Sentiment analysis has recently gained popularity in the nancial domain thanks to its capability to predict the stock market based on the wisdom of the crowds. Nevertheless, current sentiment indicators are still silos that cannot be combined to get better insight about the mood of di erent communities. In this article we propose a Linked Data approach for modelling sentiment and emotions about nancial entities. We aim at integrating sentiment information from di erent communities or providers, and complements existing initiatives such as FIBO. The approach has been validated in the semantic annotation of tweets of several stocks in the Spanish stock market, including its sentiment information.
The proliferation of user generated content in web sites and social networks,
such as Facebook, TripAdvisor or Twitter, has lead to an increased awareness
of the power of social networks for expressing opinions about products, services
and even disasters. These so-called social sensors enable real time indexing of
the social web with the aim of providing insight about the structure and activity
of social networks. They provide a vast array of application possibilities, from
monitoring brands or products to become early disaster warning systems [
In the nancial eld, social sensors can provide additional valuable
information that complements other sources of information used in fundamental analysis,
such as nancial newspapers. In particular, sentiment analysis has been one of
the most popular technologies to measure the investment mood. The sentiment
stock market indicator has become a popular indicator that is provided together
with the classical fundamental and technical stock market indicators [
fear-and-greed/ Sentiment Survey4, StockMarketSensor5, or SentimentTrader6, just to name a few.
In addition, recent research has shown that sentiment expressed in
microblogging sites such as Twitter can be applied to predict daily changes in stock
values [
Linked Data is another valuable resource that can provide nancial analysts
with an integration of available data sources in their activity [
The aim of this article is the application of a Linked Data approach to expressing sentiments and emotions about nancial concepts, which nancial analysts can use to combine opinions expressed in di erent social media sites.
The article is arranged as follows. Sect. 2 gives an overview of the vocabularies
we have de ned for modelling sentiment and opinions as well as its interlinking
with nancial vocabularies such as FIBO [
Modeling Sentiment and Emotions as Linked Data This section provides insight about the potential of Linked Data for accessing, interlinking and reasoning about business data sources. To leverage that power, it is necessary to have a robust representation model for sentiment in the nancial context. Rather than creating an ad-hoc model, the Linked Data approach is to look for models for each domain and connect them. In particular, we will need a model for nancial entities, a model for sentiment analysis results, and a model for microblogging messages. The following sections review the models (also referred to as ontologies or vocabularies) available in these domains, and Sect. 2.3 exempli es the use of the nal integrated model.
5 Available at http://www.stockmarketsensor.com/ 6 Available at http://www.sentimentrader.com/ 2.1
Financial Industry Business Ontology (FIBO) [
Financial Exchange Framework Ontology (FEF) [
The FP7 FIRST Project (Large Scale Information Extraction and
Integration Infrastructure for Supporting Financial Decision Making) has de ned an
ontology for sentiment analysis in nancial domains [
In its simple form, a FIBO de nition would be a single triple. However, FIBO is a complete ontology that enables much more powerful assertions, as will be shown later. 2.2
In this section we introduce two vocabularies, Marl and Onyx, that we have dened for providing a uniform vocabulary for expressing sentiments and emotions, respectively, according to linked data principles.
Marl [
this, Marl attempts to answer the research question of to what extent opinion information can be formalised in a uni ed way.
Marl is the result of analysing the properties that characterise opinions expressed on the web or inside various IT systems. The nal set of concepts proposed is shown in Fig. 1. It should be noted that opinions in Marl are meant to be linked to an entity. Such entity can be a FIBO Corporation, as described in the previous section. We will make use of this property in Section 2.3.
A detailed description of each particular property and an explanation of their meaning can be found in the vocabulary's speci cation 8.
Onyx [
At its core, the Onyx ontology has three main classes: EmotionAnalysis, EmotionSet and Emotion. In a standard Emotion Analysis, these three classes are related as follows: an EmotionAnalysis is run on a source (generally in the form of text, e.g. a status update), the result is represented as one or more EmotionSet instances that contain one or more Emotion instances.
The speci cation of the Onyx vocabulary 9 contains an updated description of all its elements, with some usage examples. 8 http://www.gsi.dit.upm.es/ontologies/marl 9 http://www.gsi.dit.upm.es/ontologies/onyx 2.3 First of all, let us review a simpli ed version of the integration of all the elements that we described. To keep it as simple as possible, we will avoid any provenance information (such as who or how analised the twit to extract the opinion) or information about the post itself (author, date, etc.) This simplicity will not prevent us from harnessing the potential of Linked Data.
Listing 1.1. Simple representation using FIBO ex : my Op in ion a marl : Opinion ; marl : h a s P o l a r i t y V a l u e marl : Positive ; marl : d e s c r i b e s O b j e c t ex : G S a n t a n d e r ; marl : e x t r a c t e d F r o m ex : twit1 . ex : twit1 a sioct : M i c r o b l o g P o s t ;
sioc : content " I like testing Grupo S an ta nd er ". ex : G S a n t a n d e r a fibo : I n c o r p o r a t e d C o m p a n y .
In this work, we have gathered thousands of posts from Twitter and stored them in a graph using a more complex version of this schema.
In order to provide a semantic representation of tweets, we have selected
TwitLogic [
However, the true potential of Linked Data comes into play when we use data from di erent sources. For instance, if there is another endpoint that contains opinions gathered from Twitter or other social networks, we can query their information seamlessly, provided they use Marl and FIBO as well.
If that example still seems uninteresting, we can also use disparate sources, such as DBpedia. DBpedia contains general information about many entities, which includes several corporations. To be able to query DBpedia, we just need to link our entities to a DBpedia entity. If we take our former example, this modi cation is as simple as:
Of course, this also involves named entity recognition techniques, which are covered in Section 3.2. Once this step is done, we can issue complex queries that answer questions such as: "What is the general opinion about Banks in Spain?", or "What is the relationship between year of incorporation and the number of opinions in social media?". Note that such queries could use advanced FIBO information, such as current contracts or date of incorporation.
ex : GSantander rdfs : seeAlso dbpedia : Santander_Group .
PREFIX sioc : < http :// rdfs . org / sioc / ns #> PREFIX marl : < http :// www . gsi . dit . upm . es / ontologies / marl / ns #> SELECT ? entity
COUNT (? negative_opinion ) AS ? negative_opinions
COUNT (? positive_opinion ) AS ? positive_opinions WHERE { { ? positive_opinion marl : describesObject ? entity .
? positive_opinion marl : hasPolarity marl : Positive . } UNION { ? negative_opinion marl : describesObject ? entity .
? negative_opinion marl : hasPolarity marl : Negative . } } GROUP BY ? entity 3
Financial Twitter Tracker Architecture
In this section we describe the architecture of a prototype, called Financial
Twitter Tracker, that we have developed for tracking the sentiment evolution
of nancial entities in Twitter. The core of the system is a semantic pipeline,
described below, where tweets are retrieved and analysed. As a result, tweets
are semantically annotated as stored in the semantic store Linked Media
Framework (LMF) [
The semantic pipeline for sentiment analysis consists of three tasks. First, the system connects to the Twitter API (Sect. 3.1) and retrieves tweets that match a list of prede ned keywords. Then, a semantic analysis (Sect. 3.2 is carried out. Finally the sentiment analysis is done (Sect. 3.3). 3.1
For the purpose of obtaining tweets we developed a wrapper over the services
o ered by the public Twitter API11, concretely method search bounded by dates
and keywords, which allows the retrieval of each and every tweet published within
a particular day and regarding a particular topic. Given the data set of study,
several related topics to nancial world { such as banking, telecommunication,
energy, to name a few { were established. Such data sets have been split according
to di erent languages in order to increase performance and accuracy within the
developed \sentiment analysis".
10 Available at http://github.com/gsi-upm/Sefarad
11 https://dev.twitter.com/docs/api/1.1/get/search/tweets
Data from Twitter is very heterogeneous, as it is used for di erent purposes
(e.g. reviews, factual data, personal comments), covering di erent categories
and subjects. Hence, it was necessary to carry out a categorization prior to the
data analysis itself. With such ltering in mind, the Support Vector Machine
system (SVM) was developed, taking into account the fact that it supports
high dimensional data [
As POS-tagging, Treetagger [
Another relevant aspect about the lexicon enrichment for its later storage and visualization was the extraction of entities by a NER based on Wikipedia, so that information is compared to the entities published by Wikipedia in order to work out the possible extraction from the text. Periodically the system brings the available information up to date with the new entries published on the online encyclopedia. Finally, that information is lined up with the nancial ontology FIBO to provide data in a standardized way in accordance with the semantic web principles such as RDF/OWL, allowing the integration in other technical systems that adapts the given standard. Thanks to FIBO it is possible to provide a clear meaning - without ambiguities - for the nancial terms. 3.3
The last stage of the pipeline is in charge of the sentiment and emotion analysis. With a view to quantify the \sentiment" the procedure is to perform the arithmetic mean considering all the registered values recognized in the tweet and using simple rules like inverters (i.e. not). The emotion eld can be extracted from the connection between triplets (aligned with WordNet 3.0) and WN-A ect. The outcome stems from the analysis of each tweet which was stored in a MongoDB NoSQL data base, which can handle high volume of information ful lling the big data requirements of twitter processing. 3.4
After the processing is done, all the triples are stored in an LMF instance,
which provides SPARQL and Solr [
Experimentation Throughout classi cation and Sentiment Analysis stages stages of the aforementioned pipeline, we performed experimentation with the obtained data. The classi cation step has been developed with an SVM trained for the recognition of two groups; nance and non- nance; which states whether the tweets are to continue to the next ow level or, on the contrary, are to be discarded.
Within Machine Learning there are two main discovery methods: supervised and unsupervised learning. In supervised learning, a series of manually tagged data are provided for the system training. On the unsupervised setting, it is the 12 http://d3js.org/ system itself that directly infers patterns from the raw information. The current project uses supervised learning: a random set of tagged tweets has been trained by experts in nances and added to the established groups.
The model has been trained 4 times by modifying the range of information
in order to measure and test the system. The rst approach makes use of 90% of
values for the training and 10% for the assessment, such proportions vary in the
second training to 80%-20%, 70%-30% for the third, and 60%-40% for the last
one [
Model training-Test 90%-10% 80%-20% 70%-30% 60%-40% Average precision 0,940 0,9393 0,9369 0,9290 Supported Vectors 886,4548 825,4787 757,501 674,8602
Table 1. Results using di erent training options
From these results we observe that the bigger the quantity of information used to train the model, the more precise is the outcome, and that value decreases as the volume of data saved for the assessment grows. However, it is remarkable that the more data is used to train the system, the greater is the number of supporting vectors, and, consequently, the classi er loses computational performance.
The Sentiment Analysis phase has been tested against a set of manually
annotated corpus. The evaluation was carried out in order to measure the
effectiveness. Such accuracy conforms the ability the system owns to satisfy the
feature it was developed for [
a c b d Expert Identi ed Not identi ed
F1 := 2
Recall := a +a c P recision := a +a b
c Pomissions := a + c
b Pfalsepositive := b + d
P recision Recall P recision + Recall (1) (2) (3) (4) (5) In this article we have presented a vocabulary for modelling sentiments and emotions. This vocabulary can be used to query opinions and emotions about nancial institutions and stock values across di erent web sites and information sources. The main advantage of this approach is that heterogeneous sentiment indexes can be easily integrated and used together with other vocabularies such as FIBO. We have evaluated these vocabularies in a sentiment analysis service based on Twitter for tracking nancial institutions.
As a future work, we are working on improving the visualisation and query capabilities of the interface so that non technical users, such as business analysts can take advantage of the possibilities that the Web of Data brings for exploring and consulting, sentiment about nancial institutions in large amounts of complex and heterogeneous data.
Another current line of research is the standardisation of these vocabularies for sentiment and emotion. With this aim, we are participating in the Linked Data Models for Emotion and Sentiment Analysis W3C Community Group, which takes as a baseline the vocabularies Marl and Onyx. 6
Acknowledgement This research has been partially funded by the Spanish Ministry of Industry, Tourism and Trade through the project Financial Twitter Tracker (TSI-0901002011-114) and the EUROSENTIMENT FP7 Project (Grant Agreement no: 296277)