How-to-use Guide

A quick guide to engines, methodologies, and formality.

First and foremost, Lingonect^® is not a volume translation tool!

It is for words, phrases, sentences, and short paragraphs. Avoid excessive or consecutive copying and pasting of text, as this may cause the system to stop working.

NMTs vs. LLMs

When using Lingonect^®, it is important to understand the difference between NMTs (Neural Machine Translation) and LLMs (Large Language Models). While all engines support translation, more complex methodologies, such as localization, transcreation, culturalization, analysis, correction, contextualization, and triangulation, are supported only by LLMs. Please note that if you select these methodologies for an NMT engine, you will not receive a response.

Consider each LLM as one single opinion. Every LLM has its own datasets and its own learning algorithm, so there isn't necessarily a single best one; it often depends on what you like most or understand best. Never trust only one LLM! For Western languages and cultures, models like ChatGPT, Claude, Gemini, or Mistral tend to shine; for Eastern European, something like Yandex or Alice may serve you better; and for Asian, Chinese in particular, Baidu, Alibaba, or Tencent. You are the human who decides what you need, what you like most, and what fits your purpose best. Lingonect^® recommends to always compare at least 3-4 different outputs.

NMTs always use bilingual corpora and are designed to translate by following the source text as closely as possible. While translation remains subjective, particularly in longer or more complex content pieces, translation allows for less creative interpretation than the other three methodologies.

Translation

Sticks as close as possible to the source. In some cases it has to select an entirely different expression or idiom, as it does not exist in the target, but otherwise it must be a direct reflection of the source text.

Translation + Triangulation

Every NMT and LLM translates a little differently, so with many engines on screen it can be hard to see which rendering to trust. Select Translate and then press Triangulate: Translate stays bright green and Triangulate turns a darker shade of green, just like Localize when it is combined. You can do this before you translate, or afterwards, once all the translations are in.

Once every engine has answered, all the LLM and NMT translations are compared and condensed into one clean summary above the engine results: a single recommended translation (with a copy button), built on what the engines agree on, followed by a few notes on where the translations differ and which choice was made and why, with the key terms highlighted. Underneath you can see how many engines were compared, how many distinct versions they produced and how many engines shared the most common one. The individual engine results stay on screen to compare and rate as usual.

The notes are written in your target language by default. If you would rather read them in the language you wrote in, use the Notes in switch at the top of the summary to change them to your source language (and back). Only the notes change: the recommended translation, and the winners and losers, stay exactly the same. Your choice is remembered, so your next summaries arrive in that language straight away.

Triangulation also shows you the winners and losers. Every engine's translation is scored from 0 to 100% by how closely it matches the recommended translation, character by character, ignoring punctuation and capitals. The engines that came closest are listed in green and the ones furthest from it in red, and the same badge (with its score) appears next to the engine's name in the results, with a green or red bar along the card. You can also open the full list to see how every engine scored. Losers are only named when the scores really differ, so engines that all said nearly the same thing are not branded losers. Keep in mind that the score measures agreement with the consensus, not quality: a freer but perfectly valid translation can still score low.

Every win is recorded. The Ratings page has a Wins column showing how many times each engine came closest to the recommended translation, out of how many triangulations it took part in, and each language pair shows that engine's wins for the pair under its star rating. Only the language pair and which engines won, lost or took part are stored: never your text or the translations. Wins are counted separately and never change the star ratings, which always stay your own choice.

This is a lightweight triangulation, built not to waste requests. It re-uses the translations already on screen: no engine is asked to translate again. Engines whose API implementation is still in progress, engines that failed or timed out, and NMTs that could not take part (for example when localization or a formality level is selected) are left out. The winner and loser scores are worked out in your browser and cost nothing. Identical translations are counted once, so if every engine agrees, no extra request is made at all; otherwise the comparison uses one single extra request, and turning Triangulate off and on again for the same results costs nothing. Switching the notes language uses one small request the first time and nothing after that. It works with Localize too: the recommended translation then respects your maximum character count. Like the full question-answering Triangulation below, it is available to full members only.

Localization

Localization involves adjusting translated content to fit specific formats, platforms, or media requirements. For example, when translating Google Ads copy from English to German, character limitations may require you to condense the text, as German often requires more space than English. Similarly, website localization may require adapting measurements, currencies, and formats to suit the target culture.

On Lingonect^®, the localization methodology provides a translation that is automatically adjusted to your desired length. Simply enter the maximum character count for your translation, and the tool will localize the text to fit those constraints.

After selecting localization with a character max limit, you can select translation, transcreation, or culturalization. Localization cannot be combined with analysis. If you select only localization, the system will automatically use translation as the primary methodology.

Transcreation

Transcreation captures the intent, style, and meaning of the source, then recreates it in the target. It is possible that the target text is quite different from the source, so that it best reflects the source intent.

Culturalization

Culturalisation starts from the target culture, not the source. It reimagines the text to fit the linguistic and cultural reality of the target market, drawing on its references, idioms, sensitivities, and conventions so the message belongs there rather than being adapted into it.

Analysis

Analysis does not translate. Instead it takes your source word, phrase, or sentence and explains its syntax, grammar, style, and meaning, written in the target language you have selected, so a learner can understand the source text. Because it produces an explanation rather than a translation, localization cannot be combined with analysis and the formality level has no effect.

Correction

Correction does not translate either. It proofreads your source-language input and fixes any objective grammar, syntax, spelling, punctuation, or style mistakes, returning the corrected text in the same language you wrote it in. Below a grey dotted line, it then gives a short, succinct explanation of what was corrected and why, written in the target language you have selected, just like the transliteration reading aid. Correction is an LLM-only methodology, so NMT engines will not return a result.

Unlike analysis, the formality level does apply to correction. If you leave it on Original, the corrected text keeps the formality of your input. If you pick a specific level, correction will not only fix the objective mistakes but also adjust the text to that register, for example correcting a casual note into polite, formal phrasing. Because the corrected text stays in your source language, honorific and humble become available based on your source language rather than the target.

Correction also has its own optional maximum character limit, just like localization. When you select correction, a Max characters box appears: enter a number and the corrected text will be rewritten to fit that limit; leave it blank for no limit.

Contextualization

Contextualization is a deep dive on a single word. Unlike the other methodologies, it accepts only one word at a time, written in your source language. In return you get a compact learner's entry, with all explanations written in the target language you have selected:

A concise definition and the nearest natural equivalent in the target language. If a word has more than one grammatical function — for example English need, which is both a noun and a verb — each function is defined separately so nothing is missed.
Important collocations and idioms that use the word.
Strong, relevant example sentences in the source language — each with an audio button to hear it spoken — alongside a backtranslation in the target language.
Synonyms — and, where a synonym is genuinely close, a note on how it subtly differs in nuance, register or use — plus the word's antonyms.
Full conjugations for any verb function and full declensions for any noun/adjective function the language inflects. A word that is both a noun and a verb shows both sets of tables.
Any other grammar points worth knowing that are specific to that word in the source language.

Key notes are colour-coded so the most important things stand out: green for must-know points, blue for key points, and red for important things to watch out for. Contextualization is an LLM-only methodology, and because it explains rather than translates, localization and the formality level do not apply.

Contextualization never relies on a single opinion. Every available LLM is asked for its own entry, and the best of all of them is then merged into the one entry you see — the correct forms, the widest set of genuinely useful collocations, idioms and examples, and the most natural target-language explanations, whichever model supplied them. This matters because no single model is best at every language: a model native to your target language usually writes the most natural explanations, while another may know the rarer forms or idioms. The LLMs that contributed are named underneath the entry, so you can always see where it came from. Because it asks every LLM, contextualization uses one request per LLM plus one for the merge, so it counts more against your daily request limit than a single translation, and takes longer to come back.

Triangulation

Triangulation is for when you want an answer, not a translation, and do not want to trust a single AI for it. It accepts only a question, for example “Why did the Roman Empire fall?” or “Is coffee bad for your heart?”. Write the question in any language: that is your source language. The answer comes back in the target language you have selected. If your input is not a question, you will be asked to rephrase it.

Instead of showing one answer per engine, triangulation works out what the LLMs agree on, in four steps:

Ask everyone. Every available LLM answers your question independently.
Break it down and cross-check. Each answer is split into atomic claims (single, checkable statements), and the claims are compared across all the models. Each claim is marked as consensus (most models agree), partial agreement (some agree, none object), contested (models contradict each other) or single model (only one model says it).
Let the models argue it out. The claims that are not consensus go back to every model, which critiques them: support, refute or unsure. A claim most models support is promoted to consensus; one most models refute is marked rejected; the rest stay open.
Write the answer. A single summary is written from the consensus claims. Disagreements are not thrown away: they are reported as hedged alternatives (“some sources hold…, others…”), followed by what is worth verifying.

As elsewhere, the result is colour-coded: green for consensus, blue for partial agreement, and red for disputed points and alternatives. Underneath you can open the full list of claims to see their status and exactly which LLMs backed or contradicted each one.

Two outside sources help you check the answer yourself:

Wikipedia. Triangulation looks your question up on Wikipedia, in your source language and in English, and checks every claim against the articles it finds. A claim the article confirms gets a green Wikipedia ✓ tag; one it contradicts gets a red Wikipedia ✗ tag. Each tag links to the article. The tags are for reference only: they never change a claim's status, the concordance score or the written answer. No tag simply means the article did not cover that claim, not that the claim is wrong.
Perplexity's web sources. Perplexity is the one LLM that searches the web while it answers. Claims it backed are marked “web”, and the pages it used are listed under the answer so you can open them. Remember that Perplexity is still an LLM: it counts as one opinion among the others, and its sources show where it looked, not that it read them correctly.

Every answer also gets an inter-model concordance score from 0 to 100%: how far the models agree at the level of individual claims. Each claim counts by the share of models that back it, and claims more models took a position on count more. 75% or more means a robust cross-model consensus; 50 to 75% means the core is agreed but the details vary; below 50% means the claims need checking.

High concordance is not proof. The LLMs share similar architectures and much of the same training data, so they are semi-independent sources, not fully independent ones, and can be wrong together. Read the concordance score as a measure of how consistently the models agree, not as a guarantee that the answer is true. For anything important, verify contested and low-concordance claims against an authoritative source. Triangulation is available to full members only; it is not part of the free plan. It is an LLM-only methodology; localization and formality do not apply. Because every LLM answers and then critiques, it uses roughly two requests per LLM plus a few for the analysis, so it counts more against your daily request limit and takes longer than a translation.

Selecting a methodology

When selected, the chosen methodology icon will turn bright green. If localization is selected and then another methodology is added, localization will turn a darker shade of green. Likewise, if you press Triangulate while Translate is selected, Triangulate turns darker green and summarises the translations (see Translation + Triangulation above); pressed on its own, Triangulate answers a question instead.

Formality levels

You can also select a formality level. Some formality levels are only relevant to certain languages:

Honorific and humble is only relevant to languages like Japanese, Korean, Thai, and Vietnamese. Original means that the target text reflects, as closely as possible, the source text's formality level. Ceremonial is reserved for very formal, official occasions. The rest of the formality levels should be self-explanatory.

← Back to engines