Tools map: digital humanities and computational social science

On this page, and how to read the tables

Two research fields have grown over the past twenty years around the same fact – that texts, relationships, places and behaviors now exist in digital form and can be computed. The digital humanities (digital humanities, DH) came from philology, history and literary studies: their question is what can be known about a text, an archive or a period when they are read with the help of a computer (Schreibman, Siemens & Unsworth, 2016). Computational social science (computational social science, CSS) came from sociology, political science and psychology: its question is what can be known about a society from the digital traces it leaves (Lazer et al., 2009, 2020; Edelmann et al., 2020). The tools mostly overlap, and this map presents them together – by what they do, not by the field that invented them.

The definitions of the field itself are many, and Carter (2013, p. ix-xi) surveys them candidly: "an intersection between the humanities and information technologies" (Svensson, 2010, as cited in Carter), "a nexus of fields" in which traditional humanities questions are asked about technologies and vice versa (Fitzpatrick, 2010, as cited in Carter), and two "waves" – quantitative-corpus-based and then interpretive-experimental (Schnapp & Presner, 2009, as cited in Carter). His practical conclusion is useful for this map: there is no need to understand the code in order to use it, and whoever works in the field today suffers from "almost too many options". Drucker and Albrezzi (2026, pp. 1-2) offer the simplest formula for any project – materials + processing + presentation + participation – the fourth term is new in the second edition: stakeholders whose material or cultural property is involved are active partners in decisions about access, use, description and presentation, and the professionals and students on the team are entitled to recognition and credit – and seven basic operations, each of which is a decision and not a technical step: remediation (conversion to digital), "datafication" and modeling, processing and analysis, presentation, rendering and embedding, sustainability and preservation, and ethical considerations; "technological instruments are never value-neutral" (p. 2). And their central claim is the reason every category in the map opens with "the principle before the tool": "tools and platforms change, but the principles of digital work have become clear and settled" (p. 2) – whoever knows how to build an edge list will move it to Gephi, to Cytoscape or to whatever comes after them. Dawson (2020, p. 1) extends the framework from the humanities to the social sciences: "digital" is an umbrella term for methods that collect and analyze data by computerized means – both new methods (machine learning, big data analytics) and established ones that have been adapted (questionnaire, interview, network analysis) – and every entry in her book has a section on "ethics, morals and law", from which the ethical notes in the categories below are taken.

And from the side of computational social science – two textbooks. Cioffi-Revilla (2017, p. 2) defines the field as "the interdisciplinary investigation of the social universe on many scales, from the individual actor to the largest groupings, through the medium of computation", and warns that it is not reduced to big data, network analysis or simulation – these are four areas that feed one another: automated information extraction, social networks, social complexity and social simulation (pp. 12-17). It is, in his words, an "instrument-enabled science" like astronomy after the telescope – "instruments are the main tools with which science creates new science" (pp. 3-4) – and this is why a tools map is a methodological matter and not a technical one. Salganik (2018, pp. 6-8) provides the pair that organizes the collection categories: "readymade" data (readymades – traces that companies and governments created for other purposes, which the researcher repurposes, like Duchamp) versus "custommade" data (custommades – a questionnaire or experiment created for the question, like Michelangelo), and the most interesting research is "part Duchamp and part Michelangelo". And like Drucker and Albrezzi, he too writes about principles and not platforms: "instead of a platform-by-platform approach", ten general characteristics of big data sources, which will also apply to "data sources that will be created in the future" (p. 17) – they are in category 7.

And why does this belong on a statistics site? Because statistics is the layer beneath both fields: topic modeling is factor analysis on words, a centrality measure is descriptive statistics on ties, and a simulation is tested against data with the familiar tools. Whoever can read a correlation matrix can read an adjacency matrix.

How to read the tables

Each table presents the tools in one category in five columns. License and cost in three chips: Open source – the code is open, free, can be installed and modified; Free (closed) or Free + paid – free to use but the code is closed, sometimes with paid plans or limitations; Paid – a paid license, usually with an academic price or a campus license. Interface: "Installed software" – a graphical interface on the computer; "Browser" – works on a website, with no installation; "Code" – a library or command line, requires R, Python or Java.

The links, licenses and development status of all 99 tools were checked against the tools' websites in September 2026. A tool that is no longer developed is marked in the notes; a tool that was omitted – because it could not be verified, not because it is not good.

1. Text analysis and text mining

The largest category in both fields, and also the place where the difference between them is noticeable. In the digital humanities they speak of "distant reading" (distant reading; Moretti, 2013; Jockers, 2013): word counts, concordance and topic modeling on a corpus too large to read. In computational social science they speak of "text as data" (text as data; Grimmer, Roberts & Stewart, 2022): the text is a variable, and everything else – measurement, reliability and validity – is as in any quantitative research. The tools overlap; the difference is in the question.

The principle before the tool. Drucker (2021, Section 7a): the decisions about what can be measured (parameterized), counted, sorted and displayed "are interpretive acts that shape the results of research"; in the second edition this is called parameterization – some features become a metric easily (number of words, length of a poem), and many others are "entirely legible to human perception" and yet hard to parameterize, and what is included in the data and how it is formatted also requires "critical attention" (Drucker & Albrezzi, 2026, pp. 14, 49). And the language of the tools hides this: topic modeling is described as "extracting" topics from the text, "as if the topics were already in the text rather than created as an act of reading and interpretation" (p. 51). A small, sharp example from the first edition: most tools remove "stop words" such as definite articles, but the difference between "the God" and "god" is all the difference (Drucker, 2021, Section 7a). And the reproducibility problem: repeated runs of text mining and cultural analytics "return results that have shifted and changed", because the corpus is not stable and because the processing is probabilistic (p. 54) – so the random seed, version and parameters are part of the finding. From the computational social science side, Cioffi-Revilla (2017, pp. 13-14) calls this category "automated information extraction" and sees it as a "foundational methodology" of the field, because it feeds everything else; and in his data mining process (pp. 115-117) the question comes before the tool – on a continuum between a data-driven (inductive) question and a theory-driven (deductive) question – and "the formulation of research questions should frame any data mining investigation, because the research questions condition each of the subsequent stages"; and before the analysis – scanning (OCR), cleaning, filtering, segmenting into units (the next category). Ethics and law (Dawson, 2020, pp. 79-85): "discrimination-aware" data mining – Berendt and Preibusch (2017, as cited in Dawson) show how discrimination arises from the "combined influence of human and machine inference"; and in the European Union and the United Kingdom there is a "database right" over collections that required substantial investment to gather, and computational analysis of protected material is permitted there only for non-commercial purposes – whoever mines a protected corpus must check the local law.

Hebrew has a structural problem: most tools were developed for English and assume that a word = a sequence of letters between spaces, whereas in Hebrew "ukhshehalakhnu" ("and when we went") is four units. A tool that declares "Hebrew support" usually means tokenization only; real morphological analysis requires the tools in Table 3.

2. Text encoding, OCR and digital archives

Before analyzing text, there has to be text. Three stages precede the analysis: converting a scan or manuscript to text (OCR for print, HTR for handwriting), encoding the structure and interpretation in an open standard (TEI – the standard of digital scholarly editions), and organizing the archival material itself. The FAIR rule (Wilkinson et al., 2016) – data that are findable, accessible, interoperable and reusable – starts here: an open, exportable format is preferable to a pretty tool that locks the material in.

The principle before the tool. Drucker and Albrezzi (2026, p. 15) insist on a word: data are not "data" ("what is given") but "taken" – capta – "information captured through a process of abstraction and structuring that depends on decisions that are already acts of interpretation"; so the data model – what is counted, what is a unit, what is discarded – must be documented like any methodological decision (p. 18: "data are never neutral, but always constructed from a particular point of view"). And on markup languages (p. 35): their advantage is that they remove ambiguity – and that is also their drawback, because "computers do not tolerate ambiguity" whereas most humanistic material is full of it; encoding in TEI a metaphor that is also a literal statement means choosing, and they ask whether this poses an ethical problem and answer: "absolutely". File formats and naming conventions are set at the planning stage, with the institutional partners who will preserve the material (p. 27). Ethics and law: legacy data – data created by someone else, sometimes without documentation, and the decisions about what was collected, cleaned and omitted "may not be recoverable"; whose data they are, what permissions were required to create them, and what about privacy – reusing an old catalog or database does not exempt you from these questions (pp. 15-16, 42).

3. Hebrew, Judaica and classical languages

The category in which "the world's leading tool" is not necessarily the right tool. Hebrew – biblical, rabbinic and modern – is a morphologically rich language with unvocalized spelling, and any serious analysis starts with morphological analysis that general tools do not do. Alongside the processing tools – the databases: the classical Jewish text is today encoded almost in full and open to an API, and the historical Hebrew press is fully searchable. At the end of the table – three resources for Sanskrit and the classical languages, for the researcher working comparatively.

The principle before the tool. Here Drucker and Albrezzi's principle on ambiguity (2026, p. 35) works doubly: unvocalized spelling is a built-in ambiguity in every token, and morphological analysis is a decision the tool makes for you. And for the annotated databases – whoever relies on BHSA inherits the ETCBC linguistic model; what exactly, in section 6 of the page Hebrew NLP.

4. Network analysis

social network analysis is the oldest branch of computational social science – 20th-century sociometry that gained computing power. Its question differs from the usual statistical question: not "what characterizes the individual" but "what characterizes the ties." In the humanities the same tools are used for correspondence networks, characters in a novel and citations between texts.

The practical choice: Gephi for visualization and exploration, a code library (igraph or NetworkX) for measures and statistics, and UCINET if your guide is Borgatti. For bibliometric networks – VOSviewer.

The principle before the tool. The data model of a network is a three-term formula – entity-relation-entity – that can be built in a spreadsheet and exported to any platform; the real work is to specify the type of relation precisely (not one "related to" for everything) and change over time (Drucker & Albrezzi, 2026, p. 70). Cioffi-Revilla (2017, p. 147) sharpens the distinction: "graphs are to networks what decision-theoretic models are to decision-making" – the graph is the representation, not the network; a network is a finite set of entities (nodes) and a set of relations (edges) defined on it, and these two definitions are the work. And on software (pp. 168-169): UCINET is "one of the most widely used software packages," and networks of hundreds and thousands of nodes have become feasible only in recent years – computation time and tractability are a consideration before choosing the measure. Dawson (2020, pp. 356-361) adds the validity questions: how to justify the network boundary; what missing information on people and ties does to bias the measures; and whether visualization is the right way to interpret structural properties – "can a visual representation display them in a misleading way?" (the answer: yes, which is why the measures come from igraph and not from the eye). Ethics and law (ibid.): whoever identifies their own network has given consent – the others in the network have not; anonymity can be broken by network position alone; and analysis of an organizational network can lead to dismissals or harm – "could your analysis cause harm to individuals?" is a question asked before collection.

5. GIS and spatial-historical mapping

Geographic information systems (GIS) answer the "where" question: periphery versus center, distribution of responses by statistical area, a historical map laid over a contemporary map. For the social researcher in Israel two free resources are especially important: the statistical-area layers of the Israel Central Bureau of Statistics (CBS) – the unit in which census and welfare data are published – and GovMap, the national mapping portal. Joining a survey file with a statistical-area layer is the step that moves an analysis from "by locality" to "by neighborhood."

The main tool is QGIS – a complete free alternative to ArcGIS. For spatial-statistical analysis (spatial autocorrelation, spatial regression) – GeoDa.

The principle before the tool. "All mapping systems are based on projections that translate a three-dimensional shape into a two-dimensional representation and contain distortions," and the assumptions built into the representation should be visible in the argument the map serves (Drucker & Albrezzi, 2026, p. 77); "the positivist foundations of GIS" make it hard for it to grasp dwelling and experience, "even when its apparent objectivity makes it a powerful presentation of arguments" (ibid.). And in the summary of chapter 8b (p. 82) – the limit: maps are excellent at locating data in relation to geography, and "less effective at presenting the qualitative experience of space" – a map of statistical areas says where, not what it feels like to live there. Ethics and law (Dawson, 2020, pp. 140-146): location tracking, drones and high-resolution satellite imagery are an intrusion on privacy that requires an ethical code (Dawson refers to that of GISCI); and participatory (crowdsourced) mapping can be used for marginalization by those who hold power – the question "could the analysis be used by interested parties to harm a group?" is asked in advance.

6. Agent-based models and social simulation

agent-based modeling reverses the direction of inquiry: instead of inferring a rule from data, simple rules are defined for agents and the researcher checks which macro-social phenomenon emerges from them – urban segregation from mild preferences (Schelling), consensus from mutual influence, a norm from repeated cooperation. It is a tool for theory more than for empirics, and that is precisely why it matters to the researcher who wants to test whether her explanation can produce the phenomenon at all.

NetLogo is the entry point, without competition – "low threshold, high ceiling" – with a built-in library of classic models. Those who write Python will go to Mesa; models with GIS – GAMA. The CoMSES model library is the place to look for an existing model before writing a new one.

The principle before the tool. Dawson (2020, pp. 4-9) makes the words precise: a model represents, a simulation uses the model – and the method builds "bottom up, from micro to macro," so that patterns emerge from the interaction of agents. Two of her questions require an answer in the methodology chapter: "Do the models represent the world, or the researcher's interpretation of the world?" and how to check – verification (the model works correctly) versus validation (the right model was built). Cioffi-Revilla (2017, pp. 384-390) spreads this into stages: the research questions come first and determine the model's level of fidelity – from theoretical abstraction to full empirical calibration (p. 385) – and the fidelity scale is "not related at all" to the model's scientific value (pp. 382-383); verification is the counterpart of internal validity – "a model that has not been verified is not fit for analysis" – and is done by reading the code line by line, profiling the frequency of function calls and parameter sweeps that expose anomalies (pp. 388-389); validation is pattern matching between the simulation output and empirical data – histograms, moments, time series, indices such as Gini – and "always involves matching simulated data to empirical data" (pp. 389-390). And two warnings: "simulation code is highly perishable, much more so than mathematical or statistical models" – famous simulations were lost shortly after they were written, and only the conceptual model remains of them (p. 388); and verification and validation "are not sufficient in themselves" for judging quality – alongside them stand the three criteria of Lave and March (1993, as cited in Cioffi-Revilla): truth, beauty (parsimony and form) and justice – the model's contribution to a better world (pp. 391-393). Before choosing a tool: Abar et al. (2017, as cited in Dawson) characterize 85 ABM toolkits – NetLogo is the entry point, not the only one. Ethics and law (ibid.): a model can do harm – when individuals are flagged for action based on a prediction, when the prediction rests on past behavior that has changed, and when there is no way to correct input; and whether the data were given for the purpose of modeling at all.

7. Digital data collection

Two families: the online survey – the veteran tool with a new interface – and the collection of "digital traces" (digital traces; Salganik, 2018): scraping websites, access to platform APIs, and capturing content from social networks. The first raises a question of coding and validity; the second raises, before anything else, a question of ethics and legality – the platform's terms of service, users' privacy, and ethics committee approval. Salganik (2018, chapter 6, pp. 281-354) is required reading.

For survey research in Israel – most universities hold a Qualtrics license; LimeSurvey is the open alternative when there is none. The output of all of them – an Excel file in the site's standard data template, or an SPSS file that the From SPSS to Excel page converts.

The principle before the tool. Salganik (2018, pp. 17-41) lists ten characteristics of big data sources and groups them roughly: three "generally helpful" – big, always-on, nonreactive (measurement does not change behavior) – and seven "generally problematic": incomplete ("no matter how big the data, they probably don't have the information you want"), inaccessible, nonrepresentative ("bad for out-of-sample generalizations, but useful for within-sample comparisons," p. 29), drifting (drift in population, usage and system makes long-term trends hard), algorithmically confounded ("behavior in big data systems is not natural; it is driven by the engineering goals of the systems," p. 35), dirty ("loaded with junk and spam," p. 37) and sensitive. And an overarching rule: "large datasets are a means to an end; they are not an end in themselves" (p. 17). And on the survey: "we will always need to ask people questions" – internal states such as emotion, knowledge and expectation are not in behavioral data, even perfect ones (pp. 87-88); and probability sampling is becoming hard in practice while non-probability sampling opens new possibilities – provided they are corrected (p. 99). For social networks (Dawson, 2020, pp. 349-355): the data are "unsolicited and unscripted, and often multimodal" (Brooker et al., 2016, as cited in Dawson) – text, link, image, video – so the method of analysis (content, discourse, sentiment, network, time analysis) is chosen according to the question, not according to what the tool prints; and the platforms' API policies are "increasingly restrictive," so data obtained through them "cannot be reproduced exactly because of the unknown logic of the algorithm that produced them" (Felt, 2016, as cited in Dawson) – a limitation to state explicitly. For the online survey (pp. 288-294): a unique link for each respondent prevents double completion and allows reminders; for sensitive topics – an indirect question ("how would other people respond") is sometimes better than a direct one. For webometrics (pp. 398-403): "the coverage of the web is partial" and biased in favor of countries that adopted it early (Thelwall, 2008, as cited in Dawson), and country versions of search engines favor local sites. Ethics and law: "is an online post public or private data?"; informed consent when participants do not know they are in a study; anonymity when the platform requires attribution to the author (pp. 349-355); IP address, username and referrer in log files are "personal data" under the GDPR, and storing them requires consent for a defined purpose, security and deletion when they are no longer needed (pp. 188-193) – and in Israel, the Privacy Protection Law and its regulations (what the law requires of a researcher who collects personal information, after Amendment 13 – on the Research ethics page, section 11). Drucker and Albrezzi (2026, p. 49) add for the age of big data: "the fact that data are available does not mean it is ethically permissible to use them without safeguards," because whoever posted did not necessarily intend scientific analysis; real anonymization is "hard to achieve," because people are re-identified from other attributes or by cross-matching databases; hence data minimization – collect only what the analysis requires: general sentiment does not need an exact location and a full profile. And Salganik (2018) on what has changed: the power of researchers – "the ability to do things to people without their consent or even awareness" – has grown, and the rules have not become clearer to the same extent (p. 288); the four principles of the Belmont and Menlo reports – respect for persons, beneficence (do no harm, maximize benefits and minimize harms), justice in distributing burdens and benefits, and respect for law and public interest (pp. 294-299); the practical rule "some form of consent for most research" (p. 303); informational risk – harm from disclosure of information – is "the most common risk in social research, has increased dramatically, and is the hardest to understand" (p. 307); and three pieces of advice: the ethics committee is "a floor, not a ceiling"; put yourselves in everyone else's shoes; and think of ethics as continuous rather than dichotomous (pp. 321-324).

8. Statistics and machine learning

The category this site itself belongs to, hence the important note: for every test that has a calculator here – the calculator page is the tool, in the browser and without the data leaving the computer (see the links at the end). This table is for what lies beyond: multilevel models, machine learning, full SEM, and for those who want a complete software package. The central distinction is between statistics (explanation: which variable has an effect and how much) and machine learning (prediction: what will happen to a new observation) – two different goals that require different validation, even when the model is identical. This distinction, the end-to-end pipeline, the models – from a decision tree to a neural network – and their tools are on the Machine learning for the social researcher page; the metrics, cross-validation, data leakage and non-independent data – in Performance metrics and validation; factor, component and cluster analysis – learning without a dependent variable – in Unsupervised learning; Table 8 here remains the table of statistical software.

The pair JASP / jamovi deserves special attention: both are a free SPSS replacement with a graphical interface, based on R, and also compute Bayesian statistics. PSPP is the reference software against which parts of this site are verified – free, and it runs SPSS syntax.

The principle before the tool. Dawson (2020, pp. 194-199) defines machine learning as algorithms that "learn from previous computations to make repeatable decisions and predictions without explicit rule-based programming," supervised (labeled input and output) or unsupervised – and that is exactly the difference from the calculators on this site, where the rule is visible. Drucker and Albrezzi (2026, p. 1) call processing "the black box of digital work," because most of it is done in ready-made tools whose operation is hidden, and warn that the more sophisticated the processes and the more they involve machine learning and artificial intelligence, "they often reinforce biases that are built into their model" (p. 2). Dawson also asks about cluster analysis (pp. 31-37): "can the software, the tools and the algorithms be trusted?" – and suggests checking with additional software, which is exactly this site's cross-checking principle (PSPP, R, Python on the same data). And what big data do not provide by themselves – causality: Salganik (2018, pp. 50-61) shows how to "approximate experiments that cannot be run" in two ways that the digital age has expanded, natural experiments and matching – and both begin with assumptions that must be written down, not with a tool. And for large language models, to which the second edition of Drucker and Albrezzi devotes a chapter (12a), two axes for choosing a model: open or closed, and local or remote. A closed model run on a company's server means that "your data are captured and processed on proprietary servers," and the alternative – an open model downloaded and run locally (for example in Ollama) – is for those who do not want to share the material; and a RAG-type retrieval system, which answers only from a repository the researcher has defined, is more accurate and suits sensitive or protected data (pp. 118-119). And hallucinations still happen: every output, even when it looks accurate, "should be investigated, checked and compared against known reliable sources" (p. 119). Ethics and law (Dawson, 2020, pp. 194-199): "is machine learning ethically neutral?" – when a prediction model gives rise to social decisions, fairness, accountability and transparency are required (Veale & Binns, 2017, as cited in Dawson). Drucker and Albrezzi (2026, pp. 120-121) add: "any process in which human beings are counted, sorted or considered as data risks the loss of their subjective humanity"; the provenance of models is "very opaque" – the training sets are proprietary, and unless it is a RAG system there is no knowing which sources the output relies on and how they were weighted; and under the terms of use of many tools the user agrees that their prompts and outputs will be used for further training and even passed to partners – which must not be done with participants' data.

9. Data visualization

Three levels: "no-code" tools in the browser (Datawrapper, Flourish, RAWGraphs) for a quick, publication-worthy chart; a grammar of graphics in code (ggplot2, Vega-Lite, Plotly) for a chart that can be reproduced and updated automatically; and D3.js for custom visualization that has no template. For analytical-journalistic writing – Datawrapper is the newsroom tool, and it is the only one here that explicitly declares full support for RTL and Hebrew.

For an academic figure according to APA (sections 7.22-7.28) – the site's calculator charts are already built that way; for R see the "And how do you do it in R" section on the Scatter plot page, including the solution for Hebrew punctuation (the AGG engine).

The principle before the tool. The feeling that data have a built-in visual form "is an illusion" – any series can be displayed in dozens of formats, and only some of them suit the properties of the data (Drucker & Albrezzi, 2026, p. 57); "all visualizations are interpretations, not presentations of fact" (p. 67). The common mistakes they list (pp. 60-66): misuse of area (when the value doubles and the area grows fourfold – their figure 7.5), artificial continuity (a line connecting discrete points that are not a series – figure 7.4a), and a scale stretched by a single outlier that erases the differences in the middle (figure 7.7). Dawson (2020, pp. 86-92) adds the biases that are not graphical: narrative bias (choosing the story being illustrated), visualization that supports an ideology, and exclusionary language. Ethics and law (ibid.): attribution and marking of the data sources in the chart itself, and transparency about the construction steps – "how will you ensure that the visualization does not lead to wrong conclusions?"

10. Reference management, open science and collaboration

The tools that analyze nothing yet without which research cannot be reproduced: bibliography management, preregistration, version control, collaborative writing. Two principles of open science are directly relevant to this site: a model registered in advance is reported differently from a model changed after the fact (see the path analysis page), and documented code and versions – just as every page on the site carries a version number.

The principle before the tool. Managing a digital project according to Drucker and Albrezzi (2026, pp. 126-127): "start with something that can be done"; build with room for expansion but without dependence on constant maintenance; document every workflow step by step, because without a written record "even repeating your own work can be a challenge"; open-source tools with a large user community "are likely to last longer than proprietary formats"; and plan the project's "life cycle" in advance – "not every digital project will last," and "the time to think about how to export the intellectual content is at the beginning." The first edition put it more sharply: "project documentation may be the best contribution" (Drucker, 2021, Section 12a). And on platforms (p. 110): "always check the export function" – content built inside presentation software that does not export is lost work; standards for data sharing (LoD, IIIF, CIDOC-CRM) are efficient, but expensive to implement and therefore favor wealthy institutions (p. 114). Ethics and law (pp. 128-129): the digital makes copying easy, so an understanding of copyright, fair use and Creative Commons licenses is required; "citation is an essential aspect of scholarship and should be practiced in all circumstances"; linking does not require permission, and material can be displayed in a window without storing it on your server; and the golden rule – "treat the work of others as you would want your own work to be treated."

Common mistakes

Choosing a tool before a question. The first and most common mistake: "I have a corpus, let's do topic modeling." Topic modeling answers a particular question (which clusters of words recur together) and not others (who said what to whom). First the question, then the type of data that answers it, and only then the tool – this is also the order of the columns in the checklist below.

Sending sensitive data to a server without noticing. An interview pasted into Voyant, a survey file uploaded to Flourish – are no longer on your computer. The ethics committee approval you received almost certainly said something else.

Trusting the tokenization of Hebrew. A tool that treats "ukhshehalakhnu" ("and when we went") as one word will silently produce wrong statistics – the frequencies of "halakh" ("went") will vanish into thousands of forms. For Hebrew – morphological segmentation comes before any counting (Table 3).

Building a project on a tool that is not being developed. Palladio, YAP, GRETIL – they work, but the last update was years ago. Such a tool is good for exploration; for a two-year project – make sure you can export to an open format (CSV, GraphML, TEI) so the work outlives the tool.

Scraping without asking. The platform's terms of service, users' privacy and local law come before the first line of code. Salganik (2018, pp. 294-299) offers a framework of four principles – respect for persons, beneficence, justice, respect for law and public interest – and explains why informed consent here differs from a laboratory experiment (pp. 303-306).

Being impressed by size. A million records from one platform are a nonrepresentative sample of the platform's users, whose behavior is "driven by the engineering goals of the system" – and not of the population; "large datasets are a means to an end; they are not an end in themselves" (Salganik, 2018, pp. 17, 29, 35). The question is not how many, but who, and what the system did to them.

Simulation without verification and validation. A model that runs and produces a chart is not a finding: first verification – that the code does what the conceptual model says ("a model that has not been verified is not fit for analysis") – and then validation against empirical data; and both are still not sufficient for judging quality (Cioffi-Revilla, 2017, pp. 388-393). And simulation code is "perishable" – without documentation and a repository (CoMSES) it disappears.

Not documenting the version and the parameters. Topic modeling with k = 20 and k = 40 gives two different studies; Gephi with a different layout gives two different pictures. A computational result that is not documented down to the level of version and parameter cannot be reproduced – and is therefore not a finding.

Confusing prediction with explanation. A machine learning model that predicts 92% correctly explains nothing about the cause, and a significant regression coefficient does not guarantee good prediction. Two goals, two tests of validity.

Calling data "data." What enters the tool was taken from the material through decisions – what is a unit, what is counted, what is discarded – and Drucker and Albrezzi (2026, p. 15) propose calling it capta. The mistake is not in the word but in what it hides: an undocumented data model is an unreported interpretation.

Trusting area and line. A circle whose value is double and whose area is four times as large; a line connecting categories as if they were a series; a scale stretched by one outlier. Drucker and Albrezzi (2026, p. 67): "all visualizations are interpretations, not presentations of fact" – and some graphical properties are a product of the display, not of the data.

Citing what a language model "found." A general language model answers even when it has no source, and whatever is not in its training set looks as if it does not exist – "and sometimes, when this happens, the system tries to create the requested information"; the data provenance of most models is opaque, and unless it is a RAG system there is no knowing what the output relies on (Drucker & Albrezzi, 2026, pp. 119, 121). A model's output is a starting point for checking against a source – not a source.

Assuming the network is complete. Centrality and distance measures are computed on what was collected; one missing tie can turn a "bridge" into a "structural hole." Dawson (2020, pp. 356-361): justify the network boundary, and say what is missing.

Checklist – before starting with a new tool

And what is here, on the site

Professor Stat's calculator pages cover the statistics layer of most studies in both fields – in the browser, without the data leaving the computer, with an APA table and a report sentence in English:

The full list of calculators is in the tabs at the top of the page. The other guide pages of Part A: Hebrew NLP · Machine learning for the social researcher · Performance metrics and validation · Unsupervised learning.

References

Carter, B. W. (2013). Digital humanities: Current perspective, practices, and research (Cutting-Edge Technologies in Higher Education, Vol. 7). Emerald.

Cioffi-Revilla, C. (2017). Introduction to computational social science: Principles and applications (2nd ed.). Springer. https://doi.org/10.1007/978-3-319-50131-4

Dawson, C. (2020). A–Z of digital research methods. Routledge.

Drucker, J. (2021). The digital humanities coursebook: An introduction to digital methods for research and scholarship. Routledge.

Drucker, J., & Albrezzi, F. (2026). The digital humanities coursebook: Applied concepts and critical approaches (2nd ed.). Routledge. https://doi.org/10.4324/9781003589792

Edelmann, A., Wolff, T., Montagne, D., & Bail, C. A. (2020). Computational social science and sociology. Annual Review of Sociology, 46, 61-81. https://doi.org/10.1146/annurev-soc-121919-054621

Grimmer, J., Roberts, M. E., & Stewart, B. M. (2022). Text as data: A new framework for machine learning and the social sciences. Princeton University Press.

Jockers, M. L. (2013). Macroanalysis: Digital methods and literary history. University of Illinois Press.

Lazer, D., Pentland, A., Adamic, L., Aral, S., Barabási, A.-L., Brewer, D., Christakis, N., Contractor, N., Fowler, J., Gutmann, M., Jebara, T., King, G., Macy, M., Roy, D., & Van Alstyne, M. (2009). Computational social science. Science, 323(5915), 721-723. https://doi.org/10.1126/science.1167742

Lazer, D. M. J., Pentland, A., Watts, D. J., Aral, S., Athey, S., Contractor, N., Freelon, D., Gonzalez-Bailon, S., King, G., Margetts, H., Nelson, A., Salganik, M. J., Strohmaier, M., Vespignani, A., & Wagner, C. (2020). Computational social science: Obstacles and opportunities. Science, 369(6507), 1060-1062. https://doi.org/10.1126/science.aaz8170

Moretti, F. (2013). Distant reading. Verso.

Rieder, B. (2015). YouTube Data Tools (Version 1.x) [Computer software]. https://ytdt.digitalmethods.net

Rosseel, Y. (2012). lavaan: An R package for structural equation modeling. Journal of Statistical Software, 48(2), 1-36. https://doi.org/10.18637/jss.v048.i02

Salganik, M. J. (2018). Bit by bit: Social research in the digital age. Princeton University Press. https://www.bitbybitbook.com

Schreibman, S., Siemens, R., & Unsworth, J. (Eds.). (2016). A new companion to digital humanities. Wiley Blackwell. https://doi.org/10.1002/9781118680605

Wilkinson, M. D., Dumontier, M., Aalbersberg, I. J., Appleton, G., Axton, M., Baak, A., Blomberg, N., Boiten, J.-W., da Silva Santos, L. B., Bourne, P. E., Bouwman, J., Brookes, A. J., Clark, T., Crosas, M., Dillo, I., Dumon, O., Edmunds, S., Evelo, C. T., Finkers, R., . . . Mons, B. (2016). The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data, 3, Article 160018. https://doi.org/10.1038/sdata.2016.18