Classifying the TRIZ Contradiction Problem of the Patents Based on Engineering Parameters - Technologies and Applications of Artificial Intelligence - page 348

Fig. 5. Four combination cases of the fourth step

As shown in Fig. 5, the case 1 is a standard process which is performed in any situ-
ation unless SPS i-1 does not exist. The case 2 is performed when SPS i contains any
verb to be (be Verb), or such as become, have, etc. The case 3 is performed when the
segments meet all the conditions as we list below. First, SPS i contains any verb which
belongs to the verb to be (be Verb), or such as become, have, etc. Second, SS i+1 con-
tains no noun or is an empty segment. Third, SPS i+1 contains the participle verb. The
case 3 is checked for the passive voice. The case 4 is performed when the end word of
SS i is “to” or “of”.
In summary, the VISAT algorithm can generate more meaningful termsets, be-
cause it considers the syntactic structure between words more in detail. In the CFF
process, the process Bag-of-words generates four types of outputs which are TFIDF
vectors. The VISAT process generates two types of outputs which are termsets com-
bined by two words. In total, these six types of outputs of the CFF process will be the
inputs for some classifications we proposed to judge the possible conflicting Engi-
neering Parameters in later phase.

4 Performance

In this section, we mainly conduct several experiments to evaluate the performance of
our method. We set the parameter of the similarity threshold between 0.7 and 0.8,
according to our observation. There are totally 39 Engineering Parameters in the
TRIZ theory originally, however some Engineering Parameters represent very similar
concept, and we thus group some similar Engineering Parameters into some new
classes as shown in Table 1.
The used training dataset includes only 115 patents collated manually by the re-
searchers who were from the National Taiwan University Department of Mechanical
Engineering. The distribution of the conflicting Engineering Parameters and the tech-
nical contradiction are imbalance. Most contradiction categories will only appear one
time, thus we cannot classify these patent documents on the contradiction directly.
We conduct two experiments to evaluate and discuss the performance of our
method. In our algorithm, the performance of final experiment results of MCIVC is
highly relevant to the performance of the classification result of Possible Conflicting
Engineer Parameters, thus we conduct the experiment to evaluate the performance
from the point of view of multi-label classification problem in this intermediate phase.
We also conduct another experiment and adopt different performance measurement to
evaluate the final performance of MCIVC, where it is a single-label classification
problem.