A Sentimental Wordnet
Fer when yeh've come to weigh the good an' bad— The gladness wiv the sadness you 'ave 'ad—
from Songs of a Sentimental Bloke by C. J. Dennis
This code makes a sense-based sentiment lexicon for English. It combines word-level and POS-level lexicons from various sources and a sense and sentiment tagged corpus (the NTU-MC). These scores are propogated across the semantic network using smantic links.
We include the most useful of the lexicons built:
| Method | Size | non 0 | ρ | Cover | % of 1054 | non 0 | % of 597 | Name in Paper |
|---|---|---|---|---|---|---|---|---|
| Lexicons | 16,499 | 8,179 | 0.86 | 217 | 20.59% | 153 | 25.63% | LEX |
| Corpus | 11,154 | 2,989 | 0.75 | 339 | 32.16% | 200 | 33.50% | NTUMC |
| Lexicon+Corpus | 26,325 | 10,793 | 0.81 | 471 | 44.69% | 296 | 49.58% | ALL |
| Propagated | 79,155 | 49,466 | 0.82 | 768 | 72.87% | 467 | 78.22% | ALL P¹¹ |
| SentiWordnet | 117,659 | ? | 0.63 | 1,054 | 100.00% | 597 | 100.00% |
Micro-WNOp is used for the evaluation. Size is the number of senses in the lexicon; non 0 is the number with non-zero sentiment; ρ is the correlation with Micro-WNOp, and cover is the number of entries in Micro-WNOp found in the lexicon (1,054 possible, with 597 non-zero). We also compare to SentiWordnet.
- LEX is monosemous words linked to their sentiment in word(+pos)-based hand-built lexicons
- NTUMC is words marked for sentiment in the Natural text Understanding Multilingual Corpus
- ALL is the combination of these, the best combination of manual resources
- ALL P¹¹ propogates these through synonym, derivation, pertainym and antonym links 11 times (until it stabilizes). This is the lexicon we suggest you use.
- Sentiwordnet is the first sentiment wordnet built by propogation
The format is yaml, with the key being the key in the OMW wordnet derived from PWN 3.0 (omw:1.4) in the omw-data package and the value being the sentiment: a float between +1 and -1.
omw-en-happy--go--lucky-01998260-s: 0.6666666666666666
omw-en-happy-01148283-a: 0.513
omw-en-happy-02565583-s: 0.505
omw-en-harangue-00990249-v: -0.6666666666666666
omw-en-harassed-02455845-s: -0.625
omw-en-harbinger-00974173-v: 0.0
omw-en-hard--and--fast-02506267-s: 0.0
omw-en-hard--bitten-02448623-s: 0.0
omw-en-hard--line-01026150-s: -0.3333333333333333
omw-en-hard--nosed-01940651-s: -0.1
omw-en-hard--pressed-02457558-s: -0.27The process is described in full detail in Bond and Herng (2022)
Francis Bond and Merrick Choo. 2022. Sense and Sentiment. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 61–69, Marseille, France. European Language Resources Association.
Erata: In Table 9 we say there are 457 non-zero senses in Micro-WNOp, but there are 597 non-zero senses and 457 senses where the score is zero!