Figure 2. The structure of MANN in ANFIS.
there, 𝑦 is the output vector of MANN, and 𝑦̅𝑘and 𝑦̃𝑝are two consecutive maximal elements of 𝑦,
i.e.
𝑦̅𝑘 = max 𝑦𝑖 , 𝑘 = 𝑎𝑔𝑟 max 𝑦𝑖
1≤𝑖≤𝑁
𝑦̃𝑝 = max
1≤𝑖≤𝑁
𝑦𝑖.
1≤𝑖≤𝑘−1;𝑘+1≤𝑖≤𝑁
|
Correct (%)
|
Rejection (%)
|
Error(%)
|
∆2= 0.8, ∆3= 0.5
|
76.66
|
18.84
|
2.5
|
∆2= 0.5, ∆3= 0.8
|
85.77
|
8.62
|
5.61
|
No restriction
|
91.66
|
0.01
|
8.33
|
Table 3 displays the findings of ANFIS's identification of subjectivity in restaurant domain reviews using various ∆ 2and ∆ 3values.
Table 3. The average results of a tenfold cross validation accuracy ANFIS based on 𝑻𝑭 ∙ 𝑰𝑪𝑭 for detecting subjectivity in restaurant domain reviews.
ANFIS has a greater accuracy (91.66%) than FCS (91.3%) at the expense of requiring extra vari- ables in the intermediate layer of the neural network.
Conclusion. We have defined and implemented two distinct classification system architectures, FCS and ANFIS, to the identification of sentence-level subjectivity in a database of restaurant domain reviews. We have particularly shown how to train and evaluate these approaches for objective or sub- jective categorization of texts. The objective of the study was to develop procedures that did not need lin- guistic expertise and were thus applicable to any language. The feature extraction procedure is a crucial component of these approaches. Without language-specific limitations, we focused on the study of useful aspects that increase the accuracy of the systems. As a consequence, a unique “Pruned ICF Weighting Function” with a parameter determined particularly for the subjective data set was developed.
When comparing the present system to others, it is important to note that language understanding does enhance accuracy. Since we do not utilize such information, our findings should only be compared to other techniques with comparable limitations, such as those that evaluate features based on bags of words on the same data set. Studies by Pang and Lee (2004) as well as Martineau and Finin (2009). On the same data set, Pang and Lee claim a 92% classification accuracy for sentence-level subjectivity using Nave Bayes classifiers and a 90% classification accuracy using SVMs. Using the SVM Difference of TFIDFs, Martineau and Finin (2009) showed 91.26 percent accuracy. FCS (91,3 %) and ANFIS (91,7 %) have comparable outcomes as of this moment. Nevertheless, our provided techniques have certain benefits. FCS is the quickest algorithm among supervised machine learning techniques since function(5) is minimized exclusively with regard to (𝑦 = (𝑦1, 𝑦2, … , 𝑦𝑁)) (in the given issue N=2).
References:
C.Lin, Y.He, R.Everson. «Sentence Subjectivity Detection with Weakly-Supervised Learning». Proceedings of 5th International Joint Conference on Natural Language Processing, pp. 1153‒1161, 11, 2011.
A.Mogadala, V.Varma. «Language Independent Sentence-Level Subjectivity Analysis with Feature Selection,» Proceedings of the 26th Pacific Asia Conference on Language, Information, and Computation, pp. 171 ‒ 180, 11, 2012.
J.Martineau, T.W.Finin. «Delta TFIDF: An Improved Feature Space for Sentiment Analysis», ICWSM, 2009.
R.Fuller, Neural Fuzzy Systems, Springer Charm, 1995.
B.Pang, L.Lee, «A Sentimental Education: Sentiment Analysis Using Subjectivity,» Pro-
ceedings of ACL, pp. 271 ‒ 278, 2004.
G.Murray, G.Carenini. «Predicting Subjectivity in Multimodal Conversations,» Proceedings of
the 2009 Conference on Empirical Methods in Natural Language Processing, pp. 1348‒1357, 8, 2009.
S.Mukund, R.Srihari. «A Vector Space Model for Subjectivity Classification in Urdu aided by Co-Training,» Coling 2010: Posters, pp. 860 ‒ 868, 8, 2010.
J.Martineau, T.Finin. Delta TFIDF: An Improved Feature Space for Sentiment Analysis, 2009.
G.Grefenstette Y.Q., D.A.Evans, J.G.Shanahan. «Validating the Coverage of Lexical Resources for Affect Analysis and Automatically Classifying New Words along Semantic Axes,» Computing Atti- tude and Affect in Text: Theory and Applications, pp. 93‒107, 2006.
F.Benamara, B.Chardon, Y.Mathieu, V.Popescu. «Towards Context-Based Subjectivity Ana- lysis,» Proceedings of 5th International Joint Conference on Natural Language Processing, pp. 1180 ‒ 1188, 11, 2011.
Do'stlaringiz bilan baham: |