46 languages tried, zero samples anywhere
Abaza, Afrikaans, Alemannic, Assamese, Bavarian, Bokota, Cantonese, Cebuano, Chinese, Chukchi, Classical Chinese, Coptic, Gorontalo, Gujarati, Gwichin, Haitian Creole, Ika, Japanese, Kadiweu, Kinyarwanda, Latgalian, Luxembourgish, Madi, Makurap, Malayalam, Mbya Guarani, Munduruku, Neapolitan, Nenets, Northwest Gbaya, Old English, Old Turkish, Paumari, Pesh, Runyankore, Shanghainese, Sinhala, South Levantine Arabic, Swahili, Tagalog, Telugu, Tswana, Vietnamese, Western Sierra Puebla Nahuatl, Yiddish, isiXhosa
Samples & pairs per language
Totals summed across every condition a language appears in. The chart shows the top 25 languages by minimal pairs generated; the table below covers all 142.
Language × condition matrix
Minimal pairs generated for every language/condition combination. Sort by total pairs or alphabetically; hover a cell for exact samples, pairs, ratio, and decision-tree accuracy, or click it to open that language/condition's own diagnostics page. Blank cells are combinations that were never run, not zero-pair combinations.
142 languages × 45 conditions.
View as table
Samples vs. minimal pairs, per language
Each dot is one language. The dashed line marks where pairs equal samples (1×) — languages well above it produced unusually many contrastive pairs per candidate; languages near the bottom found samples but little usable contrast. Both axes are log-scaled given the range from a handful of samples to hundreds of thousands; languages with zero pairs are nudged just above zero so they stay visible rather than dropping off a log axis. Total sums each language across every condition it appears in (dots aren't clickable here -- a language's condition breakdown lives behind its own name in the table below); open see conditions and pick a single one instead to see that condition alone, with dots linking to each language's own diagnostics page.
Where samples don't become pairs
Aggregated outcome buckets across every language/condition (from each run's own diagnostic buckets). Not a strict partition of the sample count above — one sample can land in more than one outcome across its accepted swap candidates — but it shows where the corpus loses candidates: no matching swap candidate, an inflector that couldn't produce a contrastive form, a reinflection that came out identical, or a genuinely ambiguous subject.
View as table
| Outcome | Instances |
|---|
Samples by treebank
Which treebank each language's raw candidate samples actually came from, for every minimal pair those samples could produce — not an estimate, real per-row counts straight from this repo's own decision-tree cache. Coverage depends on how far each condition's own fitting run has gotten (currently 1698 language × condition combinations across 149 languages, out of this page's full 142), so it can be a subset of the rest of this page rather than a full match. Samples, not pairs, by treebank: nothing here flags which individual rows became accepted pairs, so a treebank's share of samples is used as the closest honest proxy for its share of that combination's pairs.
Treebanks per language
Every UD treebank this pipeline is eligible to draw samples from for each attempted language (some languages restrict to a specific subset; most use every treebank the UD release has for them) — an inventory of possible sources, not a per-sample or per-pair count by treebank. Includes the 46 zero-yield languages above, since which treebank(s) a language drew from can be part of why it came back empty. For the finer, per-sample breakdown, see above.
| Treebanks |
|---|
Subject–Verb
3 conditions tracked in this category, 120 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
67 languages tried, zero samples anywhere
Abaza, Afrikaans, Alemannic, Assamese, Assyrian, Bambara, Bavarian, Bengali, Bokota, Cantonese, Cebuano, Chinese, Chukchi, Classical Chinese, Coptic, Danish, Esperanto, Gorontalo, Gujarati, Gwichin, Haitian Creole, Ika, Indonesian, Japanese, Javanese, Kadiweu, Khoekhoe, Kinyarwanda, Korean, Latgalian, Luxembourgish, Madi, Makurap, Malayalam, Manx, Mbya Guarani, Munduruku, Naga, Naija, Neapolitan, Nenets, Northwest Gbaya, Norwegian, Old English, Old Irish, Old Turkish, Paumari, Pesh, Runyankore, Scottish Gaelic, Shanghainese, Sinhala, South Levantine Arabic, Swahili, Swedish, Tagalog, Telugu, Thai, Tswana, Vietnamese, Warlpiri, Western Sierra Puebla Nahuatl, Xibe, Yiddish, Yoruba, Zaar, isiXhosa
Dropped before fitting in this category
Rows across Subject–Verb's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Subject–Participle
3 conditions tracked in this category, 48 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
53 languages tried, zero samples anywhere
Afrikaans, Albanian, Armenian, Aromanian, Assamese, Assyrian, Azerbaijani, Bambara, Belarusian, Bengali, Breton, Chintang, Dutch, Esperanto, Estonian, Faroese, German, Greek, Hausa, Hittite, Karelian, Kazakh, Kiche, Komi Zyrian, Kyrgyz, Livvi, Low Saxon, Macedonian, Malayalam, Marathi, Middle French, Moksha, Naija, Neapolitan, Nepali, Northern Kurdish, Odia, Old English, Old French, Old Georgian, Phrygian, Sinhala, Skolt Sami, Swedish, Tamil, Tatar, Umbrian, Uyghur, Uzbek, Veps, Western Armenian, Xibe, Zazaki
Dropped before fitting in this category
Rows across Subject–Participle's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Subject–Auxiliary
3 conditions tracked in this category, 86 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
89 languages tried, zero samples anywhere
Afrikaans, Akuntsu, Alemannic, Amharic, Apurina, Aromanian, Assamese, Azerbaijani, Bambara, Bavarian, Bengali, Bokota, Brahui, Buryat, Cantonese, Cappadocian, Central Kurdish, Chinese, Chintang, Chukchi, Classical Chinese, Coptic, Danish, Egyptian, Esperanto, Gujarati, Haitian Creole, Highland Puebla Nahuatl, Hittite, Hungarian, Ika, Indonesian, Irish, Japanese, Javanese, Kaapor, Kadiweu, Karo, Khoekhoe, Kiche, Kinyarwanda, Komi Permyak, Korean, Kyrgyz, Latgalian, Livvi, Luxembourgish, Madi, Makurap, Malayalam, Maltese, Manx, Moksha, Munduruku, Naga, Naija, Neapolitan, Nenets, Nepali, Northwest Gbaya, Norwegian, Occitan, Odia, Old English, Old Irish, Old Turkish, Paumari, Pesh, Phrygian, Runyankore, Scottish Gaelic, Shanghainese, Sinhala, Skolt Sami, South Levantine Arabic, Swahili, Swedish, Tagalog, Tatar, Teko, Thai, Tswana, Umbrian, Vietnamese, Western Sierra Puebla Nahuatl, Xibe, Yakut, Yiddish, isiXhosa
Dropped before fitting in this category
Rows across Subject–Auxiliary's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Object–Verb
3 conditions tracked in this category, 9 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
178 languages tried, zero samples anywhere
Abaza, Abkhaz, Afrikaans, Akkadian, Akuntsu, Albanian, Alemannic, Amharic, Ancient Greek, Ancient Hebrew, Arabic, Armenian, Aromanian, Assamese, Assyrian, Azerbaijani, Bambara, Bavarian, Beja, Belarusian, Bengali, Bhojpuri, Bokota, Bororo, Brahui, Breton, Bulgarian, Buryat, Cantonese, Cappadocian, Catalan, Cebuano, Central Kurdish, Chinese, Chintang, Chukchi, Classical Armenian, Classical Chinese, Coptic, Croatian, Czech, Danish, Dutch, Egyptian, English, Esperanto, Estonian, Faroese, Finnish, French, Galician, German, Gothic, Greek, Guajajara, Gujarati, Gwichin, Haitian Creole, Hausa, Hebrew, Hindi, Hittite, Hungarian, Icelandic, Ika, Indonesian, Irish, Italian, Japanese, Javanese, Kaapor, Kadiweu, Kangri, Karelian, Karo, Kazakh, Khoekhoe, Kiche, Kinyarwanda, Komi Permyak, Komi Zyrian, Korean, Kyrgyz, Latgalian, Latin, Latvian, Ligurian, Lithuanian, Livvi, Low Saxon, Luxembourgish, Macedonian, Madi, Makurap, Malayalam, Maltese, Manx, Marathi, Mbya Guarani, Middle French, Munduruku, Naga, Naija, Neapolitan, Nenets, Nepali, Nheengatu, North Sami, Northern Kurdish, Northwest Gbaya, Norwegian, Occitan, Odia, Old Church Slavonic, Old East Slavic, Old English, Old French, Old Irish, Old Occitan, Old Turkish, Ottoman Turkish, Pashto, Paumari, Persian, Pesh, Phrygian, Polish, Pomak, Portuguese, Punjabi, Romanian, Runyankore, Russian, Sanskrit, Scottish Gaelic, Serbian, Shanghainese, Sicilian, Sindhi, Sinhala, Skolt Sami, Slovak, Slovenian, South Levantine Arabic, Southern Kurdish, Spanish, Swahili, Swedish, Tagalog, Tamil, Tatar, Teko, Telugu, Thai, Tswana, Turkish, Ukrainian, Umbrian, Upper Sorbian, Urdu, Uyghur, Uzbek, Veps, Vietnamese, Warlpiri, Welsh, Western Armenian, Western Sierra Puebla Nahuatl, Wolof, Xavante, Xibe, Yakut, Yiddish, Yoruba, Yupik, Zaar, Zazaki, isiXhosa
Dropped before fitting in this category
Rows across Object–Verb's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Object–Participle
3 conditions tracked in this category, 1 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
95 languages tried, zero samples anywhere
Afrikaans, Albanian, Ancient Greek, Ancient Hebrew, Armenian, Assamese, Azerbaijani, Belarusian, Bengali, Bhojpuri, Breton, Bulgarian, Catalan, Chintang, Classical Armenian, Croatian, Czech, Danish, Dutch, Egyptian, English, Erzya, Esperanto, Estonian, Faroese, Finnish, French, Galician, Georgian, German, Gothic, Greek, Hausa, Hebrew, Hindi, Hittite, Icelandic, Italian, Kangri, Karelian, Kazakh, Kiche, Kyrgyz, Latin, Latvian, Ligurian, Lithuanian, Livvi, Low Saxon, Macedonian, Malayalam, Marathi, Middle French, Moksha, Naija, Neapolitan, Nepali, North Sami, Northern Kurdish, Norwegian, Odia, Old Church Slavonic, Old East Slavic, Old English, Old French, Old Georgian, Old Occitan, Ottoman Turkish, Pashto, Persian, Phrygian, Pomak, Portuguese, Punjabi, Romanian, Russian, Sanskrit, Serbian, Sicilian, Sinhala, Skolt Sami, Slovak, Slovenian, Spanish, Swedish, Tamil, Tatar, Turkish, Upper Sorbian, Urdu, Uyghur, Uzbek, Veps, Western Armenian, Xibe
Dropped before fitting in this category
Rows across Object–Participle's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Object–Auxiliary
3 conditions tracked in this category, 2 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
164 languages tried, zero samples anywhere
Abkhaz, Afrikaans, Akuntsu, Albanian, Alemannic, Amharic, Ancient Greek, Ancient Hebrew, Apurina, Arabic, Armenian, Assamese, Assyrian, Azerbaijani, Bambara, Bavarian, Beja, Belarusian, Bengali, Bhojpuri, Bokota, Brahui, Breton, Bulgarian, Buryat, Cantonese, Cappadocian, Catalan, Central Kurdish, Chinese, Chintang, Chukchi, Classical Armenian, Classical Chinese, Coptic, Croatian, Czech, Danish, Dutch, Egyptian, English, Erzya, Esperanto, Estonian, Faroese, Finnish, French, Galician, Georgian, German, Greek, Guajajara, Gujarati, Haitian Creole, Hausa, Hebrew, Highland Puebla Nahuatl, Hindi, Hittite, Hungarian, Icelandic, Ika, Indonesian, Irish, Italian, Japanese, Javanese, Kaapor, Kangri, Karelian, Karo, Kazakh, Khoekhoe, Kiche, Kinyarwanda, Komi Permyak, Komi Zyrian, Korean, Kyrgyz, Latin, Latvian, Ligurian, Lithuanian, Livvi, Low Saxon, Luxembourgish, Macedonian, Madi, Malayalam, Maltese, Marathi, Middle French, Moksha, Munduruku, Naga, Naija, Neapolitan, Nepali, Nheengatu, North Sami, Northern Kurdish, Norwegian, Occitan, Odia, Old Church Slavonic, Old East Slavic, Old English, Old French, Old Georgian, Old Occitan, Old Turkish, Ottoman Turkish, Pashto, Paumari, Persian, Pesh, Phrygian, Polish, Pomak, Portuguese, Punjabi, Romanian, Runyankore, Russian, Ruuli, Sanskrit, Scottish Gaelic, Serbian, Shanghainese, Sicilian, Sindhi, Sinhala, Skolt Sami, Slovak, Slovenian, South Levantine Arabic, Spanish, Swahili, Swedish, Tamil, Tatar, Teko, Thai, Tswana, Turkish, Ukrainian, Umbrian, Upper Sorbian, Urdu, Uyghur, Uzbek, Veps, Vietnamese, Welsh, Western Armenian, Western Sierra Puebla Nahuatl, Wolof, Xavante, Xibe, Yiddish, Yoruba, Zaar, Zazaki, isiXhosa
Dropped before fitting in this category
Rows across Object–Auxiliary's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Indirect Object–Verb
3 conditions tracked in this category, 3 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
137 languages tried, zero samples anywhere
Abaza, Afrikaans, Akkadian, Akuntsu, Albanian, Alemannic, Amharic, Ancient Greek, Ancient Hebrew, Apurina, Arabic, Armenian, Aromanian, Assamese, Assyrian, Beja, Belarusian, Bhojpuri, Brahui, Bulgarian, Buryat, Cantonese, Cappadocian, Central Kurdish, Chinese, Chintang, Classical Armenian, Classical Chinese, Coptic, Croatian, Czech, Danish, Dutch, English, Esperanto, French, Galician, German, Greek, Guajajara, Gujarati, Gwichin, Haitian Creole, Hausa, Hebrew, Highland Puebla Nahuatl, Hindi, Hittite, Hungarian, Icelandic, Ika, Indonesian, Italian, Japanese, Javanese, Kaapor, Kangri, Kazakh, Khoekhoe, Kinyarwanda, Korean, Latgalian, Latin, Latvian, Ligurian, Lithuanian, Low Saxon, Luxembourgish, Macedonian, Malayalam, Maltese, Manx, Marathi, Middle French, Naga, Naija, Nepali, Nheengatu, Norwegian, Occitan, Odia, Old East Slavic, Old English, Old French, Old Georgian, Old Occitan, Ottoman Turkish, Pashto, Paumari, Persian, Phrygian, Polish, Pomak, Portuguese, Punjabi, Romanian, Russian, Ruuli, Sanskrit, Serbian, Shanghainese, Sicilian, Sindhi, Sinhala, Slovak, Slovenian, South Levantine Arabic, Swahili, Swedish, Tagalog, Tamil, Teko, Telugu, Thai, Tswana, Tupinamba, Turkish, Ukrainian, Umbrian, Upper Sorbian, Urdu, Uyghur, Uzbek, Vietnamese, Warlpiri, Welsh, Western Armenian, Western Sierra Puebla Nahuatl, Wolof, Xavante, Xibe, Yakut, Yiddish, Yoruba, Zaar, Zazaki, isiXhosa
Dropped before fitting in this category
Rows across Indirect Object–Verb's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Indirect Object–Participle
3 conditions tracked in this category, 0 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
Dropped before fitting in this category
Rows across Indirect Object–Participle's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Indirect Object–Auxiliary
3 conditions tracked in this category, 1 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
108 languages tried, zero samples anywhere
Abkhaz, Afrikaans, Albanian, Alemannic, Amharic, Ancient Greek, Arabic, Armenian, Aromanian, Assyrian, Beja, Belarusian, Bhojpuri, Bulgarian, Buryat, Cantonese, Cappadocian, Central Kurdish, Chinese, Chintang, Classical Armenian, Classical Chinese, Coptic, Croatian, Czech, Danish, Dutch, English, Faroese, French, Galician, Georgian, German, Greek, Guajajara, Gujarati, Haitian Creole, Hausa, Highland Puebla Nahuatl, Hindi, Hittite, Hungarian, Icelandic, Ika, Indonesian, Italian, Japanese, Kaapor, Kangri, Khoekhoe, Korean, Latvian, Ligurian, Lithuanian, Low Saxon, Luxembourgish, Macedonian, Maltese, Marathi, Middle French, Naga, Naija, Nepali, Nheengatu, Norwegian, Occitan, Odia, Old East Slavic, Old French, Old Georgian, Old Occitan, Ottoman Turkish, Paumari, Persian, Phrygian, Polish, Pomak, Portuguese, Punjabi, Romanian, Russian, Ruuli, Sanskrit, Serbian, Shanghainese, Sicilian, Sindhi, Slovak, Slovenian, South Levantine Arabic, Swedish, Tamil, Thai, Turkish, Ukrainian, Upper Sorbian, Urdu, Uyghur, Uzbek, Vietnamese, Warlpiri, Western Armenian, Wolof, Xavante, Yiddish, Yoruba, Zaar, isiXhosa
Dropped before fitting in this category
Rows across Indirect Object–Auxiliary's 3 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Noun Phrase
77 conditions tracked in this category, 117 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
9 languages tried, zero samples anywhere
Assyrian, Highland Puebla Nahuatl, Hittite, Kadiweu, Komi Permyak, Malayalam, Neapolitan, Nepali, Western Sierra Puebla Nahuatl
Dropped before fitting in this category
Rows across Noun Phrase's 77 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.
Other
17 conditions tracked in this category, 33 languages combined. Pick one below for its own breakdown — the chart shows its top 25 languages by pairs; the table has all of them.
Total
6 languages tried, zero samples anywhere
Greek, Kangri, Old Georgian, Ottoman Turkish, Phrygian, Upper Sorbian
Dropped before fitting in this category
Rows across Other's 17 conditions where the agreement label itself was missing (head, subject, or both), so they never reached the swap-candidate pipeline at all — see the same chart on the Overview tab for the all-conditions total this is part of.
No conditions match this filter.