Hello, you could also write what is the purpose of your classification, how do you measure its success, how successful you are, and what do you expect to improve.

Generally, there are a lot of steps and possibilities in text classification, including data collection, preparation, transformation, processing with classification algorithms, changing the parameters of these algorithms etc. It is therefore not easy to give you a specific and reliable answer.

Hi Munawar,

You can use Random Forest or other Weka or R or Rattle techniques

Best wishes,

Osman

I do not suggest SVM, at least I havent had good result with it (with python and Nltk). 

accuracy of classification is very much depends on the quality of the input data. make sure you elaborate the pre processing and remove the junk data out of your corpora.

In my opinion, there are two issues to consider: improving classification results as Precision and Recall and how to interpret classification results.

The most effective way to improve text classification numerically is to extract most optimal features. In text classification, these features are frequently BoW. However, depending on text to classify, there might be also other features that yield best classification such as stylometric, deictic, grammatical. In some text domains, it can be insufficient to interpret classification results using only Precision and Recall. I used in my thesis own heuristics: a classes-similarity measure, cost-based measure, classes-number measure.

You might be interested in my thesis where I present results of many experiments regarding choice of the classifier and extraction of the features (s. below).

Thesis Opinion Mining and Lexical Affect Sensing

It depends on the nature of the things you want to classify it. Generally, in text classification, in my opinion, the most influential factor to improve the accuracy is the quality of the features.

As for as you are using SVM try Linear kernel which performs better results in text classification if you have not already tested your model with different kernels.

Thanks

Munir

Please, check this link as user (not developer)

http://scikit-learn.org/dev/_downloads/scikit-learn-docs.pdf

Dear Munawar

I suggest doing preprocessing step first to clean your data then apply RBF kernel to transform your data to high dimension space after that apply twin SVM classifier and you will get a better result.

Regards,

Please read this paper,

Conference Paper HDLTex: Hierarchical Deep Learning for Text Classification

With Autoweka most probably you get good results and an indication of the best algorithm to use. Bad is that you cannot replicate the results manually in Weka.

First step is to clean your data. If you are using one-hot encoding to represent your input data, please convert your sparse matrix to dense. Then try svm.

Hi there...Just wanna share my experience in text mining, but I'm not sure how my research application of text mining will help yours. In my research, there are a few factors that may influence classifier accuracy.

1. The pre-processing phase may decrease the dimensional of your word vectors;

2. If you are using term weighting scheme to weighted every single term, different scheme may produced different results;

3. The content of stop word list and root word list also may influence the accuracy of your classifier (in your case is SVM) to perform efficiently.

My research is applying algorithms from Danger Theory of AIS (DCA and dDCA) and aforementioned factors do influence classifiers to true classify at high rate of accuracy.

Hope that this will help u. Good luck.

Dear Munawar

By now you might have found a way to improve the outcome of sentiment analysis.

You have started with BoW . Good.

In Sentiment analysis, the words to be interpreted in conjuction with few scenarios:

- people use slang in expressing sentiments. This has to be caught and linked with corresponding english terms generally used.

- A dictionary maintenance of meta data could help

- Double negative identification to mean positive could be identified

- as suggested, weight could be associated with frequently used on powerful words.

- The context of discussion - say product review, movie review, news review etc to be dealt with associated descriptive terms

Best Wishes

Depend on:

- Preprocessing,

- Feature Extraction,

- Classification algorithm.

I think that you could try to perform Exploratory Data Analysis (EDA) first. After understanding the characteristics of what you are working with, you have to decide which data preparation and preprocessing steps are needed. You may need to apply any of the dimensionality reduction techniques and use some adaptive feature engineering techniques. Finally, you could try using the ensemble learning to enhance you performance.

By the way, Accuracy is not always reflects the efficiency of classification algorithms; other metrics could indicate the efficiency of the used classification algorithm rather than Accuracy. e.g. the AUC, Precision, Recall, Miss Rate, False Omission Rate, etc.

I Agree With Mohamed Elhadad

Use dense vectors - for example, word2vec/doc2vec are pretty efficient. If you use Python, install the package gensim for that.

Regarding classifiers, Linear SVC & Neural Networks usually work pretty well for text classification, or even more complex nn like LSTM.