XKD-Dial: Diagrammatic AI Language

• XKD-Dial is a diagrammatic language that serves as an engineering schematic for AI systems, integrating architecture, data management, and evaluation.
• It employs modular dialects like DIAL-SYS and DIAL-NN, offering tailored representations for system-level functions and neural network components.
• XKD-Dial aims to improve interpretability, reproducibility, and communication by standardizing heterogeneous AI diagrams into a coherent schematic format.

XKD-Dial, as described in "The Diagrammatic AI Language (DIAL): Version 0.1," denotes DIAL (The Diagrammatic AI Language), a proposed diagrammatic notation for AI systems intended to function as an “engineering schematic” for AI Systems. It is presented as an initial specification rather than a finished standard, with the explicit aim of supporting community dialogue toward a common diagrammatic language for AI systems. The proposal addresses the absence of a consistent model for visually or formally representing AI-system architecture and introduces a symbolic, typed, and extensible notation organized into dialects, notably DIAL-SYS and DIAL-NN (Marshall et al., 2018).

1. Motivation and Problem Setting

The central motivation for XKD-Dial is the claim that AI systems currently lack a consistent model for visually or formally representing architecture. In practice, AI papers mix arbitrary diagrams, algorithms, formulae, and natural language, with little consistency in abstraction level or notation. The paper associates this heterogeneity with problems of interpretability, correctness, completeness, transparency, dialogue, and reproducibility (Marshall et al., 2018).

The stated objective is to reduce the communication burden between AI researchers and engineers by providing a shared “engineering schematic” for AI systems. In this formulation, diagrammatic representation is not treated as a cosmetic aid, but as a medium for expressing the core functional elements of a system in a way that is more coherent than heterogeneous combinations of prose, mathematics, and ad hoc figures.

The paper also positions DIAL against UML. It argues that UML is not a good fit for AI-system communication because it is often too general, too complex, and not widely used by AI practitioners for representing AI-specific structure. This suggests that DIAL is intended not as a generic modeling language, but as a notation specialized for AI workflows, resources, and transformations.

2. Scope, Design Goals, and Representational Commitments

DIAL is intended as a self-contained description of the core functional elements of an AI system, with special attention to datasets, data management infrastructures, gold standards, and evaluation metrics, all of which are treated as first-class citizens of the representation (Marshall et al., 2018). This is a distinguishing design choice: the notation is not limited to model internals, but includes resources and evaluation structure that are often left implicit in conventional architecture figures.

The design goals are stated explicitly. DIAL should focus on functional components and data transformations; represent systems at a level suitable for complex multi-component AI systems; avoid privileging any one AI technique, especially not just neural networks; use non-linguistic symbolic language for recurring primitives; be extensible to software engineering and data management aspects; and communicate performance of components. In this sense, XKD-Dial is conceived as a system-level language rather than a notation for a single paradigm or model family.

The design approach combines bottom-up design and top-down design. The bottom-up component extracts recurring features from existing diagrams and translates commonly textual elements into diagrammatic form. The top-down component imposes high-level requirements for completeness and interpretability. A plausible implication is that the language attempts to balance descriptive adequacy for current practice with normative pressure toward a more standardized representation.

3. Formal Organization: Dialects, Grammar, and Diagrammatic Conventions

DIAL is organized as a set of dialects. The paper identifies DIAL-SYS as the core high-level language for AI systems and DIAL-NN as an extension for neural network components. It also lists future dialects: DIAL-DB, DIAL-LOG, DIAL-ML, DIAL-PGM, DIAL-BIZ, and DIAL-SEM (Marshall et al., 2018). This dialect structure indicates that the framework is modular and intended to expand across adjacent technical domains rather than remain confined to a single notation layer.

The diagrammatic conventions are simple but explicit. Features that are parts of the architecture are in circles, whereas full components are in rectangles. In addition, DIAL symbols are typed with subscripts and superscripts, used to describe recurring classification tasks, associated classes or data types, tensor dimensionality, and common data models or file formats. These conventions are part of the language’s effort to encode recurring AI concepts in a non-linguistic symbolic form.

A concise summary of the dialectal structure given in version 0.1 is as follows:

Dialect	Role
DIAL-SYS	Core high-level language for AI systems
DIAL-NN	Extension for neural network components
DIAL-DB, DIAL-LOG, DIAL-ML, DIAL-PGM, DIAL-BIZ, DIAL-SEM	Future dialects listed in the paper

A recurrent misconception would be to treat DIAL-NN as the language’s center of gravity. The paper argues the opposite: DIAL is designed to avoid privileging any one AI technique, and neural-network notation is only one extension within a broader system language.

4. Symbolic Vocabulary and Typed Notation

DIAL-SYS defines a table of core symbols for system-level composition, data movement, and AI operations. Among the listed elements are Direct sum $\bigoplus$, Concatenation $+\!\!+$, Tensor product $\bigotimes$, Set $\{elements\}$, Data flow $\rightarrow$, Data flow (both ways) $\leftrightarrow$, Data persistence $\longmapsto$, System interface (e.g. service, API) $\multimap$, Composition $a \circ b$, Join $\bowtie$, Similarity / Relatedness $+\!\!+$0, Embedding projection $+\!\!+$1, Word2vec $+\!\!+$2, Classification $+\!\!+$3, Ranking operator $+\!\!+$4, Top n elements $+\!\!+$5, Deductive Reasoning $+\!\!+$6, Verification $+\!\!+$7, Datasets, Data resources, Gold Standard, Knowledge Base of functions, and Accuracy $+\!\!+$8 (Marshall et al., 2018).

The language also introduces explicit notation for data types. Examples given in the paper include Text $+\!\!+$9, Passage $\bigotimes$0, Sentence $\bigotimes$1, Character $\bigotimes$2, Term $\bigotimes$3, Word $\bigotimes$4, Dialogue term $\bigotimes$5, Sense $\bigotimes$6, Clustered word $\bigotimes$7, Image $\bigotimes$8, Query $\bigotimes$9, Answer $\{elements\}$0, Facts $\{elements\}$1, Rules $\{elements\}$2, and Classification outcome $\{elements\}$3. Example task and annotation notations include $\{elements\}$4, $\{elements\}$5, $\{elements\}$6, $\{elements\}$7, $\{elements\}$8, $\{elements\}$9, and $\rightarrow$0.

DIAL-NN adds neural-network-specific symbols, including Loss function $\rightarrow$1, Activation function, Softmax, Attention, RNN Layer (e.g. LSTM), BiLSTM Layer, GRU Layer, Convolutional Layer, Recursive Neural Network, Support Vector Machine, Ground truth of sentiment classification $\rightarrow$2, Hidden layer (forward) $\rightarrow$3, and Hidden layer (backward) $\rightarrow$4. This layering shows how XKD-Dial accommodates both system-level composition and architectural detail without collapsing them into a single undifferentiated notation.

5. Worked Examples in Version 0.1

The paper provides three worked examples intended to demonstrate how the notation functions across heterogeneous AI systems (Marshall et al., 2018).

The first example is a question answering system over unstructured text expressed in DIAL-SYS. It contains two main cycles. The KB construction cycle takes documents, performs OIE, NER, and EL, serializes into an RDF-NL file, indexes it using an inverted index with tf-idf, and exposes the knowledge base as a service. The semantic parsing cycle processes a natural-language query, applies POS tagging, performs lexical answer type (LAT) detection), performs syntactic parsing, applies SRL, uses Q-learning with KB actions to learn a sequence of operations $\rightarrow$5, and outputs the answer $\rightarrow$6. The example is explicitly intended to show that DIAL can represent both NLP pipelines and KB-driven reasoning.

The second example is a DIAL-NN representation of the model from Zou et al. It depicts two Bi-LSTM layers, attention at word level and sentence level, and joint training through creation of a word-level and sentence-level attention lexicon. The paper also notes that hyperparameters and model accuracy are shown in tables. This example illustrates how DIAL-NN extends system notation into neural-network structure.

The third example is a DIAL-SYS representation of Zhao et al. for text entailment. The diagram shows two Siamese projections with shared weights, binary-tree LSTMs, a dual-attention model between premise and hypothesis, and recursive entailment computation. The recurrence is described as follows: at each hypothesis node $\rightarrow$7, $\rightarrow$8 is calculated recursively given the meaning representation at this tree node $\rightarrow$9, the meaning representation of every node in the premise tree $\leftrightarrow$0, $\leftrightarrow$1, and the entailment from $\leftrightarrow$2’s children, $\leftrightarrow$3. The final entailment probability is obtained through a tanh activation function for softmax. The paper notes that the neural-network detail is not fully expanded in this dialect, which also exemplifies one of the framework’s acknowledged limitations.

6. Claimed Benefits, Limitations, and Open Development

The paper explicitly claims that DIAL is meant to improve interpretability, correctness, and completeness (Marshall et al., 2018). More specifically, it argues that DIAL can make AI systems easier to design, reason about, and learn from; increase the consistency and efficiency of communication about AI systems; provide a coherent perspective of the whole system; help readers understand the core methods, resources, and interdependencies; support faster understanding of papers as a kind of nano-publication; reduce the need to parse heterogeneous diagram/text combinations; and support reproducibility and better scientific dialogue. The authors also emphasize that DIAL should be cognitively efficient rather than informationally dense.

At the same time, version 0.1 is presented as incomplete. The paper states that it is not yet complete, does not fully answer all target questions yet, and still lacks some necessary symbols. It acknowledges that embedding text directly in images is not ideal. It also states that DIAL-NN does not yet include some lower-level architecture detail and that DIAL-SYS does not yet fully incorporate temporal elements or human-in-the-loop interaction. These statements are central to understanding XKD-Dial: it is a proposal for standardization, not an already stabilized standard.

The future directions are correspondingly broad. The paper calls for community feedback and stewardship; extension of DIAL-SYS and DIAL-NN; development of future dialects for data management, logic, machine learning, probabilistic graph models, business/application contexts, and semantic knowledge representations; and incorporation of temporal elements, execution performance, human interaction points, crowdsourcing, and more low-level neural architecture details. The “Request for Comments (RfC)” framing makes clear that the project is intended to evolve as a community standardization effort.

In encyclopedic terms, XKD-Dial is best understood as an early, explicitly provisional proposal for a standardized diagrammatic engineering language for AI systems. Its significance lies less in the finality of version 0.1 than in the representational agenda it articulates: a shared symbolic language for complex AI systems that integrates components, data transformations, datasets, gold standards, and evaluation into a single coherent schematic form.