Browsing by Subject "Text classification"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Publication
    Investor sentiment in blogs : design of a classifier and validation by a portfolio simulation
    (2016) Klein, Achim; Kirn, Stefan
    How can investment recommendations available on the web significantly improve stock selection? This dissertation shows how online investment recommendations can automatically be analyzed, aggregated, and used to achieve a return above the market’s. To this respect, it is crucial to understand how investment recommendations affect returns. Therefore, the dissertation examines the effects of direct and indirect investment recommendations from blogs in the form of investor sentiments (i.e., opinions) on the expected development of stock prices. Blogs have made it possible for everyone to publish articles on the web. The studied blog platforms Seekingalpha and Blogspot host a wealth of semi-professional stock analyses, investor opinions, company rumors, and stock recommendations. The dissertation’s study uses about 77,000 articles from Seekingalpha and about 198,000 articles from Blogspot over a five-year period (2007-2011). A novel text classification method is developed for the automatic classification of blog articles in a positive vs. negative sentiment. To achieve a high classification accuracy, experiments were carried out to configure this method. The text classification method uses machine learning techniques, which learn from manually classified articles from a novel corpus. Using behavioral finance theory, hypotheses are developed about the effects of investor sentiments on a portfolios returns. To test these hypotheses, a monthly selection of stocks of the Dow Jones Industrial Average into a portfolio was simulated (i.e., backtested). The selection is made by means of the ranking of the monthly aggregated overall sentiment of all articles regarding a specific stock. The results show that a return above the market’s can be achieved with aggregated investor sentiments from the Seekingalpha platform. In most cases, the achieved return exceeds the return of a momentum portfolio based solely on past returns. For the platform Blogspot, results are weaker. Overall, it seems advisable for investors to select a small number of stocks based on the most positive and most negative monthly investor sentiments from professional blogs.
  • Thumbnail Image
    Publication
    Sentiment analysis in electronic negotiations
    (2017) Körner, Michael; Schoop, Mareike
    The thesis analyzes the applicability of methods of Sentiment Analysis and Predictive Analytics on textual communication in electronic negotiation transcripts. In particular, the thesis focuses on examining whether an automatic classifier can predict the outcome of ongoing, asynchronous electronic negotiations with sufficient accuracy. When combined with influencing factors leading to the specific classification decision, such a classification model could be incorporated into a Negotiation Support System in order to proactively intervene in ongoing negotiations it judges as likely to fail and then to give advice to the negotiators to prevent negotiation failure. To achieve this goal, an existing data set of electronic negotiations was used in a first study to create a Sentiment Lexicon, which tracks verbal indicators for utterances of positive and, respectively, negative polarity. This lexicon was subsequently combined with a simplified, feature-based representation of electronic negotiation transcripts which was then used as training data for various machine learning classifiers in order to let them determine the outcome of the negotiations based on the transcripts in a second study. Here, complete negotiation transcripts were classified as well as partial transcrips in order to assess classification quality in ongoing negotiations. The third study of the thesis sought to refine the classification model with respect to sentence-based granularity. To this end, human coders were classifying negotiation sentences regarding their subjectivity and polarity. The results of this content analysis approach were then used to train sentence-level subjectivity and polarity classifiers. The fourth and final study analyzed different aggregation methods for these sentence-level classification results in order to support the classifiers on negotiation granularity. Different aggregation and classification models were discussed, applied to the negotiation data and subsequently evaluated. The results of the studies show that it is possible to a certain degree to use a sentiment-based representation of negotiation data to automatically determine negotiation outcomes. In combination with the sentence-based classification models, negotiation classification quality increased further. However, this improvement was only found to be significant for complete negotiation transcripts. If only partial transcripts are used – specifically to simulate an ongoing negotiation scenario – the models tend to behave more erratic and classifcation quality depletes. This result yields the assumption that polarized utterances (positive as well as negative) only carry unequivocal information (with respect to the outcome) towards the end of the negotiation. During the negotiation, the influence of these utterances becomes more ambiguous, hence decreasing classification accuracy on models using a representation based on sentiments. Regarding the original goal of the thesis, which is to provide a basic means to support ongoing negotiations, this means that supporting mechanisms employed by a Negotiation Support System should focus on moderation techniques and resolving of potentially conflicting situations. Approaches that could be used to employ further conflict diagnosis in interaction with the negotiators are given in the final chapter of the thesis, as well as a discussion of potential recommendations and advice the system could give and lastly, approaches to visualize the classification data to the negotiators.