Update README.md
Browse files
README.md
CHANGED
@@ -973,3 +973,59 @@ configs:
|
|
973 |
path:
|
974 |
- zza.json.gz
|
975 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
973 |
path:
|
974 |
- zza.json.gz
|
975 |
---
|
976 |
+
# Paraphrases of Sentences from Tatoeba
|
977 |
+
|
978 |
+
Downloaded: 2024-07-08
|
979 |
+
|
980 |
+
**Description:**
|
981 |
+
- Large multilingual paraphrase dataset
|
982 |
+
- If two sentences translate to the same sentence in a foreign language, then those two sentences could be paraphrases.
|
983 |
+
|
984 |
+
**Creation Process:**
|
985 |
+
1. Translation pairs of sentences are first linked in a large undirected graph.
|
986 |
+
2. The connected components of the graph contain sentences with similar meanings.
|
987 |
+
3. Sentences in the same language in the same component are taken to be paraphrases.
|
988 |
+
4. Only clusters of 2-20 sentences are counted. Any larger cluster is likely to come from irrelevant sentences linked by bad translation.
|
989 |
+
5. Only languages with at least 10 clusters are included here.
|
990 |
+
|
991 |
+
**Format:**
|
992 |
+
- GZipped JSON files
|
993 |
+
- Each line contains one cluster of paraphrase sentences in JSON format
|
994 |
+
|
995 |
+
**Dataset Configs:**
|
996 |
+
- [ISO 639-3 language codes](https://en.wikipedia.org/wiki/List_of_ISO_639-3_codes) as listed on Tatoeba
|
997 |
+
|
998 |
+
**Sample entry:**
|
999 |
+
|
1000 |
+
```json
|
1001 |
+
{
|
1002 |
+
"id": 5,
|
1003 |
+
"sentences": [
|
1004 |
+
"Today is June 18th and it is Muiriel's birthday!",
|
1005 |
+
"It's 18th June today, and also Muiriel's birthday!",
|
1006 |
+
"Today is June 18th, the birthday of Muiriel!",
|
1007 |
+
"Today is June 18 and it is Muriel's birthday!"
|
1008 |
+
]
|
1009 |
+
}
|
1010 |
+
```
|
1011 |
+
|
1012 |
+
`id`: The cluster ID of the sentences. Clusters with the same ID in other languages mutually translate to this set of sentences.
|
1013 |
+
|
1014 |
+
`sentences`: List of sentences belonging to this cluster. That is, those sentences are paraphrases of each other.
|
1015 |
+
|
1016 |
+
# Limitations
|
1017 |
+
|
1018 |
+
Many important things are lost in translation. Even if two sentences translate to the same thing in a foreign language, they might not be good paraphrases.
|
1019 |
+
|
1020 |
+
- Context and connotations ("What's up, Mike?", "What's wrong, Mike?")
|
1021 |
+
- Gendered common nouns ("I need to get rid of it.", "I need to get rid of her.", "I need to get rid of him.")
|
1022 |
+
- Ambiguous cases that could be the same or different.
|
1023 |
+
- "I'll shoot it down.", "I'll kill it with a gun."
|
1024 |
+
- "What kind of people do you prefer?", "What's your type?", "What kind of guys do you like?"
|
1025 |
+
|
1026 |
+
Also some languages have important features such as:
|
1027 |
+
|
1028 |
+
- [Familiar vs. formal forms](https://en.wikipedia.org/wiki/T%E2%80%93V_distinction)
|
1029 |
+
- Verb tenses, modality, and moods (so a sentence pair can be paraphrases in one language but not in others)
|
1030 |
+
- Vocabulary, noun distinctions, and cultural concepts (so foreign translations will lose the original meaning of the sentence)
|
1031 |
+
- Idioms, dialects, slang, jargon
|