Tool AT-W46REV. 2026-A

Keyword sorter: which search terms belong on one page?

Terms that differ only in word form, order or place belong on one page. If two terms share the topic but not the intent, you need two. The keyword sorter reduces every line to its word stem, splits German compounds and reads the search intent from signal words. This produces a page plan with notes on cannibalisation, which you can further edit as a CSV.

  • free
  • no sign-up
  • sorts in your browser

Paste in a list, read off clusters and page plan

Sorter · clusters, search intent and page plan

The sorter reads wording, word stem and signal words. It cannot see what Google really shows for a term; for open and borderline cases, the search results decide.

Step 1Which search terms do you want to sort?

Separate volume and URL with a semicolon, comma or tab. Exports from Keyword Planner or Search Console can be pasted in directly; the header row is recognised. The example already filled in is invented; its numbers are made up.

Step 2, optionalService area, brands and strictness

Separate with commas. The sorter already knows larger towns and the Vogtland; enter smaller places here. A brand on its own in a term counts as navigational.

How strict should the clusters be?

Result

—

Paste in a list, one term per line.

Rules and word lists as of 14 November 2026. Check titles and descriptions for the pages with the snippet preview; if you merge pages, the redirect planner plans the redirects.

Sorted in your browser. The list is not sent or stored.

From search term to cluster

A keyword list from the planner looks tidy, but rarely is. The same question appears five times, once in the plural, once with a place, once as a compound word. Build pages from that and you end up writing three texts that compete for the same spot on Google. So the sorter tidies up the words first and only then compares.

Word stem instead of word form

Every word is reduced to its stem: “Webdesigner” and “Webdesign” become “webdesign”, “Steuerberatung” and “Steuerberater” become “steuerberat”. For this the sorter uses the Snowball stemmer for German, which goes back to Porter’s method: endings are removed step by step, depending on the length of what remains Porter 1980. For all 35,053 words of the published test vocabulary, our implementation produces the same stem as the original. That this pays off in German is well documented: even simple stemming helps retrieval, and splitting compounds helps considerably more Braschler and Ripplinger 2004, in one trial by about 25 per cent Monz and de Rijke 2002.

Splitting compounds into parts

“Zahnarztpraxis”, “Kinderzahnarzt”, “Webdesignkosten”: German glues words together, and a plain word comparison would see three unrelated terms. The sorter splits a compound into known words of at least three letters, with “s” or “es” as a joining element, and chooses the split with the highest geometric mean of frequencies Koehn and Knight 2003. “Known” means what occurs in your list, plus a basic word list. If a compound ends in a signal word such as “-kosten” or “-praxis”, that is split off: “Webdesignkosten” is web design with a question about price.

A split can go wrong because the sorter does not check meaning; this is exactly where language models are otherwise needed Riedl and Biemann 2016. That is why the whole word is always kept, and the parts count only 0.5 in the comparison. “Zahnarzt” and “Kinderzahnarzt” are therefore half alike, not identical.

„webdesigner in plauen"webdesignerCORESTEM webdesigninFILLERplauenPLACE→ LOCAL · SERVICE PAGE WITH PLACE REFERENCE„was kostet eine homepage"wasQUESTION WORDkostetSIGNALeineFILLERhomepageCORESTEM websit→ COMMERCIAL · SERVICE OR COMPARISON PAGE„zahnarztpraxis plauen"zahnarztCORESTEM zahnarztpraxisSIGNALplauenPLACE→ LOCAL · SERVICE PAGE WITH PLACE REFERENCE

Fig. 1 · Core outlined and filled, signal and question words hatched, places dotted, filler words dashed. “Homepage” stands for Website, Webseite and Homepage.

Meaning the same without the same words

“Homepage” and “Webseite” share no letter that a stemmer could find. That is typical of short texts such as search queries; a word comparison then gives zero Sahami and Heilman 2006. The sorter therefore knows a short list of equivalences, for example website = webseite = homepage = internetseite = internetauftritt = webpräsenz; onlineshop = webshop = internetshop; email = mail; smartphone = handy. What is not on the list it does not recognise; that is what the notes on close clusters are for.

Similarity and strictness

What remains of each term is the core, that is the topic without filler words, signal words and places. The sorter compares two cores word by word: each core word counts with its best match on the other side; the same stem counts as 1, a compound part only as 0.5. The sum divided by the number of all core words gives a value between 0 and 1. The most general terms become the lead term of a cluster first; every further term joins the cluster whose lead term is most similar to it, once the threshold is reached. Pure word comparisons have limits with texts this short; click data from the search engine improves them Wen et al. 2002. The sorter does not have that. So it tells you where two clusters lie close together, and there a look at the search results decides.

Strictness of the clusters
StrictnessThresholdWhat is groupedExample (similarity)Clusters in the example
tight1only the same core (by word stem, equivalence and compound)“zahnarzt plauen” and “zahnarztpraxis plauen” (1)12
medium0,6core mostly matches“webdesign” and “webdesign wordpress” (0.67)10
wide0,4a shared main topic is enough“wordpress hosting” and “wordpress theme” (0.5)9

Clusters in the example: the tool’s preset with 23 terms. Medium is the default.

What the searcher wants: five intents and an open remainder

The usual classification dates from 2002: someone searching wants to know something, reach a particular page or do something, such as buy or download Broder 2002. In its guidelines for raters, Google calls the same classes Know, Website and Do and adds Visit-in-person, a visit in person. The sorter follows this and handles two classes separately because for businesses they determine the page type. Commercial means: someone is comparing before a decision, an intent that research has studied on its own Ashkan et al. 2009. Local means: someone is looking for a provider nearby.

Search intents in the sorter
IntentThe searcher …Signal words (extract)ExamplePage type
Informationalinformational · Google: Knowwants to know or learn somethingquestion word at the start, anleitung (guide), tutorial, tipps (tips), ratgeber (guide), erklärung (explanation)“wie funktioniert seo” “wie” informational, “funktioniert” informational.Guide article
Navigationalnavigational · Google: Websitewants to reach a particular websitebrand alone, domain, login, einloggen (log in), anmelden (sign in), anmeldung (sign-in)“jimdo login” “login” navigational, “jimdo” brand only, navigational.Strengthen an existing page
Transactionaltransactional · Google: Dowants to do something: buy, book, downloadkaufen (buy), bestellen (order), buchen (book), mieten (rent), leihen (borrow), leasen (lease)“vorlage kostenlos download” “vorlage” transactional, “kostenlos” transactional, “download” transactional.Offer or product page
Commercialcommercial investigationis comparing before a decisiontest, testsieger (test winner), vergleich (comparison), vergleichen (compare), vs, versus“homepage baukasten vergleich” “vergleich” commercial.Service or comparison page
Locallocal · Google: Visit-in-personis looking for a provider nearbyplace name, postcode, nähe (nearby), umgebung (surroundings), umkreis (radius), öffnungszeiten (opening hours), anfahrt (directions)“seo agentur plauen” “agentur” commercial, “plauen” place. Place reference takes precedence over commercial.Service page with place reference
Openno signal wordcannot be read from the wordingnone“zahnarzt” No signal word. Whether a guide or an offer is meant is shown by the search results.Search results decide

Examples and reasons are calculated by the same logic as the tool. All signal words are listed in the section “For AI assistants” and in the tool’s data file.

How the sorter decides

  1. Signal words count. Strong signal words such as “kaufen”, “Vergleich” or “Anleitung” score 2 points, weak ones such as “kostenlos” or “Agentur” 1. A question word at the start counts 2, a domain or a brand on its own 3. The features follow a large analysis of search logs Jansen et al. 2008, carried over to German.
  2. Highest score wins. On a tie: navigational before transactional before commercial before informational. For “was kostet eine homepage”, “was” points to informational and “kostet” to commercial; it stands at 2 to 2. The price question takes precedence; the result is commercial and marked “mixed”.
  3. Place before purchase and comparison. If a term has a place reference and otherwise a purchase or comparison intent, it counts as local; the page for it is a service page with a place reference. Informational terms with a place stay informational.
  4. A brand on its own means navigational. Only what refers to a single website counts as navigational, such as a company name Rose and Levinson 2004. With a topic added, the brand is part of the topic: “amazon” is navigational, “amazon kindle preis” commercial.
  5. No signal, it stays open. “Zahnarzt” alone does not reveal whether someone is looking for a practice or has toothache. In Jansen’s analysis, such terms were automatically put under “informational”. We leave them open because many seemingly informational queries are in fact looking for a provider Rose and Levinson 2004.
  6. Questions become FAQs. If a term begins with a question word, the page plan suggests it as a question for the FAQ or as a subheading. Questions put to search engines can be classified fairly reliably by intent and type of question word Cambazoglu et al. 2021.

Why word lists and not a language model? Because rules built on keywords did not do worse in recent comparisons than learned methods and can be run quickly for every query Alexander et al. 2022. Above all, though, every decision can be reviewed. That matters, because intent is never entirely certain: in a large study of about 50,000 queries, neither outside raters nor the searchers themselves identified the intent reliably Lewandowski et al. 2012. Jansen’s rules reached 74 per cent agreement with hand classification Jansen et al. 2008. And how intents are distributed depends strongly on the topic, even with German search engines Lewandowski 2006. A fixed share of “mostly informational” would therefore be wrong for a list of tradespeople.

In the example the 23 terms are distributed as follows: informational 3, navigational 2, transactional 1, commercial 9, local 7, open 1.

When a place counts as a place

Around 13 per cent of search queries pursue a goal in a particular place. A good quarter, however, contain place words without meaning a place, for example as part of a company name Gan et al. 2008. In German, some towns also have the names of ordinary words. The sorter therefore counts “Essen”, “Halle”, “Hof”, “Siegen”, “Hagen” and similar names as a place only after “in”, “bei”, “aus” or “Raum”.

It knows around 80 larger towns and cities, the federal states, the Vogtland and the places around it. Enter smaller places in your service area in the “Your places” field. And note: a search can be meant locally even without a place name. For “Zahnarzt” or “Heizungsbauer”, Google mostly shows nearby results, and such queries can be recognised well from search logs Welch and Cho 2008. The sorter cannot see that from the wording; the map box in the search results gives it away.

Place or word?
TermIntentReason
pizza bestellen in essenLocal“bestellen” transactional, “essen” place. Place reference takes precedence over transactional.
essen bestellenTransactional“bestellen” transactional.
feier in halleLocal“halle” place.
halle mietenTransactional“mieten” transactional.
heizungsbauer 08228Local“08228” place.

One page per place?

The temptation is great: “Webdesign Plauen”, “Webdesign Zwickau”, “Webdesign Auerbach”, one page each with the place name swapped. Google explicitly names this as an example of doorway pages in its spam policies. The sorter therefore puts terms that differ only in the place into one cluster and points it out when a cluster names several places. A page of its own for a place is worthwhile if you have something of your own to say there: a location, projects there or a separate process. More important than any place page is a complete Google Business Profile; the Business Profile check shows what is missing there.

One topic, one page: spotting cannibalisation

Cannibalisation arises when two pages of the same website compete for the same term. Google then shows one, then the other, and neither ranks as high as a single strong page would. It usually happens gradually: a blog article on the topic, later a service page, plus an old promotion page that was never deleted. The sorter finds three kinds, in this order.

Notes in the page plan
LevelWhat the sorter seesWhat to do
CannibalisationA term appears in the list with two URLs, or a cluster is spread across several URLs.Choose one page as the target; merge the others and hand them over by redirect, or delimit them clearly.
CheckOne URL is meant to serve several clusters, or two clusters lie just below the threshold.Compare the first ten results for both main terms. If they are largely the same, it belongs on one page.
PlanningThe same topic in two intents, several places in one cluster, years.Separate guide and service page and link them; place pages only with a real connection; keep a visible date up to date.

The sorter sees the first level only if you give it the URLs. The easiest way is Search Console: in the “Performance” report, filter for a search query and check in the “Pages” tab whether more than one URL appears. The line then looks like this: steuerberater rodewisch;300;https://beispiel.de/steuerberatung. For merging, what Google writes about the canonical URL applies: a permanent redirect is the strongest signal, rel="canonical" the second best. Which redirects are needed is planned by the redirect planner; the snippet preview checks the titles of the new pages.

What the page plan contains

  • Main term. The term with the highest volume; without volume, the most general. It belongs in the title and main heading.
  • Page type. From the intent with the most weight in the cluster: guide, service page with or without a place, offer page, or no new page for brand searches.
  • Suggested URL. Built from the core, guides under /ratgeber/. If your list already contains a URL, the plan gives that.
  • Secondary terms and questions. The remaining terms of the cluster belong in subheadings and text, questions in the FAQ. They do not need pages of their own.

How to turn the list into a plan

  1. Collect terms. From Keyword Planner, Search Console, the questions your customers ask on the phone. It is precisely the last that every tool lacks.
  2. Sort. Paste in the list, enter the places in your service area and your company name. Medium is right for most lists; with very similar services, “tight” helps.
  3. Check open and close cases. For every open main term and every close pair, search once yourself. Ten minutes that often save an unnecessary page.
  4. Compare with what exists. Which clusters already have a page, which do not, where do two compete? The CSV has the column “URL in the list” for this.
  5. Set the order. First the clusters with cannibalisation, then those with high volume and a clear intent. Local service pages usually bring small businesses enquiries fastest.

Where the sorter stops

  • Wording only. It does not check what Google shows for a term. If two terms share the first results, they belong together, even if the sorter separates them.
  • German first. Stemming and compounds are built for German; it also reads English signal words such as “best” or “buy”, but does not merge English word forms.
  • Lists, not dictionaries. Places, brands, equivalences and basic words are selection lists. What is missing there you add in the fields, or it goes unrecognised.
  • Volumes are taken over. None is determined. Without figures the plan orders by number of terms.

How we build pages for search terms in the Vogtland is described under SEO in Rodewisch. Why a website does not appear despite good texts is explained by the guide Website not found on Google, and what changes with AI searches is under SEO, AEO and GEO. A verdict on the whole website, covering its technical setup, content and visibility in AI assistants, comes from the SpotLens check. More tools for being found are under SEO and visibility.

What the research says

The rules in the sorter are not gut-feel decisions. Six findings support the most important ones; the full list with verification notes is further down.

  • Most queries are informational, but not all are clear-cut. Over 80 per cent informational, around 10 per cent each navigational and transactional, and a quarter stayed vague or ambiguous Jansen et al. 2008. That is why the sorter marks mixed cases.
  • Reading intent is hard, even for people. Neither raters nor the searchers themselves identified it reliably Lewandowski et al. 2012. That is why every term has its reason attached.
  • Explainable rules are enough. Keyword rules were not outperformed by learned models Alexander et al. 2022.
  • German words need stemming and splitting. Both improve search in German considerably Braschler and Ripplinger 2004; compound splitting by frequency is simple and effective Koehn and Knight 2003.
  • Short queries share few words. Pure word comparisons cluster search queries poorly Wen et al. 2002; synonyms without a shared word part fail completely Sahami and Heilman 2006. That is why there are equivalences and notes on close clusters.
  • Place words are not always places. A quarter of queries contained place words without place intent Gan et al. 2008.

Using this tool from an AI assistant

Assistants such as Claude or ChatGPT that speak the Model Context Protocol (MCP) can call this page’s calculation directly instead of rebuilding it. The server is public, needs no sign-in and stores nothing. The answer always includes the link to this page, the date of the data and the limits of the result.

Endpoint
https://aureum-tech.com/mcp
Tool
werkzeug_keyword_sortierer
Transport
Streamable HTTP, read-only
Inputs
NameRequiredMeaning
begriffeyesThe term list, at most 1000 entries. German; English signal words (buy, best, how) are also recognised.
strengenoHow tightly clusters are formed: eng (tight) = only the same core (by word stem, equivalence and compound) (threshold 1); mittel (medium) = core mostly matches (threshold 0.6); weit (wide) = a shared main topic is enough (threshold 0.4). Default: medium.
ortenoOptional: your own places in your service area that should count as a place reference, e.g. ["Treuen", "Lengenfeld"].
markennoOptional: your own and other brand names. If a brand stands alone in a term, it counts as navigational.

Questions where an assistant should call it

  • “Which of these 200 keywords belong on the same page, and which page do I need for each group?”
  • “Is ‘webdesign kosten’ a commercial or an informational search query?”
  • “Here is my Search Console export with query and page: where are my pages competing with each other?”
  • “Sort these search terms for a tax adviser in Plauen by search intent and make me a page plan.”

How the connection works and what MCP can do in your own business is explained on the page about MCP integrations.

Common questions about keywords, clusters and search intent

What is keyword clustering?

Clustering groups search terms that belong on the same page. “Webdesign Plauen”, “Webdesigner Plauen” and “Webdesign in der Nähe” (web design nearby) are one topic with one intent, so one page. “Was ist SEO” (what is SEO) and “SEO Agentur Plauen” share the topic but reflect different needs: a guide and a service page. The keyword sorter groups by word stem and compound parts, kept apart by intent; in the example on the page, 23 terms become 10 clusters.

What is keyword cannibalisation?

Cannibalisation means that several pages of the same website target the same search term and compete for the same ranking. The surest sign is a term for which two URLs appear alternately in Search Console. If you enter the URLs in the third column, the sorter reports such cases first. The solution is almost always one target page that the others redirect or link to.

Which search intents are there?

Since 2002, research has distinguished informational, navigational and transactional searches; Google calls them Know, Website and Do and adds Visit-in-person for a visit in person. In an analysis of over 1.5 million queries, over 80 per cent were informational and around 10 per cent each navigational and transactional. The sorter adds commercial intent (comparison before buying) and local intent.

Do singular, plural and other word orders belong on one page?

Yes. Google understands “Webseiten erstellen” and “Webseite erstellen” as the same question, and two texts for it would only compete with each other. The sorter reduces both to the same word stem (“webseit”) and puts them in the same cluster. Terms are kept apart only when the intent changes, for example from “Kosten” (costs) to “Anleitung” (guide).

Do I need a separate page for every place?

Only if you really have something to offer there, such as a location, projects or a process of its own. Near-identical pages that differ only in the place name are treated by Google as doorway pages and therefore as spam. For a service area, one service page that names the places is usually enough, together with a well-kept Google Business Profile.

Where do I get the search volume?

Keyword Planner in Google Ads gives monthly searches in ranges or rounded; for existing pages, Search Console shows the impressions per search query. Both exports can be pasted in directly; the header row is recognised. The sorter does not determine any volume itself; without figures it orders the clusters by number of terms.

Where does the keyword sorter stop?

It reads only the wording: word stem, compounds, signal words and places. It cannot see what Google actually shows for a term, and it knows synonyms without a shared word part only from a short list. Whether the clusters hold up against the real search results, what the pages look like and which redirects are needed, we work out with you; the overall verdict on your website comes from SpotLens.

What the tool is based on

The reference values and notes above rest on these 17 scientific papers. Each has been checked via its DOI or the publisher’s page (last on 14 November 2026): it exists, it has not been retracted, and its finding still holds for the question today.

  1. Broder, A. (2002): A taxonomy of web search. ACM SIGIR Forum 36 (2), pp. 3–10

    Finding A web search is often not driven by a need for information. The searcher wants to reach a particular page (navigational) or do something, such as shop or download a file (transactional).

    In the tool The sorter’s three basic classes: informational, navigational, transactional.

  2. Rose, D. E.; Levinson, D. (2004): Understanding user goals in web search. Proceedings of the 13th International Conference on World Wide Web (WWW 2004), pp. 13–19

    Finding Navigational searches are rarer than assumed. Only what refers to a single authoritative website, such as a company name, counts as navigational. A large share of seemingly informational queries aim to find out where a product or service is available.

    In the tool A brand on its own counts as navigational, a brand with a topic does not. Terms without a signal word are not automatically treated as guides but stay open.

  3. Jansen, B. J.; Booth, D. L.; Spink, A. (2008): Determining the informational, navigational, and transactional intent of Web queries. Information Processing & Management 44 (3), pp. 1251–1266

    Finding A classifier with fixed features (question words, company and domain names, short queries, purchase and download words) classified over 1.5 million queries: over 80 per cent informational, around 10 per cent each navigational and transactional. It agreed with classification by hand in 74 per cent of cases; the rest was vague or ambiguous.

    In the tool The signal word lists follow these features. The sorter marks ambiguous terms as “mixed” instead of hiding them.

  4. Lewandowski, D. (2006): Query types and search topics of German Web search engine users. Information Services & Use 26 (4), pp. 261–269

    Finding 1,500 queries from the German search engines Fireball, Seekport and Metager, classified after Broder: all three types occur on a large scale, and their distribution differs clearly by topic.

    In the tool No fixed standard for all terms: the sorter classifies each term by its own signal words.

  5. Lewandowski, D.; Drechsler, J.; von Mach, S. (2012): Deriving query intents from web search engine queries. Journal of the American Society for Information Science and Technology 63 (9), pp. 1773–1788

    Finding Around 50,000 queries, classified by raters and cross-checked with click data and a survey: neither outside raters nor the searchers themselves stated the intent reliably.

    In the tool Every classification shows its reason. For open and mixed terms the tool points to the search results instead of feigning certainty.

  6. Alexander, D.; Kusa, W.; de Vries, A. P. (2022): ORCAS-I: Queries Annotated with Intent using Weak Supervision. Proceedings of the 45th International ACM SIGIR Conference (SIGIR 2022), pp. 3057–3066

    Finding Search queries were labelled using keyword rules and heuristics after Broder (informational subdivided more finely). Learned models did not outperform these rules, and rules can be run quickly for every query.

    In the tool The reason for open word lists instead of a model: equal quality, and every decision can be read back.

  7. Cambazoglu, B. B.; Tavakoli, L.; Scholer, F. et al. (2021): An Intent Taxonomy for Questions Asked in Web Search. Proceedings of the 2021 Conference on Human Information Interaction and Retrieval (CHIIR 2021), pp. 85–94

    Finding A classification by intent, type of question word and expected answer was developed from 1,000 real questions put to a search engine. It is finer than older schemes and still achieves higher agreement between raters.

    In the tool Question words at the start make a term a question; the page plan suggests such terms for the FAQ.

  8. Ashkan, A.; Clarke, C. L. A.; Agichtein, E.; Guo, Q. (2009): Classifying and Characterizing Query Intent. Advances in Information Retrieval (ECIR 2009), Lecture Notes in Computer Science 5478, pp. 578–586

    Finding The study focused mainly on commercial intent. Ad clicks, features of the query and the content of the results page together identified it well.

    In the tool Commercial is a class of its own. Because the results page also gives away the intent, the tool advises looking at the hits when in doubt.

  9. Gan, Q.; Attenberg, J.; Markowetz, A.; Suel, T. (2008): Analysis of geographic queries in a search engine log. Proceedings of the First International Workshop on Location and the Web (LocWeb 2008), pp. 49–56

    Finding Of 36 million AOL queries, around 13 per cent pursued a geographic goal. A further good 24 per cent contained place words without meaning a place, for example in company names.

    In the tool Place names that are also words (“Essen”, “Halle”, “Hof”) count as a place only after “in”, “bei” and similar words.

  10. Welch, M. J.; Cho, J. (2008): Automatically identifying localizable queries. Proceedings of the 31st Annual International ACM SIGIR Conference (SIGIR 2008), pp. 507–514

    Finding Some queries without a place are “localisable”: results from the user’s surroundings fit better. A classifier recognised such queries from the search log with 94 per cent accuracy.

    In the tool Note in the result: an open term such as “zahnarzt” may still be meant locally; whether Google shows a map gives it away.

  11. Wen, J.-R.; Nie, J.-Y.; Zhang, H.-J. (2002): Query clustering using user logs. ACM Transactions on Information Systems 20 (1), pp. 59–81

    Finding Because search queries are so short, a pure comparison of words is poor for clustering. Which documents users click for a query helps; the combination of words and clicks worked best.

    In the tool The sorter compares word stems and compound parts instead of bare words. For close clusters it recommends comparing the search results, because it lacks the click data.

  12. Sahami, M.; Heilman, T. D. (2006): A web-based kernel function for measuring the similarity of short text snippets. Proceedings of the 15th International Conference on World Wide Web (WWW 2006), pp. 377–386

    Finding Short texts such as search queries often share not a single word although they mean the same, for example “AI” and “Artificial Intelligence”. A word comparison then gives zero; context from search results closes the gap.

    In the tool The list of equivalences (Homepage, Webseite, Website …) and the limit that other synonyms are not recognised.

  13. Braschler, M.; Ripplinger, B. (2004): How Effective is Stemming and Decompounding for German Text Retrieval?. Information Retrieval 7 (3–4), pp. 291–316

    Finding For German texts, stemming helps even with simple methods. Careful splitting of compounds improves search considerably.

    In the tool Every term is reduced to word stems, compounds are split into their parts.

  14. Monz, C.; de Rijke, M. (2002): Shallow Morphological Analysis in Monolingual Information Retrieval for Dutch, German, and Italian. Evaluation of Cross-Language Information Retrieval Systems (CLEF 2001), Lecture Notes in Computer Science 2406, pp. 262–277

    Finding Stemming and the splitting of compounds improved search in German by around 25 per cent, in Dutch by up to 69 per cent.

    In the tool Evidence that the effort involved in handling German compounds is worthwhile, especially for short search terms.

  15. Koehn, P.; Knight, K. (2003): Empirical methods for compound splitting. Proceedings of the 10th Conference of the European Chapter of the ACL (EACL 2003), pp. 187–193

    Finding Compounds are split into every sequence of known words of at least three letters, with “s” or “es” as a joining element. The split with the highest geometric mean of word frequencies is chosen. With knowledge from parallel texts, the method reached 99.1 per cent accuracy.

    In the tool The sorter’s splitting follows this method; the frequencies come from your list and a basic word list.

  16. Riedl, M.; Biemann, C. (2016): Unsupervised Compound Splitting With Distributional Semantics Rivals Supervised Methods. Proceedings of NAACL-HLT 2016, pp. 617–622

    Finding The parts of a compound are usually similar in meaning to the whole; this rules out false splits. Without linguistic knowledge, the method comes close to rule-based and supervised methods in German.

    In the tool The sorter does not check meaning. That is why the whole word is kept and the parts count only half; a wrong split can therefore not merge clusters.

  17. Porter, M. F. (1980): An algorithm for suffix stripping. Program 14 (3), pp. 130–137

    Finding Endings are removed in several steps, each depending on the shape of the rest of the word. The simple method did somewhat better than a considerably more elaborate system.

    In the tool Basis of the Snowball stemmer for German, with which the sorter merges word forms.

Standards, authorities and manufacturer data

Mon–Fri 8:00–17:00

Do you have your page plan? Send us the CSV or the copied result. We check the clusters against the real search results, write the pages and set up the redirects, for businesses in Rodewisch and across the whole Vogtland.

+49 3744 4399760[email protected]

Mon–Fri 8:00–17:00 · Straße der Jugend 5, 08228 Rodewisch

Discuss your page plan
+49 3744 4399760Mon–Fri 8:00–17:00
CallCallback

PARTS LIST · SHEET 1/1

What this page is made of

Values measured at build time. With JavaScript, your browser measures them again live.

ITEMPARTTYPESIZE
01index.htmlDocument≈ 29 kB
02styles.cssStylesheet≈ 20 kB
03scripts.jsScript≈ 13 kB
04archivo-variable.woff2Font≈ 58 kB
05ibm-plex-mono.woff2Font≈ 30 kB
06favicon.svgImage< 1 kB

Advertising or analytics cookies: none. Without your consent, the page loads nothing from third-party servers apart from Cloudflare, our delivery service. Fonts, scripts and images are hosted by us. That is how we build your IT, too.

OUTSIDE OFFICE HOURS

Feel free to call anyway

Our regular office hours are Monday to Friday, 8:00–17:00.

In the evening and at weekends the number is forwarded to us, and often someone will pick up. We cannot promise that, though.

If nobody answers, leave a short message on the voicemail: your name, your number and what it is about. We will call back as soon as we can, on the next working day at the latest.

The call costs nothing. If someone has to start work straight away outside office hours, the surcharge in the price list applies.

Call now+49 3744 4399760

Prefer to write? Use the contact form or e-mail [email protected]. We will read it on the next working day.

DIAGNOSIS · AT-K2

What do you need help with?

↵ SELECTESC CLOSECALL: +49 3744 4399760

Custom settings

Which services may load?

Strictly necessary

The choice you make here and your settings for colour scheme and motion. These are kept as small entries on your device and are never sent to a server. They are needed for the site to work and therefore cannot be switched off.