Named entity recognition¶

LIMA includes predefined pipelines for named entity recognition (NER) in English and French, available out of the box in the legacy eng and fre configurations.

Pipeline	Input	Output
`ner-rules`	Plain text	CoNLL-03
`ner-rules-pretok`	CoNLL-U	CoNLL-03

The output can be switched to CoNLL-U in the configuration of the conllDumperNer process unit (in lima-lp-eng.xml and lima-lp-fre.xml).

Rule-based NER is implemented with ModEx rules, whose sources are in lima_linguisticdata/SpecificEntities/<lang>. You can add your own entity types by writing a ModEx.

Neural NER

LIMA previously also provided neural (ner-deep) and combined (ner-fusion) NER pipelines, based on TensorFlow. TensorFlow support has been removed; these pipeline names are kept for compatibility but currently run the rule-based recognizer only. A libtorch-based NER unit (RnnNER) exists but no models are published for it yet.

Usage¶

analyzeText -l eng -p ner-rules input_file.txt
analyzeText -l fre -p ner-rules-pretok input_file.conllu

Processing units¶

	`ner-rules`	`ner-rules-pretok`
Input
`flattokenizer`	✓
`conllureader`		✓
Pre-processing
`simpleWord`	✓	✓
`defaultProperties`	✓	✓
Rule-based NER
`SpecificEntitiesModex`	✓	✓
`sentenceBoundariesUpdater`	✓	✓
Output
`conllDumperNer`	✓	✓

Evaluation¶

The pipelines are evaluated on pre-tokenized input (ner-rules-pretok).

English — CoNLL-03 (eng.testb)English — WikiNER (aij-wikiner-en-wp2)French — WikiNER (aij-wikiner-fr-wp2)

processed 46435 tokens with 5616 phrases; found: 4440 phrases; correct: 2984.
accuracy:  92.45%; precision:  67.21%; recall:  53.13%; FB1:  59.35
              LOC: precision:  64.31%; recall:  84.99%; FB1:  73.22  2202
             MISC: precision:  90.32%; recall:   3.99%; FB1:   7.65  31
              ORG: precision:  57.07%; recall:  20.10%; FB1:  29.73  580
              PER: precision:  74.31%; recall:  75.47%; FB1:  74.88  1627

processed 3499655 tokens with 296413 phrases; found: 211853 phrases; correct: 128203.
accuracy:  92.12%; precision:  60.52%; recall:  43.25%; FB1:  50.45
              LOC: precision:  64.89%; recall:  55.78%; FB1:  59.99  72577
             MISC: precision:  57.84%; recall:   1.84%; FB1:   3.56  2144
              ORG: precision:  41.70%; recall:  36.84%; FB1:  39.12  41009
              PER: precision:  65.30%; recall:  64.00%; FB1:  64.64  96123

processed 3499679 tokens with 251726 phrases; found: 141976 phrases; correct: 102255.
accuracy:  92.67%; precision:  72.02%; recall:  40.62%; FB1:  51.95
              LOC: precision:  68.93%; recall:  45.60%; FB1:  54.89  74057
             MISC: precision:  45.09%; recall:   4.74%; FB1:   8.57  4096
              ORG: precision:  53.15%; recall:  33.15%; FB1:  40.84  15262
              PER: precision:  84.94%; recall:  54.04%; FB1:  66.06  48561

The rule-based recognizer processes about 6,400 tokens per second on a single thread.