Physics language and language use in physics—What do we know and how AI might enhance language-related research and instruction

Language is considered an evolutionary adaptation, carrying out important functions like communicating 'who did what to whom' [43, p. 81]. Language is recognized as an outstanding feature of the evolution of humans [44]. Accordingly, it has been characterized to be a 'main evolutionary contribution of humans' [45, p. 611] and 'a means by which humans collectively establish and maintain coherence with one another and the grander world' [46, p. 32]. Humans developed languages to transmit 'unlimited [..] information among individuals' [45, p. 611], and to represent and make sense of their experiences, coordinate action, and think together [3, 47–49]. Language developed (and probably can only develop) in social interaction [50, 51].

Communication purposes afford that languages develop specific means to convey information. To transmit information on high-dimensional, real-world phenomena efficiently and effectively, languages are patterned and exhibit universal laws that characterize language use. Communication in language is a process where mutual understanding is rewarded, and communication has to be optimized with regards to time, information content, and other commodities [52]. The principle of least effort in communication states that 'once a word has been used, it takes less effort to use it again for similar meanings than to come up with a different word. On the other hand, people want language to be unambiguous, which they can accomplish by using different words for similar but nonidentical meanings' [53, p. 271]. Utterances such as words or phrases should 'maximize the amount of information while minimizing the cost of sending that information.' [53, p. 271]. The minimization of communicative effort and the simultaneous optimization afford regularities in languages [54–59]. Similarly, rules and universal features of languages can be determined [45]. Any known language can be described by a few simple super-rules such as subject-verb-object ordering for English (and many other languages) [43, 51]. With the mentioned super-rule and the overall subject-verb-object rule, a child can generate novel sentences according to environmental affordances, and thus communicate with the world.

Communication through language requires recipient and receiver to (re-)construct the meaning of what was communicated, either in a situation where spoken language is utilized or in some form of written communication [40]. Meaning construction is inherently constrained by the affordances inherent in both spoken and written language, and situated in a specific context defined by communicative norms [60] and rhetorical goals [47]. In a more collective dimension (subsection 2.2.1), language is linked to disciplines and ways of knowing which define respective communities. As such, a command of respective languages as a mode of representation of ways of knowing [12] is required to make sense of the discipline, and a shared experiential basis facilitates meaning construction. In a more individualist dimension (subsection 2.2.2), language, constrained by specific human affordances, enables humans to make sense of the disciplinary ways of knowing and functions as a resource for understanding and reasoning.

2.2.1. Collective dimension

2.2.2. Individual dimension

3.1.1. Studying physicists' writing

3.1.2. Studying students' writing in physics (and science) classrooms

3.1.3. Listening to physicists in the lab

3.1.4. Studying verbal interactions in physics classes

3.1.5. Understanding historical roots of terms in physics

Language is a resource to make sense of phenomena and processes, and reason in physics. As such, an important function of language in physics and beyond is to enable the description and modelling of physical experiences, phenomena, systems, and processes in nature [13, 104]. Languages encode a specific ontology that influence how people think about physical entities, systems, and processes [105]. This is also called implicit or intuitive ontology. Intuitive ontology refers to one's conception of the basic categories of existence [106]. Ontology in physics comprises matter, processes, and physical states [13]. For example, physicists might state that 'alpha particles (matter) leak through (process) a potential barrier (matter)' [13]. The physical state is denoted by the fact that the alpha particles are contained in the nucleus [13]. Physical concepts are considered to be attributed to either one ontological category by learners, and may change the ontological category in phases of conceptual change [107].

Confusion or misattribution of ontological categories was argued to comprise a major hurdle in developing an appropriate understanding of physical concepts [108]. Such as when 'abstract physics concepts [..] tend to be attributed with properties or behaviors of material substances.' [109, p. 1]. While substance-based reasoning is certainly productive for the energy concept, for other concepts such as force, light, heat, or electricity (e.g. electrical current) it can create conceptual confusion [109]. '[M]any scientific concepts are 'constraint-based interactions' [..], which is a subcategory of processes in which a defined system behaves according to the principled interaction of two or more constraints' [107, p. 9]. Electrical current, in this conceptualization, is a constraint-based interaction. Contrary to events, it does not involve a specific beginning and ending. Moreover, even though it involves charged particles as matter, it does not fall into the ontological category matter [108]. Constraint-based interaction is a subcategory of process, and it is determined by a known set of constraints [108]. Physics concepts that are classified by experts as constraint-based interactions are then often incorrectly classified as material substances by novices [110]. However, it has also been noted that ontological attribution must not be thought of as static, but rather as complex and dynamical [107] and that experts' ontologies exhibit more local coherence.

Even false ontological commitments can lead to productive reasoning. French physicist Sadi Carnot, in his derivation of the ideal heat engine, utilized the caloric metaphor (heat as an elastic fluid) to determine the maximum amount of motive force [111]. Whereas some experts characterize heat as disordered/random energy, others refer to it as energy in transit, however, it cannot be energy per se because it is not a state function [13]. It was then argued that the caloric metaphor is outdated and should not be used anymore in teaching physics and in physics language. Some suggested to entirely abandon heat as a noun from physics vocabulary to stress its function as a process function [112]. Similarly, 'heat transfer' and 'heat capacity' invoke a substance ontology, and should be used cautiously. Others have even advocated for the abandonment of the verb 'to heat' [113].

Of great relevance in physical systems are so-called relational processes, expressed preferably through the verb 'to be.' Identifying and attributional relational processes are then distinguished [105]. Physical states are oftentimes represented through identifying relational processes where grammatical circumstance of location replaces the second identifier, such as 'the electron is in the ground state.' In fact, these grammatical structures closely resemble language on mental states ('I am in love', [105]). Ontologically, however, the verb 'to be' is also problematic, because it identifies individuals with abstractions [43]. This is pervasive in physics language. The electron is a fermion/wave/particle/... [114]. All sorts of ontological confusions may arise. After all, waves are mathematical abstractions that exhibit certain properties (e.g. spreading of excitation of physical variables across time and space, better thought of as events rather than objects [10]). However, some students readily imagine there 'being' waxing and waning levels of the physical quantity quite literally in space and time, or expect sound waves to propel dust particles around the room [10].

It is also important to recognize the constraints that language puts on descriptions for physical systems and processes. It is recognized that '[n]atural language is very limited in its ability to describe continuous variation, shape, and movement in space' [1]. Language can only qualitatively describe relationships among variables, but for more complex functional relationships researchers resort to mathematical equation, pictures, graphs, tables, etc. Arguably much of scientific reasoning happens through mental modelling and carrying out thought experiments on the internal models [115]. Language, then, works in conjunction with the other modalities to make sense of models. However, internal models are not represented as natural language [43].

In sum, ontologies in language provide the necessary resources to represent physical processes and phenomena through languages. Given the everyday use of terms ('energy consumption'), ontological confusion restricts students to learn the physical meaning of concepts, however, utilizing different ontologies for concepts can function also as a productive means for reasoning processes. Being aware of ontologies and ontological confusion enables researchers and students alike to gain a deep understanding of physics concepts as represented in language.

Metaphors and analogies in physics function as productive modes of reasoning about physical systems and are closely linked to ontologies [100]. One of the major productive resources in physics language are conceptual metaphors, which are oftentimes rooted in analogies. In fact, metaphors enable us to communicate about abstract physical entities, systems, and processes. Conceptual metaphors involve mappings from concrete knowledge structures 'incorporating notions of object, space, movement, and force' (concrete, conceptual source domain) to 'abstract concepts such as time, cause, depression, and love for which no (or little) experience-based schematic structure can be described' (abstract, conceptual target domains) [11, p. 79]. Physical systems in mechanics might often resemble concrete sensory experiences (such as a falling apple that can be directly observed). However, to represent systems in electricity, magnetism, optics, acoustics, radioactivity, quantum physics, atomic physics, particle physics, etc system descriptions are largely metaphorical and unaccessible to direct sensory experiences. ⁴ 'Concrete domains are those conceptual schemata that derive directly from the sensorimotor experience of interacting with the physical and social worlds, whereas abstract domains are those schemata for which no such experiences can be identified' [11, p. 79]. Conceptual metaphors are a productive means to make abstract domains understandable through language. The sentence 'a net force causes an object to accelerate' [100], see: uses the conceptual metaphor: force is an agent, however, it gives no account of the idea (even hinders understanding) of forces being interactional. Conceptual metaphors are widely employed in physics and the natural sciences, and they oftentimes reflect 'embodied experience' and are used to draw inferences about abstract domains [100].

In thermodynamics, researchers identify Location Event Conceptual Metaphor. Within this framework, they differentiate concepts such as (a) 'Thermodynamic States Are Locations,' (b) 'Thermodynamic States Are Movements,' and (c) 'Spontaneous Change is Directed Movement.' In (a), the static location of states in phase space can be referred to. In (b), the changes in a state, going from point A to point B in phase space, are referred to, and in (c) natural directionality of thermodynamic processes is encapsulated [11]. Conceptual metaphors in quantum mechanics relate, among others, to the potential well. Talking about the potential-'well' evokes associations and assumptions that make reasoning more easy [100]. However, that also might introduce oversimplification and confusion. 'Swackhamer [..], building on the work of Lakoff and Johnson on conceptual metaphor [..], analyzed the energy as "stuff" conceptual metaphor and identified three statements derived from the metaphor which guide the development of energy concepts and an understanding of energy conservation with students: (1) As an attribute, energy is viewed as a possession that can be "stored" or "contained" in a "container," namely, a physical system. (2) Energy can "flow" or be "transferred" from one container to another and so can cause changes. (3) Energy maintains its identity after being transferred.' [104, p. 2] In fact, the 'stuff' conceptual metaphor for energy is productive in nature [104]. Energy is a concept that refers to something quantitative (coming in quantities) that can be balanced. This is beneficial, because one can use metaphors to indicate that an object has a certain amount of energy (as opposed to force), which is easier to understand [89].

Conceptual metaphors are closely linked to grammatical metaphors and ontologies in systemic functional grammar. Science and physics education researchers utilized systemic functional grammar to analyze language-based meaning construction in physics learning [1–3]. Systemic functional grammar recognizes the specific functions of language such as interpreting experiences, communicating ideas, and meaning making. As such, language is recognized to always have specific rhetorical functions that are important to consider in order to construct meaning from utterances. Grammatical metaphors seek to make abstract concepts understandable. In systemic functional grammar, grammatical participants (nouns, noun groups), grammatical processes (verbs), and circumstances (adverbial or prepositional phrases) are differentiated. Grammatical participants are then mapped to ontological categories [105]. Moreover, physical models consist of (1) models of objects, (2) models of interactions between objects, (3) models of systems of objects, and (4) models of processes [105]. Then, (5) physical properties, and (6) state functions are physical variables used to describe the system or the objects in it. These physical concepts are then mapped to an ontology tree, which is called a lexical ontology, and every physical model described in language has an ontology, and this ontology is encoded in the grammar of the sentence [105, p. 5]. If the lexical ontology does not match the grammatical ontology, grammatical metaphor is present. The sentence 'heat [medium] flows [process] from the environment to the gas' [105, p. 5]. The meaning of the sentence (lexical ontology) refers here to the energy that is transferred into the system. But grammatically, heat is functioning as the participant, i.e. as the concrete manifestation of matter that is flowing (grammatical ontology). Grammatical and lexical ontology differ in this sentence, which—on this basis—can be identified to be metaphorical [105]. 'Conceptual metaphors often give abstract concepts an existence as concrete objects or things' [100, p. 4] and allow us to understand it, quantify it, identify it as a cause, etc. However, there is the risk that conceptual metaphors are taken at their face value and are not explicitly reflected to be metaphors in everyday language [69, 105], which might hinder genuine understanding of the involved processes, systems, and models. For example, when learners hear or say 'gravity pulls' and encounter textbook definitions where forces are introduced as agencies/influences they are less likely to develop a physically accurate, interactional conception of forces [116].

(Conceptual) metaphors are rich resources in language to capture physical phenomena and processes, however, learners of physics need to get acquainted with the specific conceptual metaphors that are used in this discipline. As with the other features of physics language and language use in physics, this process will likely benefit from explicit guidance rather than expecting students to learn this from merely participating in physics-related communication situations.

Besides the underlying, metaphorical nature of physics language, the meaning of words itself and the context in which words appear are important considerations when trying to understand physics language (use). Words can have different meanings in physics depending on the context in which they are used: 'What is perhaps most confounding [in physics language use], we even use words which we define precisely yet their meanings change with context, or even with speaker' [5, p. 672]. Context itself has a certain granularity as well, local and global, and external and internal [72]. Meaning in reference to context can depend on topic (e.g. mechanics) or discipline (e.g. chemistry). For example, work in the work-energy theorem context has a different meaning as in the second law of thermodynamics [5, 117]. Arons (referenced in [117]) suggests to use 'pseudowork' for the former, to either show the connection between both and make clear that they are distinct. After all, seven types of work can be differentiated [117]. It is thus important to signify the topic in which work is used (often gleaned from the context). Moreover, meaning might also differ depending on discipline. A physicist and chemist might attribute different meaning to the term molecule, given their different traditions in their disciplines to define the word [5].

Without context it is impossible to specify the meaning and predication of a word. Moreover, language is always relational, and fixed definitions can only be approximations of the particular meaning of a word in context, which resonates with Wittgenstein's concept of language-games. Thomas Kuhn stresses this aspect in his seminal work 'The structure of scientific revolutions' as follows: '[The] 'definition' of an element is no more than paraphrase of a traditional chemical concept. Like 'time,' 'energy,' 'force,' or 'particle,' the concept of an element is the sort of textbook ingredient that is often not invented or discovered at all. [...] The scientific concepts to which they [the definitions] point gain full significance only when related, within a text or other systematic presentation, to other scientific concepts, to manipulative procedures, and to paradigm applications.' [118, p. 142]. After all, dictionaries are poor sources for definitions, given the context-dependence of word usage [116]. Not only the semantics matters, but rather also the locations (syntax) of how words are used.

Among the most deceiving contexts in which meaning differs is between physics language and everyday language: ⁵ '[M]any of the words we define precisely carry the same meaning but less precisely in their everyday usage. Other words carry different, often contradictory meanings' [5, p. 672]. Common sense terms in physics can be deceptive, as, for example, neither the terms mixture, suspension, or solution in ordinary language have the same meaning in scientific language [88]. Moreover, everyday language even uses physics concepts in contexts that invoke false associations. For example, heat is often used in everyday contexts (at times also in physics textbooks) in ways that invoke the caloric (substance-based) metaphor as in: 'Winters are milder near the coast because the ocean holds a lot of heat' [112, p. 107].

Word meaning is nothing static, but rather dynamically depends on the context (and communication situation). This enables learners to productively use language and create their own meanings, however, it also causes interference of physics language with everyday language. In one form or another, it is thus important to raise such issues in learning settings. Creating the illusion that the meaning of terms in physics is static can thus create problems when students use physics language outside of classrooms.

Physics education researchers also posited that experiences in the physical world interact with the ways humans construe and reason about physics processes and phenomena. Experiences are considered as a resource for developing phenomenological primitives (abstractions of common events), i.e. frames or schemata, in which physical phenomena are interpreted [101]. Characteristic for this knowledge system is: 'causal schematization in terms of agents, patients, and interventions ('causal syntax'); a tendency to focus on static characterizations of dynamic events, including the global form of trajectories; and a relatively rich phenomenology of balancing and equilibrium' [101, p. 105]. For example, 'people develop on the basis of their everyday experience remarkably well-articulated naive theories of motion' [119, p. 301]. However, they would not derive the laws of motion as laid out in physics theory, because these were refined over centuries with special purposes (e.g. to predict planetary motion). Phenomenological primitives are models for core intuitions and considered an elemental cognitive structure that is a resource to guide interpretation of experiences [10, 101]. These findings suggest that language-sensitive instruction includes opportunities to lay out one's experiences and understanding (in own words) for physical process and phenomena.

Suggestions for language-sensitive instructional design come from more general language research, and specific discipline-based educational research in science and physics classrooms. General educational research on language acquisition suggests that the more abstract the concepts and phenomena are (i.e. away from students' everyday experiences), the more explicit scaffolding is necessary [121]. Moreover, everyday language (and other non-disciplinary semiotic resources, see [39, 122]) and experiences should play a role in physics classrooms, as this is less alienating to students and reduces the barrier when discussing physics phenomena with the language they are already familiar with [123]. Addressing students' experiences was also found to be an important contributor to building cumulativity in classroom communication, helping break authoritative patterns of teacher-students interactions [124]. Utilizing what students already are able to (i.e. produce everyday language) as a resource for physics learning can thus enhance classroom communication.

Increasing clarity and elaboration also raises reading comprehension for science texts [125]. Evidence from intervention studies suggests that language-sensitive learning materials can help students in their development of conceptual understanding [126]. For example, the Science Writing Heuristic engages students in immersive argumentative writing (producing language), and evidently supports argumentative skills, metacognition, and other important outcomes [126–128]. Science writing is a means to prepare students for future careers in science as writing as a means of communication is an important part of the scientific discovery process [1, 129].

Language is but one disciplinary mode of representing ways of knowing in physics. Learning and deep understanding of physics concepts rather hinges on critical coordination of different disciplinary modes [12]. It is therefore important to utilize language in conjunction with other modes such as mathematics, observations, etc in order to enable students to understand how these different modes in conjunction enable understanding, and how certain modes afford only certain insights and inferences.

Metacognitive prompts, e.g. on text composition, can positively influence distant variables such as academic outcomes. Language-sensitive instructional design in physics thus involves the posing of questions and strategies for enabling students to revise their positions and understanding rather than reliance on short responses [83]. Moreover, prompting students to explicitly write how they would qualitatively solve a physics problem was found to be an effective strategy to train their physics problem solving capabilities [130]. This then allows instructors to better diagnose the students' understanding of the problem situation and provide more specific guidance. Making explicit conceptual and strategic knowledge through language is in itself a valuable resource [131].

In physics learning environments, students should get plenty of space to comprehend physics language and produce it by themselves, optimally with feedback and guidance from instructors [1]. Empirical research suggests, however, that too few opportunities are established for students to comprehend and especially produce language [132, 133]. Moreover, assessing language products requires instructors to invest much time to individually guide students [134].

Given the prevalence of unstructured, e.g. unlabelled, language data [142], physics education researchers employed tools for data-driven discovery such as NLP with the help of unsupervised ML. Unsupervised ML in PER has been used to extract distinct themes from students' interview transcripts on explaining the seasons [145]. To validate the findings, specific indicators were established to determine the meaningfulness of the automatically extracted themes for human analysts. Validation criteria included if the themes '(a) were interpretable in terms of the theory, (b) captured knowledge at an appropriate grain size, (c) captured combinations of elements, and (d) captured the dynamics of individual interviews' [145, p. 633]. It was found that many of the extracted topics well matched theoretical expectations (from knowledge in pieces theory and common sense physics reasoning). In another study that sought to group unstructured data, an unsupervised ML model named Latent Dirichlet Allocation (LDA) was used to extract themes that were discussed throughout the Physics Education Research Conference [146]. In this research, NLP methods helped to preprocess the articles in a meaningful way. For example, the authors combined NLP methods to reduce word forms, and detect language patterns that account for larger patterns of words. Different means for validating the model were used, such as coherence and perplexity of the topics, face validity, and comparison with other qualitative methods. After validating the topics, emerging themes in PER or overall important topics could be singled out from a large, unstructured language data set.

Unsupervised ML has also been utilized in intelligent tutoring systems in physics. Both symbolic and statistical NLP techniques are used to intelligently parse input, interpret it, and generate responses in tutorial dialogues. A well-known implementation is AutoTutor that is based on a strategy named expectation and misconception-tailored dialogue [147]. Besides the requirements for researchers to specify the prompts and expected answers and misconceptions (non-normative ideas), AutoTutor utilizes statistical NLP, namely latent semantic analysis. Latent semantic analysis is a means of reducing the dimensionality of the language data. As such, it allows researchers to calculate the similarity (i.e. distance in high-dimensional space) of documents. Latent semantic analysis was then used to match expectations and misconceptions in students' responses, hence allowing the tutoring system to track the current level of understanding of students in the solution space [137]. Of course, specification of expected answers and misconceptions was necessary in the form of curriculum scripts. It is important to note that in these unsupervised approaches established theory in PER and educational sciences formed the background against which the validity of the findings was evaluated. If the algorithms picked up on novel patterns, these might not have been recognized by the researchers.

In sum, unsupervised ML approaches can be utilized to explore patterns in complex language data. This can help researchers in PER to browse through data repositories of unprecedented size and extract information from them. For instance, the disciplinary structure, encompassing the topics addressed over time, could be analyzed using existing physics journals as a foundation. Moreover, unsupervised ML approaches could facilitate adaptive guidance in tutoring systems and thus foster physics learning processes.

Some of the language data is structured, i.e. annotated, coded, or labelled. Oftentimes, researchers spend parts of their research for developing and validating coding rubrics and coding data. For this data, supervised ML methods can be used. A wide variety of supervised NLP-based software tools exist that can help in analyzing constructed responses in PER. For example, the program LightSIDE was used to extract feature lists from text data [148]. The seemingly simplest text feature, namely the words that occur in a response, was selected to predict category membership. Validation of the ML model was performed through comparison with human raters. Human-machine agreement was then calculated as a proxy for performance of the ML algorithm. Mean response length was 1–2 sentences, which is quite typical for constructed responses by students for a simple prompt. The authors analyzed responses on Newton's first law, e.g. in the ball-track-question, where they asked what happens to the movement of a ball, if a certain track was extended. The ML model could then be employed to automatically determine whether students provided answers regarding whether the movement would slow down, speed up, and so forth. In another question concerning crash test dummies, the ML model successfully discerned whether a student's response was purely descriptive, or included an explanation related to forces (as the question prompted students to explain). As such, the ML model could be used to request a rewrite if students did not explain the phenomenon. Furthermore, the ML model performed reasonably well in identifying distinct groupings based on various strategies for removing a stuck coin from a graduated cylinder. The authors ascertain that self-validation proved to be a viable method for approximating results. However, as expected, cross-validation results, which involve testing the ML model on unseen data, consistently yield lower outcomes in comparison to self-validation results. Furthermore, in line with expectations, more examples for a category in the training data result in improved classification accuracy for that group. The authors conclude that 'larger data sets being more desirable suggests that collaborations that could combine smaller response sets may present a significant advantage in further exploring ML in physics education' [148, p. 13].

Supervised ML was also used to predict students' score in the Physics Measurement Questionnaire, based on their constructed responses. NLP was used to preprocess responses (vectorization) in order to feed them into a ML model (logistic regression, random forest, k-nearest neighbors). All ML models performed about as well as human raters. Moreover, they found that most predictive words were well aligned with expected words in pointlike and setlike paradigm [149]. In a similar manner, supervised ML has been used to analyze middle school students' explanations of thermodynamics phenomena [150]. The authors provided then guidance to the students based on an automatically generated (ML-based) score. One motivation for this usage was that students' misconceptions include that metals 'conduct,' 'absorb,' 'trap,' or 'hold' cold better than other objects [150]. The authors used the NLP program c-rater to automatically identify distinct concepts in students' explanations of a thermodynamic problem ('A metal spoon, a wooden spoon, and a plastic spoon are placed in hot water. After 15 s, which spoon will feel the hottest and why?'), such as 'The metal spoon feels like a different temperature than its actual temperature,' or 'Metal conducts or heats up or transfers heat (NO comparisons OR rate)'. Overall, supervised ML is oftentimes utilized to classify constructed responses into pre-defined categories and automate this process, which raises scalability of analyses of language products in PER.

Supervised ML approaches in PER can help researchers to automate their research and instruction. For example, validated coding rubrics could be fed into ML models to train automated classifiers that could then be used either for diagnostic purposes in large scale studies or in physics learning environments to classify students' responses and provide them feedback.

Capturing the statistics (e.g. recurring patterns and themes) of natural language and predicting next words in a sequence can be defined as the task of language modelling. Modelling tools became very sophisticated with advances in AI research [26]. In particular, artificial neural networks excelled in language modelling [26, 151–153]. An important goal in these approaches is to represent words as continuous vectors where each word has a vector based on the context that it appears in, following the distributional hypothesis that words that appear in similar contexts have similar meaning [75, 79]. Contextualized embeddings were trained with specific language model architectures, established as LLMs. LLMs are typically complex neural network architectures that have the overall objective to predict next words in a sequence [154]. As such, the trained LLMs can be used to generate a response to a prompt (input), so-called generative LLMs, or they are used as intermediate means to represent language data, where these representations are used in other ML models to classify text. A popular example of generative LLMs include GPT (generative pre-trained transformers). These LLMs can be used to generate new text, given some input and prompts, such as in the example of ChatGPT. Other LLMs include BERT (bi-directional encoder representations from transformers) [153], which can be used to retrieve contextualized embedding vectors for input.

Training LLMs has become pivotal to the success of contemporary NLP systems. LLMs are trained on large repositories of language data, such as the Common Crawl of the Internet (410B tokens), WebText2 (19B tokens), Wikipedia (3B tokens), and probably other, undeclared sources (proprietary information of respective private companies). It was found that generative LLMs such as GPT or BERT trained on these large repositories could well produce authentic and oftentimes correct language utterances given some prompt. However, the early versions of LLMs had to be fine-tuned to particular tasks in order to perform well on them. These problems were solved with increases in training steps (compute time), model size, and also prompting techniques [34, 155]. LLMs such as GPT-3 became what was called few-shot or zero-shot learners [155]. They were capable of completing tasks with only a few or no examples demonstrating how to solve such tasks. For example, chain-of-thought prompting was used to enable LLMs to perform simple reasoning tasks [156]. In chain-of-thought prompting, the LLM is commonly provided input-output pairs (questions and answers). Rather than instantly producing the answer, in chain-of-thought prompting the model is required to produce intermediate reasoning steps. Google Research applied chain-of-thought prompting to PaLM (pathways language model) and reached state-of-the-art performance on GSM8k, a data set of 8.5K high quality linguistically diverse grade school mathematics word problems [157], with 58%. The famous conversational AI ChatGPT (based on the predecessors, Generative Pre-trained Transformers, GPT-3.5) from OpenAI was trained both with supervised and other forms of learning (e.g. with human feedback). ChatGPT scores well in MBA exams, medical exams, and verbal-linguistic IQ [158]. However, in advanced mathematics results are more mixed [158].

LLMs became promising tools for language analysis in PER. By merely processing large amounts of language data, LLMs performed (at least) above chance in some extent reasoning tasks about physical common sense [159]. Novel training methods such as prompting allowed LLMs to increase performance in tasks of quantitative as well as common sense reasoning in domains such as physics and mathematics [34]. With further fine-tuning the LLMs on mathematics and physics data, the models surpassed performance of other LLMs to solve quantitative reasoning problems (including derivations of the solution) to unprecedented accuracy, surpassing human average in national mathematics exams [34]. An acknowledged issue in LLMs is memorization, meaning the ability to solve problems by merely recalling samples from the training data (thus characterizing LLMs as 'glorified lookup tables', or 'stochastic parrots' [160]), rather than through genuine synthesis based on understanding the underlying concepts. The authors have, to a certain extent, addressed this concern by demonstrating that the problems presented in the test data were not present in the training data and that adjustments to prompts, whether in wording or numerical content, correlated with variations in solution performance.

LLMs could also be used to enhance unsupervised ML approaches, e.g. to inform topic detection [161]. For this purpose, the feature of LLMs to transform an input sentence into a contextualized, dense embedding within a high-dimensional vector space was employed. These embeddings can then be used as a feature to be forwarded into another ML model which performs clustering. In this case, an LLM was used to encode sentences in pre-service physics teachers written responses (here: reflections). Afterwards, dimensionality reduction and clustering algorithms were used to extract topics in the reflections. Similar validity indicators as in [145] were utilized, and it was found that pre-service physics teachers' reflections entailed rather general and more physics-specific topics, which is consistent with research on pre-service teacher's noticing [162, 163]. Again, substantive theory in PER and educational science was required to validate the extracted topics and findings. As with unsupervised ML, LLMs could also enhance supervised ML through efficient data transformations. Wulff et al [164] used ML algorithms to classify pre-service physics teachers' written reflections into categories provided by a reflection-supporting model. The performance and generalizability could be substantially boosted by using LLMs to transform the input data [161]. Moreover, the LLM-based model could undergo additional fine-tuning for various contexts, allowing for the classification of written reflections from non-physics students [165]. This offers new possibilities in PER to share models and thus scale research.

Tschisgale et al [166] showed how both unsupervised and supervised ML and NLP could work in tandem with human experts to discover patterns in physics students' written problem solutions. The authors applied what is called 'Computational Grounded Theory' and used unsupervised ML on the basis of LLMs in a first step to discover patterns in the problem solutions. The patterns were cross-validated with human experts. Finally, a classifier was trained to automatically detect the identified patterns in unseen data and the final model presents a validated ML model that could be utilized in teaching and learning settings, across research sites.

Recently, many physics education researchers investigated the performance of the generative AI model GPT (especially versions 3.5, and 4.0, as implemented in conversational AI ChatGPT) for solving physics problems. Applications are versatile, given that GPT as a generative AI that continues input text can parse and process all kinds of open-ended and closed physics problems (as of October 2023 image processing capabilities are limited). West [33] found that ChatGPT was capable of physics conceptual problem solving as presented by the widely used Force Concept Inventory (FCI) [167]. In fact, ChatGPT dramatically improved its performance from version 3.5 to 4.0 with a gain g = 86.7%, 'which is more than twice the average effect of a pedagogically designed, interactive physics course' [33, p. 2]. This progress led to an almost perfect expert-like solution probability (note that graphical representations had to be transcribed for GPT to parse them). However, occasionally very basic errors that are not expert-like occurred [28, 33]. A deep, socratic conversation with ChatGPT on the question 'A teddy bear is thrown into the air. What is its acceleration in the highest point?' [29, p. 1] also produced knowledge statements that were unreliable, despite the linguistically advanced manner in which they were presented [29]. ChatGPT was then found to excel in short-form physics essays, i.e. perform on par with students [168]. Instruction with ChatGPT could increase students' perception of ChatGPT, and it is expected that ChatGPT can play an important role in instruction [30, 33]. For example, Küchemann et al [31] showed that ChatGPT could assist physics students with task generation, and Kieser et al [32] demonstrated that ChatGPT could be used as a tool for educational data augmentation: ChatGPT was successfully used to simulate students with certain pre-conception (e.g. impetus) related to forces. This offers novel opportunities to pilot test instruments in physics, augment data, probe instructional materials, and even generate instructional materials.

In sum, LLMs offer potentials to improve accuracy for unsupervised and supervised ML approaches in PER. They help represent language data in a more nuanced way (contextualized embeddings), and capture more complex (non-linear) relationships that can be harnessed for automated classification. In addition, generative LLMs offer avenues for language production when prompted, which has been demonstrated to be beneficial in both research and teaching within PER.

Language-related research and emerging AI-technologies have played and will play an important role in physics instruction. Given the analytical power and automation potentials of ML, NLP, and LLMs, these technologies can enhance adaptive instruction, and spare the valuable time of instructors who can be (partly) freed from repetitive tasks such as essay scoring. Given the importance of language as a semiotic resource and the scarcity of opportunities for students to produce language in physics classrooms, many language-involved activities and applications are conceivable such as interviews with simulated historical figures or adaptive problem solving. Moreover, teacher dashboards become possible with the help of ML, NLP, and LLMs that can track language use in classrooms. This might open up classroom communication by identifying and breaking cycles of authoritative communication or similar strategies that narrow students opportunities to actively engage in physics learning environments. It is then the responsibility for researchers and instructors to design these technologies in ways that they enhance the learning processes of students, rather than eliminating essential language-related learning opportunities.

Physics education researchers also explored the caveats for these applications such as false confidence in factually wrong answers by LLMs (also known as hallucinations), which always have to be critically monitored, especially when implemented in educational settings with students. It was discovered that LLMs acquired similar biases to those of humans, reflecting the fact that their training data encompass diverse sources from the Internet [35]. These issues pose serious challenges to educational applications, as students might be confronted with unproductive problem solving strategies, or worse, biases and stereotypes related to their social identities. After all, language data on the Internet (i.e. training data for LLMs) is full of harmful language. However, considering bias will also involve trade-offs, as we know from educational research that teachers expose certain biases as well. For example, studies reported gender biases in physics classrooms—such as teachers' implicit, and gendered theories on giftedness [169]. If students interact with LLMs and do not expose their gender orientation, this might reduce bias in guidance. Also, average human tutors were found to struggle to adequately judge a student's understanding and adapt their guidance [170]. Hence, even if automated tutoring systems achieve accuracies of only 70%, they may outperform inexperienced human tutors not only in terms of accuracy but also in considerations of time and cost [148].

Researchers laid out in what ways general ML nowadays can assist in different phases of scientific discovery and automate them [171]. When utilizing ML, NLP, and LLMs in PER, we found that researchers used various ways to ensure model validity. Mostly, substantive domain expertise was necessary to interpret and validate output of ML models. As such, the established body of knowledge in PER on language and language use in physics and beyond is crucial to direct the development of these AI-technologies. As of now, ML and NLP rather enable researchers to outsource repetitive tasks and enable adaptive, automated guidance in their learning environments. For example, automated transcription through ML models of verbal interactions in classrooms can yield new insights into interactional patterns to help refine our understanding of classroom communication, or language use in informal (learning) environments as well. These technologies could also be applied to navigate existing data repositories in order to reconstruct the evolution of meanings associated with physics terminology over time. LLMs could even be trained to predict 'optimal' physics terminology to ease communication and understanding, and lower interference with everyday language. LLMs might discover entirely new and different concepts in the first place. ML and NLP have already been used in PER to suggest curriculum sequences [172]. However, standards for ML model validity in PER should be established in order to ensure robust theory development.

A fundamental question relates to the actual capabilities ⁶ that the LLMs acquired from merely browsing through language data. While some researchers posit that LLMs such as GPT-4 show 'sparks of artificial general intelligence' ⁷ , and challenge our understanding of learning and cognition [173], others remain more sceptical and question the claim that LLMs are somewhere near artificial general intelligence or even an efficient architecture to process language in the first place [174, 175]. Be that as it may, cognitive semantics, social semiotics theory and multimodality theory suggest that physical embodiment in the world and exploratively engaging in language-games based on various semiotic resources (not only language) is crucial to develop a deep understanding of concept as used in physics. Based on these premises, it is contested that LLMs do acquire a (deep) understanding of physics concepts. Browning and LeCun [176] recognize the limitations of language and conclude: 'these systems [LLMs] are doomed to a shallow understanding that will never approximate the full-bodied thinking we see in humans.' Rather LLMs can be considered as performant interpolation machines, i.e. they excel at generating similar language which has been seen during training. Given that it is likely that LLMs trained on the Common Crawl (dump of the Internet) saw some form of concept inventories during training, this sets the high performance on these concept inventories into perspective. While some researchers demonstrated that LLMs could generalize beyond physics problems that they have seen in the training set [28, 177], the reach of this transfer is often narrow, given that names and objects are changed. Solving these transformed problems would not necessarily account for deep physics understanding, and researchers should be particularly careful to attribute processes such as reasoning or argumentation capabilities to these models, i.e. anthropomorphize them.

A purely language-based learning approach for LLMs is hardly capable of learning representations for perceptual features such as smell or colors [178], reinforcing the concern that—without physical grounding, i.e. embodiment—LLMs cannot learn adequate representations of words and concepts [178, 179]. Quite interestingly to this point, some colors can be meaningfully internally represented to some extent by LLMs [180], and also meaningful internal representations of board position in the board game Othello could be acquired from merely predicting moves in the game [181]. Yet, physics understanding is contingent on other semiotic resources than language such as visualization and mathematics. Given the limitations of language itself, it is reasonable to assume that LLMs in their current form do not possess anything resembling a physics-related understanding of the world, i.e. formed internal representations that transcend mere linguistic associations and encapsulate physics. While it has been documented that LLMs demonstrate improved performance with increased training steps, larger training data sets, and more extensive parameter configurations [182, 183], the inherent constraints of language and the absence of embodiment likely hinder LLMs from developing a deep, expert-level unerstanding of physics. In terms of Wittgenstein: what these LLMs most likely learned is to play different language-games without any reflection on the structure of communication and semiotic processes. Once, language processing and processing of perceptual sensor information is combined with opportunities for AI to engage with the physical world (robotics and multimodal models), new possibilities emerge as to what these systems can possibly learn in terms of a physics-related internal representation of the world. Current LLMs are oftentimes black-boxes, in which model decisions are not understood well [184]. To what extent these enhanced, multimodal foundation models [185] then form similar ontologies and conceptual metaphors as present in physics language is an open question. AI researchers began to integrate robotics and concept formation (e.g. schema networks) to derive knowledge (and, eventually, language) from merely observing and interacting with the physical world (real or simulated), with some successes in generalizing and understanding causality of events [186].

Ethical and ecological concerns, along with the need for transparency, pose specific challenges when it comes to the utilization of LLMs in language-related research in physics [187]. Also, biased model decisions pose concerns especially when LLMs are employed in educational settings or for high-stakes testing [187]. It is important for researchers to assure that their models do not advantage certain groups, especially related to gender or race [166, 188]. LLMs are seldom open sourced ⁸ and involve private contracts to use private data sets for training. At times, even private information might be revealed through these models with specific prompting. Researchers recommend open sourcing the LLMs in order to better understand model decisions. Moreover, research in the explainable AI tradition should seek to illuminate why LLMs generate certain outputs. Finally, it is important to also consider the CO₂ footprint of LLMs [189]. Training and inference cost large amounts of CO₂ (since energy systems are not entirely renewable in most countries) and financial investments in cloud computing infrastructure [190]. These need to be balanced with the pedagogical benefits of instructing students more adaptively and in an individualized manner.