Publications

Abstract

Tokenization significantly influences language models (LMs)’ performance. This paper traces the evolution of tokenizers from word-level to subword-level, analyzing how they balance tokens and types to enhance model adaptability while controlling complexity. Despite subword tokenizers like Byte Pair Encoding (BPE) overcoming many word tokenizer limitations, they encounter difficulties in handling non-Latin languages and depend heavily on extensive training data and computational resources to grasp the nuances of multiword expressions (MWEs). This article argues that tokenizers, more than mere technical tools, should drawing inspiration from the cognitive science about human language processing. This study then introduces the “Principle of Least Effort” from cognitive science, that humans naturally seek to reduce cognitive effort, and discusses the benefits of this principle for tokenizer development. Based on this principle, the paper proposes that the Less-is-Better (LiB) model could be a new approach for LLM tokenizer. The LiB model can autonomously learn an integrated vocabulary consisting of subwords, words, and MWEs, which effectively reduces both the numbers of tokens and types. Comparative evaluations show that the LiB tokenizer outperforms existing word and BPE tokenizers, presenting an innovative method for tokenizer development, and hinting at the possibility of future cognitive science-based tokenizers being more efficient.

Abstract

Open Science is becoming a mainstream scientific ideology in psychology and related fields. However, researchers, especially early-career researchers (ECRs) in developing countries, are facing significant hurdles in engaging in Open Science and moving it forward. In China, various societal and cultural factors discourage ECRs from participating in Open Science, such as the lack of dedicated communication channels and the norm of modesty. To make the voice of Open Science heard by Chinese-speaking ECRs and scholars at large, the Chinese Open Science Network (COSN) was initiated in 2016. With its core values being grassroots-oriented, diversity, and inclusivity, COSN has grown from a small Open Science interest group to a recognized network both in the Chinese-speaking research community and the international Open Science community. So far, COSN has organized three in-person workshops, 12 tutorials, 48 talks, and 55 journal club sessions and translated 15 Open Science-related articles and blogs from English to Chinese. Currently, the main social media account of COSN (i.e., the WeChat Official Account) has more than 23,000 subscribers, and more than 1,000 researchers/students actively participate in the discussions on Open Science. In this article, we share our experience in building such a network to encourage ECRs in developing countries to start their own Open Science initiatives and engage in the global Open Science movement. We foresee great collaborative efforts of COSN together with all other local and international networks to further accelerate the Open Science movement.

Abstract

Ancient Chinese poetry is constituted by structured language that deviates from ordinary language usage [1, 2]; its poetic genres impose unique combinatory constraints on linguistic elements [3]. How does the constrained poetic structure facilitate speech segmentation when common linguistic [4, 5, 6, 7, 8] and statistical cues [5, 9] are unreliable to listeners in poems? We generated artificial Jueju, which arguably has the most constrained structure in ancient Chinese poetry, and presented each poem twice as an isochronous sequence of syllables to native Mandarin speakers while conducting magnetoencephalography (MEG) recording. We found that listeners deployed their prior knowledge of Jueju to build the line structure and to establish the conceptual flow of Jueju. Unprecedentedly, we found a phase precession phenomenon indicating predictive processes of speech segmentation—the neural phase advanced faster after listeners acquired knowledge of incoming speech. The statistical co-occurrence of monosyllabic words in Jueju negatively correlated with speech segmentation, which provides an alternative perspective on how statistical cues facilitate speech segmentation. Our findings suggest that constrained poetic structures serve as a temporal map for listeners to group speech contents and to predict incoming speech signals. Listeners can parse speech streams by using not only grammatical and statistical cues but also their prior knowledge of the form of language.

Abstract

Language users process utterances by segmenting them into many cognitive units, which vary in their sizes and linguistic levels. Although we can do such unitization/segmentation easily, its cognitive mechanism is still not clear. This paper proposes an unsupervised model, Less-is-Better (LiB), to simulate the human cognitive process with respect to language unitization/segmentation. LiB follows the principle of least effort and aims to build a lexicon which minimizes the number of unit tokens (alleviating the effort of analysis) and number of unit types (alleviating the effort of storage) at the same time on any given corpus. LiB’s workflow is inspired by empirical cognitive phenomena. The design makes the mechanism of LiB cognitively plausible and the computational requirement light-weight. The lexicon generated by LiB performs the best among different types of lexicons (e.g. ground-truth words) both from an information-theoretical view and a cognitive view, which suggests that the LiB lexicon may be a plausible proxy of the mental lexicon.

Abstract

Chunking in language comprehension is a process that segments continuous linguistic input into smaller chunks that are in the reader’s mental lexicon. Effective chunking during reading facilitates disambiguation and enhances efficiency for comprehension. However, the chunking mechanisms remain elusive, especially in reading given that information arrives simultaneously yet the written systems may not have explicit cues for labeling boundaries such as Chinese. What are the mechanisms of chunking that mediates the reading of the text that contains hierarchical information? We investigated this question by manipulating the lexical status of the chunks at distinct levels in four-character Chinese strings, including the two-character local chunk and four-character global chunk. Male and female human participants were asked to make lexical decisions on these strings in a behavioral experiment, followed by a passive reading task when their electroencephalography (EEG) was recorded. The behavioral results showed that the lexical decision time of lexicalized two-character local chunks was influenced by the lexical status of the four-character global chunk, but not vice versa, which indicated the processing of global chunks possessed priority over the local chunks. The EEG results revealed that familiar lexical chunks were detected simultaneously at both levels and further processed in a different temporal order – the onset of lexical access for the global chunks was earlier than that of local chunks. These consistent results suggest a two-stage operation for chunking in reading–– the simultaneous detection of familiar lexical chunks at multiple levels around 100 ms followed by recognition of chunks with global precedence.