This presentation focuses on the inefficient text input dilemma and digital communication barriers faced by people with cerebral palsy in human-computer collaboration, deeply analyzes their hand-eye motor coordination strategies during text input, and demonstrates the AR intelligent interactive keyboard independently developed by the team based on this research, aiming to achieve immersive and efficient input through visual reconstruction. It also prospectively discusses virtual keyboard interaction strategies and accessible text input design paradigms for cerebral palsy patients.

The speech recognition accuracy of dysarthric patients is far lower than that of normal populations. This study explores whether multimodal LLMs can improve this issue—by fusing acoustic signals with the contextual understanding capabilities of large language models, and fully utilizing cross-modal complementary information to compensate for the acoustic deficiencies of dysarthric speech.

Automatically converting academic papers into presentation videos requires synchronously generating slides, speech, subtitles, cursor animations, and presenter avatars across multiple modalities while ensuring semantic consistency among them. Existing methods treat each modality as an independent generation task, using the original paper text as the respective input source for separate inference, leading to cross-modal semantic drift, and any local modification requires re-running the entire pipeline.

This presentation introduces PaperContract, a structured contract-driven academic video generation framework. Its core idea is to introduce a persistent structured contract between the paper and downstream modalities—a unified specification for the narrative text, visible content, and visual assets of each slide. All modules use this contract as the sole information source, fundamentally eliminating redundant semantic inference. During generation, slides are produced through a QA-guided verification-refinement loop; speech is synthesized only once per slide, and word-level timestamps are obtained through forced alignment to build a shared temporal contract that synchronously drives subtitles, cursor positioning, and Talking Head rendering, ensuring cross-modal synchronization by design rather than post-hoc alignment. During interactive modification, users issue modification instructions in natural language, and the system only updates the affected units in the contract and selectively invalidates corresponding caches, while unaffected slides, audio, and animations are directly reused, significantly reducing modification costs.

Dysarthria is a motor speech disorder caused by neurological damage. Due to the acoustic differences from standard speech and the scarcity of data, directly using general speech recognition models often leads to performance degradation. Combining existing research, we mainly investigate model architecture adaptation, acoustic feature analysis, curriculum learning training strategies, and other directions to improve dysarthric speech recognition performance.

With the deepening of information accessibility construction in China, captions have become essential infrastructure for hearing-impaired groups and elderly people to access audio and video information. However, in real communication scenarios with high immediacy and information density, traditional speech recognition technology often faces challenges such as long social latency, easy loss of core information, and lack of readability in layout. This paper introduces the "Information Accessibility Structured Caption Generation Evaluation for Hearing-Impaired Groups" task at CCL 2026 (CCL26-Eval). This evaluation focuses on the complete technical pipeline from audio and video input to highly usable structured caption generation. To approximate real-world applications, the task establishes a "PC track" for exploring performance limits and a "mobile track" under constrained resources, and subdivides into two subtasks: basic caption generation and structured readable caption generation. To address the high complexity of spoken expression in real corpora and noise in core information annotation, this evaluation constructs a hybrid evaluation system of "automatic metrics with manual verification and arbitration." This evaluation aims to promote AI caption technology that comprehensively considers transcription accuracy, response speed, and information structured presentation, providing a unified benchmark and high-quality multi-scenario datasets for model research and development and edge deployment in the information accessibility field.

As international technological competition increasingly focuses on artificial intelligence, China's independently developed large model DeepSeek has attracted international media attention, providing a key case for examining the construction and reconstruction of China's science and technology image within the existing international discourse landscape. Using critical metaphor analysis as the theoretical framework, this study builds a self-constructed DeepSeek-themed English and Chinese news corpus (January 2025 to May 2026, containing approximately 150 reports each from People's Daily and The New York Times), and employs the MIPVU metaphor identification procedure, corpus quantitative statistics, and critical discourse qualitative analysis to compare how Chinese and American media construct China's science and technology image through conceptual metaphors. Different metaphor choices influence audience cognition and judgment of China's technological development through mapping differences and framing effects, thereby participating in the shaping of national science and technology image. The study aims to answer three core questions: What core conceptual metaphors are used by both sides? How do these metaphors frame China's technological development and reveal what value orientations? What kind of technological discourse power and geopolitical games do the systematic differences in metaphor choices reflect? It is expected to clarify the metaphor distribution maps of both countries, reveal the cognitive models and value differences behind metaphor differences, and provide specific discourse strategies for constructing Chinese science and technology narratives with cultural autonomy and international persuasiveness. This study is both an empirical deepening of critical metaphor analysis in the field of artificial intelligence discourse and provides theoretical support and practical insights for understanding and optimizing the global communication mechanism of China's science and technology image.

Dysarthric speech recognition is an important issue in accessible human-computer interaction and intelligent rehabilitation assistance. Due to unclear pronunciation, decreased speech intelligibility, and significant individual differences, existing automatic speech recognition models still have high error rates in pathological speech scenarios. For fixed-vocabulary dysarthric speech recognition tasks, this study uses self-supervised speech models as the foundation, constructs a CTC and word classification dual-head learning framework, and introduces a clarity-aware acoustic-edit distance vocabulary rescoring method to enhance the model's ability to correct approximate words and substitution errors. Experiments are conducted on the UASpeech dysarthric speech dataset, and evaluation is performed from the perspectives of overall recognition rate, speaker level, and different clarity levels. Results show that combining word-level supervision and clarity-aware decoding strategies helps improve the robustness of pathological speech recognition. This research can provide technical support for accessible communication, intelligent voice interaction, and rehabilitation assessment for people with disabilities.