当前位置: X-MOL 学术Lang. Learn. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
To Bilingualism and Beyond! Modeling Bilingualism Requires Looking Beyond Language: A Commentary on “Computational Modeling of Bilingual Language Learning: Current Models and Future Directions”
Language Learning ( IF 3.5 ) Pub Date : 2022-11-09 , DOI: 10.1111/lang.12530
Viorica Marian 1
Affiliation  

Language is a symbolic system, and, like other symbolic systems (computer languages, math), it lends itself well to expansion, both within and across systems. Human minds can accommodate multiple symbolic systems simultaneously: They can understand a natural language, perform arithmetic, read musical notes, and perform a variety of tasks in which symbols are used (Marian, 2023).

In that sense, the multilingual language system does not possess a unique cognitive architecture that is categorically different, but rather multilingualism is the prototypical state of the human mind. When computational models of multilingualism go beyond existing monolingual language models, they frequently do so by focusing on how to differentiate the different languages, whether to represent them in integrated or separate ways, and how to control their use during input and output. There are multiple solutions for addressing each of these problems, but framing the questions around these issues may be missing the bigger picture, one in which language(s) cannot be separated from other mental phenomena.

A word's lexical form is tightly connected to its semantic representational features as well as to perceptual information, to memory, to affect, to other nonmodular mental states. Form overlap across words guides not only recognition within the language system, but also the visual perception of the surrounding scene (Marian & Spivey, 2003), cognitive control (Blumenfeld & Marian, 2011), attention processes more generally (Chabal & Marian, 2015), and even long-term memory (Marian et al., 2021). These rich connections and interactions among cognitive functions are also found in neuroscience. The brain is not modular; instead, a broad whole-brain network is involved in processing symbolic systems, including the two languages of a bilingual, a network that emerges and continuously organizes itself with every new incoming piece of information.

Reviews of computational models of language in bilinguals are in agreement that language learning is a dynamic, interactive, developmental process whose study requires an interdisciplinary approach. Li and Xu's most recent review in this vein considers a range of models, covering Bayesian modeling, multimodal learning, and network science modeling, and can serve as an introductory primer for students and novices in the field who are not familiar with individual models or previous reviews of this area. It reflects the field's overall focus on prioritizing language learning, while also discussing models of visual word recognition, like BIA, BIA+, and Multilink, the development of which shaped the field. Computational models of bilingual spoken language processing, like the Bilingual Language Interaction Network for Comprehension of Speech (Shook & Marian, 2013), are rarer, but just as necessary.

Existing models, however, focus on distinct individual aspects of language, such as learning, visual processing, auditory comprehension, or translation, rather than on a broad integrated framework that can accommodate the full spectrum of tasks managed by the bilingual cognitive network. The limitations of such separate models underscore the need for computational accounts of bilingualism to shift from discrete models focused on separate tasks to larger integrated models that more accurately reflect human language.

Those who model bilingualism agree that interdisciplinary approaches building on knowledge from computational neuroscience, natural language processing, and first language acquisition are needed to move the field forward. However, formulating precisely how that should happen is more difficult.

One thing that is clear, however, is that the computer metaphor the authors alluded to is no longer appropriate. To wit, the computer is a poor metaphor for the human mind. It dates to the 1950s and was relatively popular in the decades that followed, but it no longer reflects modern understanding of neuroscience, computer science, and cognitive science more broadly. The sooner researchers abandon the mind-as-a-computer lens, the sooner it will be possible to advance computational modeling of the mind, including the bilingual mind, beyond the constraints of a computer metaphor.

To understand bilingualism and to model it successfully, modeling efforts need to move away from considering language in isolation and to instead integrate it into a broader cognitive framework. Bilingualism is not just about language, although that is its most immediately salient component and the one those who study bilingualism tend to focus on. Bilingualism also shapes perception, memory, learning, emotion, decision making, and other functions. Successful modeling of bilingualism requires the recognition of a broader cognitive network in constant flux, where symbolic systems impact the entire cognitive architecture of the human mind.

Much about modeling language learning in bilinguals is not specific to bilingualism nor even to language learning, but applies more broadly to learning in general. What exactly is specific to language, and what is specific to bilingualism or multilingualism, if anything, is an open question. Asking and answering these broader questions is likely to hold the key to the next generation of computational models of bilingualism.

The field of bilingualism stands on the precipice of the next big shift in computational modeling of multiple languages, but exactly how to cross it is less clear. As modelers look around trying to figure out how to “solve” the problem of modeling bilingualism, the solution will need to extend beyond two languages, to cognitive functions more broadly.

The task may seem overwhelming; accomplishing it will be no small feat indeed, but at this point the field is more ready for it than ever before. Modern computing capabilities are increasingly able to accommodate massive amounts of data and to handle ambitious modeling efforts. The last few decades have seen smaller individual problems being worked out within the modeling of perception, comprehension, reading, learning, memory, and other domains. Although putting the different pieces of the puzzle together is a daunting task, the zeitgeist is right for it. Even if the full picture does not reveal itself at once, and we can only start by combining a few pieces at a time in an incremental manner (like perception and language, or language and memory—note that none of these are distinct categories!), a fuller picture will begin to emerge over time.

Ultimately, modeling bilingualism means moving beyond modeling the learning of two languages to modeling the brain's capacity for multiple symbolic systems, a defining feature of the human mind.



中文翻译:


双语及超越!双语建模需要超越语言:对“双语学习的计算模型:当前模型和未来方向”的评论



语言是一种符号系统,并且与其他符号系统(计算机语言、数学)一样,它非常适合在系统内部和跨系统进行扩展。人类思维可以同时容纳多个符号系统:他们可以理解自然语言、进行算术、阅读音符以及执行各种使用符号的任务(Marian, 2023 )。


从这个意义上说,多语言语言系统并不具有绝对不同的独特认知架构,而是多语言是人类思维的原型状态。当多语言的计算模型超越现有的单语语言模型时,它们经常关注如何区分不同的语言,是否以集成或单独的方式表示它们,以及如何在输入和输出过程中控制它们的使用。解决这些问题都有多种解决方案,但围绕这些问题提出问题可能会忽略更大的图景,在这个图景中,语言不能与其他心理现象分开。


单词的词汇形式与其语义表征特征以及感知信息、记忆、影响和其他非模块心理状态紧密相关。单词之间的形式重叠不仅指导语言系统内的识别,还指导周围场景的视觉感知(Marian & Spivey, 2003 )、认知控制(Blumenfeld & Marian, 2011 )、更普遍的注意力过程(Chabal & Marian, 2015 ) ),甚至是长期记忆(Marian et al., 2021 )。认知功能之间丰富的联系和相互作用也存在于神经科学中。大脑不是模块化的;相反,一个广泛的全脑网络参与处理符号系统,包括双语者的两种语言,这个网络会随着每条新传入的信息而出现并不断自我组织。


对双语者语言计算模型的回顾一致认为,语言学习是一个动态的、交互式的、发展的过程,其研究需要跨学科的方法。 Li 和 Xu 在这方面的最新评论考虑了一系列模型,涵盖贝叶斯建模、多模态学习和网络科学建模,并且可以作为不熟悉单个模型或以前的模型的学生和该领域新手的入门读本。对该领域的评论。它反映了该领域对优先考虑语言学习的总体关注,同时还讨论了视觉单词识别模型,例如 BIA、BIA+ 和 Multilink,这些模型的发展塑造了该领域。双语口语处理的计算模型,如用于理解语音的双语交互网络(Shook & Marian, 2013 ),虽然比较少见,但同样必要。


然而,现有模型侧重于语言的不同个体方面,例如学习、视觉处理、听觉理解或翻译,而不是广泛的集成框架,该框架可以容纳双语认知网络管理的全部任务。这种单独模型的局限性强调了双语计算帐户需要从专注于单独任务的离散模型转向更准确地反映人类语言的更大集成模型。


那些为双语建模的人一致认为,需要建立在计算神经科学、自然语言处理和第一语言习得知识基础上的跨学科方法来推动该领域的发展。然而,准确地制定如何实现这一目标则更加困难。


然而,有一点是明确的:作者提到的计算机比喻不再合适。也就是说,计算机是人类思维的一个糟糕的比喻。它可以追溯到 20 世纪 50 年代,并在随后的几十年中相对流行,但它不再更广泛地反映现代对神经科学、计算机科学和认知科学的理解。研究人员越早放弃将思维视为计算机的视角,就越有可能超越计算机隐喻的限制,推进思维(包括双语思维)的计算模型。


为了理解双语并成功建模,建模工作需要摆脱孤立地考虑语言,而是将其整合到更广泛的认知框架中。双语不仅仅是语言,尽管这是其最直接的显着组成部分,也是研究双语的人往往关注的部分。双语还塑造感知、记忆、学习、情感、决策和其他功能。双语的成功建模需要认识到更广泛的不断变化的认知网络,其中符号系统影响人类思维的整个认知结构。


关于双语者的语言学习建模的许多内容并不特定于双语,甚至也不特定于语言学习,而是更广泛地适用于一般学习。语言到底有什么特殊性,双语或多语言有什么特殊性(如果有的话)是一个悬而未决的问题。提出并回答这些更广泛的问题可能是下一代双语计算模型的关键。


双语领域正处于多种语言计算建模下一次重大转变的边缘,但具体如何跨越它还不太清楚。当建模者环顾四周,试图找出如何“解决”双语建模问题时,解决方案将需要扩展到两种语言之外,扩展到更广泛的认知功能。


这项任务可能看起来很艰巨;实现这一目标确实是一项不小的壮举,但目前该领域比以往任何时候都已经做好了准备。现代计算能力越来越能够容纳大量数据并处理雄心勃勃的建模工作。在过去的几十年里,较小的个人问题在感知、理解、阅读、学习、记忆和其他领域的建模中得到解决。尽管将拼图的不同部分拼凑在一起是一项艰巨的任务,但时代精神是正确的。即使完整的画面不会立即显现出来,我们也只能从以渐进的方式一次组合几个部分开始(例如感知和语言,或语言和记忆 - 请注意,这些都不是不同的类别!) ,随着时间的推移,一幅更全面的图景将开始显现。


最终,对双语进行建模意味着超越对两种语言学习的建模,转而对大脑处理多种符号系统的能力进行建模,这是人类思维的一个决定性特征。

更新日期:2022-11-10
down
wechat
bug