Error Analysis

The errors are the same as those found in the others systems: mainly coordination and noise. The ten first errors for NP are: (the correct NP is underlined)

• fixed/VBG [ leading/edges ]
• [ an/DT aerospace/NN ],/, [ electronics/NNS ] ,/,
  automotive/JJ and/CC [ graphics/NNS concern/VBP⁶ ]
• [ SHEARSON/NNP LEHMAN/NNP HUTTON/NNP Inc/NNP ] ./.
• [ development/NN and/CC property/NN management/NN ]
• [ its/PRP$ total/JJ loan/NN ] and/CC[ real/JJ estate/NN reserves/NNS ]
• [ the/DT pilots/NNS ] and/CC [ management-led/JJ buy-out/NN ]
• [ Republic/NNP Airlines/NNPS and/CC Flying/NNP Tiger/NNP]
• [ US/PRP ] [ Facilities/NNP Corp./NNP ] (3 times)

These examples illustrate well the two main problems: error of tagging and the coordination structure as Table 11 confirms it. The main source of errors is due to errors of tagging or errors of bracketing (the structure extraction from the original corpus is not trivial). Among the other errors, we find the traditional structures that are problematic (gerund, apposition). For these errors, the length of the context plays a small importance and its increase provides no improvement.

Table 11: Typology of the 200 first errors.

Error types	#	%
tagging/bracketing errors	57	28.5%
coordination	45	22.5%
gerund	15	7.5%
adverb	13	6.5%
appositives	13	6.5%
quotation marks, punctuation	10	5 %
past participle	9	4.5%
that (IN)	6	3%