Genocide at the Security Council

The words that sit near a term

Which words appear near this term far more often than chance would put them there?

Term Window 6,092 occurrences, 60,050 words in window

Drawing…

View the leading collocates as a table

Word	Near	G²	Log ratio
crimes	2,558	17,826	+6.52
humanity	1,110	9,474	+7.74
war	1,093	5,317	+4.93
rwanda	573	3,378	+5.70
cleansing	359	3,196	+8.04
ethnic	393	2,255	+5.58
srebrenica	246	2,192	+8.04
tutsi	163	1,898	+10.66
rwandan	196	1,566	+7.31
crime	327	1,380	+4.44
denial	163	1,189	+6.77
prevention	295	1,029	+3.87
responsible	248	953	+4.15
committed	315	926	+3.44
punishment	134	882	+6.23
adviser	105	617	+5.69
twentieth	86	564	+6.20
acts	183	517	+3.35
atrocities	118	514	+4.54
anniversary	117	510	+4.54
perpetrators	141	479	+3.80
victims	177	458	+3.15
violations	179	376	+2.75
ideology	62	372	+5.78
perpetrated	82	358	+4.55
protect	146	352	+3.01
glorification	42	338	+7.36
populations	102	324	+3.62
prosecution	74	315	+4.46
mass	99	301	+3.52

How to read this

Each point is a word that appears within the chosen window of the term — corpus linguists call such words collocates. Further right means more confidence that the word turns up near the term at a different rate from the rest of the corpus. Further up means a larger difference: a log ratio of +3 is eight times the corpus rate, +7 is more than a hundred times.

The words worth attention are high and right. A word far right but low is simply common enough that even a small difference can be measured, which says more about how often it occurs than about how the Council speaks.

Most collocates crowd into the lower left, where labels collide and the chart hides some of them. Scroll on the plot to zoom, drag to pan, or drag the bar under the axis; Reset zoom returns to the full view. Zooming filters nothing out and the axes never move, so a close-up remains comparable with the whole.

What it does not show

A wider window asks a different question rather than a better one: ±5 words catches the phrase the term sits in, ±15 catches the argument around it. Compare the two; do not average them.

Function words — the, of, and — are removed using a published list. Words belonging to the setting, such as council, resolution and president, are deliberately kept, because whether they sit close to the term is one of the things being asked.

The same table as a cloud

What shape does that neighbourhood have, and does it hold for a single speaker or a single decade?

Drawn from Term Window Words At least nothing is drawn for a set of fewer than 20 speeches

Drawing…

40 words drawn, out of 100 held for the whole corpus: 0 fall below the minimum frequency and 60 fall beyond the number of words asked for. The table below holds those same 40 words, in the same order.

View the same words as a table

Word	Near	G²	Log ratio
crimes	3,295	21,723	+6.26
humanity	1,516	12,618	+7.65
war	1,527	7,086	+4.77
rwanda	752	4,174	+5.46
cleansing	409	3,390	+7.62
ethnic	458	2,360	+5.16
srebrenica	273	2,248	+7.58
tutsi	172	1,876	+10.21
rwandan	219	1,604	+6.84
prevention	436	1,470	+3.78
crime	373	1,347	+3.97
denial	176	1,153	+6.23
committed	395	999	+3.11
adviser	165	973	+5.73
responsible	292	963	+3.73
punishment	137	785	+5.60
violations	289	620	+2.79
glorification	71	590	+7.63
perpetrators	184	561	+3.53
twentieth	93	541	+5.67
atrocities	136	512	+4.09
acts	211	479	+2.89
victims	220	477	+2.81
protect	210	473	+2.88
anniversary	128	469	+4.01
criminals	89	427	+4.89
prosecution	102	408	+4.27
serious	207	367	+2.46
perpetrated	92	342	+4.06
ideology	65	338	+5.20
tutsis	32	316	+9.07
impunity	148	315	+2.77
justice	230	308	+2.07
responsibility	212	303	+2.15
prosecute	67	292	+4.55
mass	114	283	+3.07
convicted	51	274	+5.33
populations	109	269	+3.06
dieng	32	261	+7.51
commemoration	40	256	+6.12

How to read this

Every word drawn here is a row of the table beneath it, and one decision governs both: what the cloud leaves out, the table leaves out too. Size carries the log ratio — how many times a word's rate beside the term exceeds its rate in the rest of the corpus. Which words appear at all is decided by the other measure, log-likelihood.

A large word is therefore not a common word. The list is chosen by confidence and drawn by size of effect, so a word used two hundred times can sit beside one used thirty times and be the smaller of the two. At least raises the minimum frequency, if you would rather the cloud stopped rewarding words the corpus barely contains.

Every word links to its lines in the concordance and can be reached by keyboard. The scale is relative to what is on screen: the largest word in view holds the largest log ratio in view, so compare within one cloud rather than between two.

What it does not show

Words are counted exactly as they appear. crime and crimes are two separate words here, each carrying part of the evidence for the same idea and each smaller than a merged entry would be. A step that merges such forms exists and runs, but it is deliberately not used yet: switching it on would move figures already published here before the hand-check of the word list is finished. Until then, a cloud of this corpus is partly a picture of English word endings rather than of the Council, and the words to trust least are those with the commonest variants.

Position and colour carry nothing. Two words sit next to each other because the layout found room there, not because they occur together; the network figure below is the one that answers that question. Area carries nothing either: size is set on a word's height, so a long word takes up far more of the picture than a short one with the same log ratio.

149 function words are removed using config/stopwords.txt, and a word occurring fewer than 5 times beside the term never enters the table at all. Words belonging to the setting — council, resolution — are deliberately kept.

The same word in two mouths

Do different speakers, groups or decades use it to do different work?

Compare by Left Right

Align

nothing is drawn for a set of fewer than 20 speeches

Rwanda 187 speeches · 960 occurrences

rwanda +6.64
tutsi +11.79
denial +6.85
convicts +10.39
ideology +7.06
fdlr +7.39
fugitives +6.38
rwandan +6.61
suspects +6.68
masterminds +10.41
tutsis +10.26
twentieth +6.52
perpetrators +4.26
crimes +3.23
survivors +5.31
post-genocide +13.45
crime +3.73
committed +3.13

United States 163 speeches · 322 occurrences

crimes +5.79
humanity +6.93
srebrenica +8.61
rwandan +8.01
war +4.12
rwanda +5.07
atrocities +5.40
kabuga +8.20
responsible +4.29
accused +5.94
financier +12.69
committed +3.31
felicien +9.06
ethnic +4.41
darfur +3.71
denial +5.87
alleged +5.34
violations +3.10

View both profiles as a table

Profile	Word	Near	G²	Log ratio
Rwanda	rwanda	271	1,951	+6.64
Rwanda	tutsi	122	1,666	+11.79
Rwanda	denial	45	337	+6.85
Rwanda	convicts	24	290	+10.39
Rwanda	ideology	37	288	+7.06
Rwanda	fdlr	31	255	+7.39
Rwanda	fugitives	36	246	+6.38
Rwanda	rwandan	33	236	+6.61
Rwanda	suspects	30	218	+6.68
Rwanda	masterminds	17	206	+10.41
Rwanda	tutsis	16	191	+10.26
Rwanda	twentieth	27	190	+6.52
Rwanda	perpetrators	47	188	+4.26
Rwanda	crimes	69	186	+3.23
Rwanda	survivors	33	178	+5.31
Rwanda	post-genocide	11	165	+13.45
Rwanda	crime	49	162	+3.73
Rwanda	committed	62	159	+3.13
United States	crimes	139	843	+5.79
United States	humanity	63	480	+6.93
United States	srebrenica	37	367	+8.61
United States	rwandan	30	273	+8.01
United States	war	53	203	+4.12
United States	rwanda	32	163	+5.07
United States	atrocities	18	100	+5.40
United States	kabuga	10	94	+8.20
United States	responsible	23	93	+4.29
United States	accused	13	81	+5.94
United States	financier	5	76	+12.69
United States	committed	24	67	+3.31
United States	felicien	6	63	+9.06
United States	ethnic	15	63	+4.41
United States	darfur	16	53	+3.71
United States	denial	8	49	+5.87
United States	alleged	9	49	+5.34
United States	violations	19	48	+3.10

How to read this

Two lists of neighbouring words, each worked out from its own set of speeches at ±8 words, but both measured against the same whole-corpus background. The two columns share one scale, so a longer bar is a larger log ratio wherever it appears.

By rank puts each column's own strongest words in its own order. Rows do not line up, and what the figure shows is that the two sets of words differ. By word gives each word a single row with both sides' figures on it, so what the figure shows is how far apart the two speakers are on the same word. A dash means the word does not appear in that speaker's list at all.

Try Rwanda against any other speaker. Most delegations return the three crimes named in the Rome Statute of the International Criminal Court — genocide, crimes against humanity, war crimes. Rwanda returns a vocabulary of denial and prosecution.

What it does not show

These sets of speeches differ enormously in size, which is why the speech count is printed under each heading. A list drawn from fifty speeches is a sketch, not a portrait.

Comparing two periods mixes up who was speaking with when they spoke: Council membership turns over, and so does the agenda.

Compared with a like-for-like speech

Setting aside what the debate was about, what marks out a speech that says genocide?

Comparison 3,104 of 3,273 speeches found a partner (94.84%)

#	Word	In these speeches	G²	Log ratio
1	genocide	5,187	5,306	+12.76
2	crimes	8,162	2,159	+1.56
3	rwanda	4,624	1,051	+1.41
4	humanity	2,093	936	+2.33
5	justice	5,272	458	+0.78
6	court	3,553	445	+0.96
7	war	4,901	357	+0.70
8	atrocities	1,291	332	+1.53
9	genocidal	295	302	+8.62
10	rwandan	784	299	+2.04
11	criminal	3,727	283	+0.72
12	srebrenica	549	240	+2.29
13	victims	2,529	240	+0.82
14	cleansing	722	235	+1.82
15	impunity	2,189	235	+0.88
16	accountability	2,044	208	+0.85
17	mass	924	201	+1.37
18	ethnic	1,488	200	+1.01
19	committed	3,084	180	+0.62
20	atrocity	388	170	+2.29
21	icc	1,136	165	+1.06
22	ictr	1,333	161	+0.94
23	responsible	1,478	155	+0.87
24	tutsi	184	146	+4.36
25	violations	2,446	142	+0.61
26	rights	5,410	138	+0.39
27	sexual	2,523	137	+0.59
28	arrest	870	133	+1.09
29	prevention	1,866	133	+0.69
30	human	5,433	129	+0.38

How to read this

The table rests on 3,104 complete pairs, drawn from 3,273 speeches that use the word. Each of those is paired with a speech that does not use it but shares its year, agenda item and speaker group. What survives the comparison is closer to the vocabulary of the idea than of the occasion.

Switch to the whole corpus to see what the pairing removed. Watch bosnia, herzegovina and tribunals: near the top of the unpaired table, and gone once year and agenda item are held constant.

What it does not show

100 groups could not be filled. These are debates in which nearly everyone used the word, so no comparable speech was left over. They are left short rather than filled from elsewhere, which would have tilted the table quietly towards the crisis years.

The whole-corpus column is not a result. It is the comparison the pairing exists to improve on, shown so that the improvement can be checked.

Because the partner speech is drawn at random, the pairing was repeated across 20 consecutive draws, and the range those draws produced is available for every word in the table.

Which terms travel together

Does this vocabulary hold together in groups, or is it just a list?

Period a line is drawn where at least 20 speeches use both terms

Drawing…

View the strongest network edges as a table

Source	Target	Shared speeches	nPMI
crimes against humanity	war crimes	2,253	0.72
genocide	crimes against humanity	1,544	0.63
genocide	war crimes	1,462	0.56
mass atrocity	prevention of genocide	51	0.51
genocide	ethnic cleansing	536	0.50
impunity	icc	2,415	0.50
crimes against humanity	icc	1,206	0.49
denial	glorification	132	0.47
responsibility to protect	prevention of genocide	63	0.46
war crimes	icc	1,276	0.46
genocide	tribunals	768	0.46
genocide	mass atrocity	253	0.44
ethnic cleansing	extermination	63	0.43
crimes against humanity	impunity	1,761	0.42
genocide	atrocity	926	0.41
mass atrocity	responsibility to protect	104	0.41
crimes against humanity	atrocity	939	0.41
war crimes	impunity	1,958	0.40
genocide	extermination	109	0.40
ethnic cleansing	crimes against humanity	378	0.40
war crimes	tribunals	748	0.40
war crimes	atrocity	1,070	0.40
atrocity	icc	961	0.38
ethnic cleansing	war crimes	407	0.38
crimes against humanity	mass atrocity	199	0.38
genocide	glorification	152	0.38
ethnic cleansing	responsibility to protect	148	0.37
atrocity	prevention of genocide	97	0.37
genocide	icc	755	0.36
genocide	responsibility to protect	293	0.36

How to read this

Each circle is a term from the word list, sized by how many speeches use it and coloured by its register. A line joins two terms that appear in the same speech, and its thickness shows how much more often the two turn up together than two unrelated terms of the same frequency would.

Drag to rearrange, scroll to zoom, hover over a line for its numbers.

What it does not show

Two terms count as linked if they appear anywhere in the same speech, even four hundred words apart and in unrelated sentences. This is a map of vocabularies used on the same occasion, not of phrases.

The thickness measure is adjusted for how often each term occurs. Without that adjustment, a term appearing in only thirty speeches would dominate the picture, because the raw measure rewards rarity.

A phrase is never drawn as evidence of association with a word already inside it: mass atrocity and atrocity are not linked on the strength of the second sitting within the first.