HHU at SemEval-2017 Task 5: Fine-Grained Sentiment Analysis on Financial Data using Machine Learning Methods

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 832–836, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics HHU at SemEval-2017 Task 5: Fine-Grained Sentiment Analysis on Financial Data using Machine Learning Methods Tobias Cabanski, Julia Romberg, Stefan Conrad Institute of Computer Science Heinrich Heine University D¨usseldorf D-40225 D¨usseldorf, Germany tobias.cabanski@hhu.de {romberg,conrad}@cs.uni-duesseldorf.de Abstract In this Paper a system for solving SemEval-2017 Task 5 is presented. This task is divided into two tracks where the sentiment of microblog messages and news headlines has to be predicted. Since two submissions were allowed, two dif- ferent machine learning methods were de- veloped to solve this task, a support vec- tor machine approach and a recurrent neu- ral network approach. To feed in data for these approaches, different feature extrac- tion methods are used, mainly word rep- resentations and lexica. The best submis- sions for both tracks are provided by the recurrent neural network which achieves a score of 0.729 in track 1 and 0.702 in track 2. 1 Introduction Analysing texts from the finance domain is a task that can help market traders to make important decisions because research has shown that senti- ments and opinions can affect market dynamics (Goonatilake and Herath, 2007). In the mean- time, the internet contains a huge corpus of fi- nance news from news websites or social media platforms. Natural language processing methods can be used to analyse this data as, for instance, sentiment analysis. To improve the understand- ing of the special characteristics of the domain, SemEval-2017 provides Task 5. 2 Related Work In the task of sentiment analysis, machine learning methods are widely used. One approach is shown in (Agarwal et al., 2011) where a sentiment anal- ysis on twitter messages is performed by combin- ing different features. It was found out that the use of multiple features processed by a support vector machine leads to good classification scores. The work of (Yadav, 2016) shows that recurrent neural networks can provide a good performance for this task. A powerful system can be created even with doing only a little preprocessing on the text data. 3 Task Description The SemEval-2017 Task 5 is divided into two tracks which each consider a different data basis (Cortis et al., 2017). The objective of both tracks is the prediction of a sentiment score with refer- ence to a company or a stock in a given piece of text. The sentiment score is a number within the interval [−1, 1] with −1 denoting a very bearish sentiment and +1 denoting a very bullish senti- ment. Track 1 is focused on microblog messages about stock market events. The data corpus consists of 1710 annotated messages taken from StockTwits1 and Twitter2, whereas a cashtag and the related spans are given in each case. Track 2 refers to financially relevant news head- lines. The annotated data given for the system’s training comprises 1156 headlines. For every headline-score pair the corresponding company name is specified. Spans are not …

Content truncated to 500 words