FEUP at SemEval-2017 Task 5: Predicting Sentiment Polarity and Intensity with Financial Word Embeddings

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 904–908, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics FEUP at SemEval-2017 Task 5: Predicting Sentiment Polarity and Intensity with Financial Word Embeddings Pedro Saleiro1,2, Eduarda Mendes Rodrigues1, Carlos Soares1,3, Eug´enio Oliveira1,2 Faculdade de Engenharia da Universidade do Porto1, LIACC2, INESC-TEC3 Rua Dr. Roberto Frias, s/n, Porto, Portugal {pssc,eduarda,csoares,eco}@fe.up.pt Abstract This paper presents the approach devel- oped at the Faculty of Engineering of Uni- versity of Porto, to participate in SemEval 2017, Task 5: Fine-grained Sentiment Analysis on Financial Microblogs and News. The task consisted in predicting a real continuous variable from -1.0 to +1.0 representing the polarity and intensity of sentiment concerning companies/stocks mentioned in short texts. We modeled the task as a regression analysis problem and combined traditional techniques such as pre-processing short texts, bag-of-words representations and lexical-based features with enhanced financial specific bag-of- embeddings. We used an external collec- tion of tweets and news headlines men- tioning companies/stocks from S&P 500 to create financial word embeddings which are able to capture domain-specific syntac- tic and semantic similarities. The result- ing approach obtained a cosine similarity score of 0.69 in sub-task 5.1 - Microblogs and 0.68 in sub-task 5.2 - News Headlines. 1 Introduction Sentiment Analysis on financial texts has received increased attention in recent years (Nardo et al., 2016). Neverthless, there are some challenges yet to overcome (Smailovi´c et al., 2014). Financial texts, such as microblogs or newswire, usually contain highly technical and specific vocabulary or jargon, making the develop of specific lexical and machine learning approaches necessary. Most of the research in Sentiment Analysis in the financial domain has focused in analyzing subjective text, labeled with explicitly expressed sentiment. However, it is also common to express finan- cial sentiment in an implicit way. Business news stories often refer to events that might indicate a positive or negative impact, such as in the news title “company X will cut 1000 jobs”. Economic indicators, such as unemployment and future state modifiers such as drop or increase can also pro- vide clues on the implicit sentiment (Musat and Trausan-Matu, 2010). Contrary to explicit expres- sions (subjective utterances), factual text types of- ten contain objective statements that convey a de- sirable or undesirable fact (Liu, 2012). Recent work proposes to consider all types of implicit sentiment expressions (Van de Kauter et al., 2015). The authors created a fine grained sentiment annotation procedure to identify polar expressions (implicit and explicit expressions of positive and negative sentiment). A target (com- pany of interest) is identified in each polar expres- sion to identify the sentiment expressions that are relevant. The annotation procedure also collected information about the polarity and the intensity of the sentiment expressed towards the target. How- ever, there is still no automatic approach, either lexical-based or machine learning based, that tries to model this annotation scheme. In this work, we propose to tackle the afore- mentioned problem by taking advantage of unsu- pervised learning of …

Content truncated to 500 words