on their settings. In the binocular conditions, we found that when the lens and camera separation were less than the viewer's IPD, they exhibited compression of perceived slant relative to baseline. The reverse pattern was seen when the lens and camera separation were larger than the viewer's IPD. These results were well explained by a geometric model that considers shifts in convergence due to lens and display misalignment with the eyes, as well as the relative contribution of monocular cues.

The magnifying optics of virtual reality (VR) head-mounted displays (HMD) often cause undesirable pincushion distortion in the displayed imagery. Eccentrically increasing magnification radially displaces image-points away from the optical axis, causing straight lines to curve outwards. This, in turn, should affect the 3D perception of surface shape by warping binocular and monocular depth cues. Previous research has shown that distortion-induced biases in perceived slant do occur in static images. However, most use cases in VR involve moving images. Here we evaluate the impact of motion on biases in perceived slant. An HMD was used to present flat, textured surfaces that varied in slant and were either stationary, or translated laterally by the observer. In separate studies we varied the degree of distortion and evaluated the impact on perceived slant at several locations along the surface. We found that, irrespective of whether the surface was moving or stationary, distortion introduced significant bias into local slant estimates. The pattern of results is consistent with the surface appearing to be concave (as if viewing the inside surface of a bowl), as predicted from the warping of binocular and monocular cues. Importantly, the intermediate distortion level produced the same, but weaker, pattern of biases seen in the fully-distorted condition. When an appropriate level of pre-warping was applied, slant perception was veridical. Overall, our results highlight the importance of sufficiently correcting for optical distortions in VR HMDs to enable veridical perception of surface attitude.

This paper describes validity testing of the NeuLog NUL-217 GSR measurement device. This was accomplished by comparing the NeuLog device to readings from the Biopac Student Lab Systems EDA system. The results of this research found that measurements from the NeuLog device are correlated with and comparable to the Biopac system. The absolute skin conductance levels typically differed between the two systems. For most psychological and human factors research the dynamic skin conductance responses are more important than absolute levels. The timing and relative magnitude of changes typically tracked well across the two systems indicating that the Neulog system is suitable for the purposes of its intended use in psychological and technological research.

Bipedal walking is generally considered to be the most natural and common locomotion technique in the physical world, for humans, and the most presence-enhancing form of locomotion in virtual reality (VR). However, there are significant differences in the way people walk in VR compared to their walking behaviour in the real world. For instance, previous studies have shown a significant decrease of gait parameters, in particular, velocity and step length in the virtual environment (VE). However, those studies have only considered short periods of walking. In contrast, many VR applications involve extended exposures to the VE and often include additional cognitive tasks such as way-finding. Hence, it remains an open question whether velocity during VR walking will further slowdown over time or if users of VR will eventually speed-up and adapt their velocity to the VE and move with the same speed as in the real world.

Cyberbullying can have a deep and long lasting impact on its victims, who are often adolescents. Accurately detecting cyberbullying helps prevent it. However, the noise and errors in social media posts and messages make detecting cyberbullying very challenging. In this paper, we propose a novel pronunciation based convolutional neural network (PCNN) to address this challenge. Upon observing that the pronunciation of misspelled words in informal online conversations is often unchanged, we used the phoneme codes of the text as the features for a convolutional neural network. This procedure corrects spelling errors that did not alter the pronunciation, thereby alleviating the problem of noise and bullying data sparsity. To overcome class imbalance, a common problem in cyberbullying datasets, we implement three techniques that include threshold-moving, cost function adjusting, and a hybrid solution in our model. We evaluate the performance of our models using two cyberbullying datasets collected from Twitter and Formspring.me. The results of our experiment show that PCNN can achieve improved recall and precision compared to baseline convolutional neural networks.