A social media monitoring platform used by business clients has to face interesting and sometimes unexpected issues arising from real texts processing, in particular dealing with the task of sentiment analysis of word-of-mouth communication. In this paper we describe some of the solutions adopted by BlogMeter, a proprietary listening platform that helps agencies and brands to discover what is said online about brands, people, topics and companies. We present some real life case studies, some of the linguistic resources used in the semantic annotation pipeline, and we suggest some topics for future investigations.
A social media monitoring platform used by business clients has to face
interesting and sometimes unexpected issues arising from real texts processing, in
particular dealing with the task of sentiment analysis of word-of-mouth
communication. As everybody knows, Sentiment Analysis (SA) is both a topic in
natural language processing which has been investigated since several years and
a tool for social media monitoring which is used in business services. Two
classical and essential references on this topic are [
In this paper we will use the term "sentiment analysis" as a broad term, that includes the narrower terms "opinion mining" and "mood". When we use "opinion mining" we refer to the identi cation of a belief or estimation or judgement expressed upon an object or target (a comment upon something, in simple words). When we use "mood" we refer to the mood state or emotional state communicated in a portion of text.
We will describe some of the solutions adopted by BlogMeter, a proprietary listening platform, and we will present some real life case studies and some of the linguistic resources used in the semantic annotation pipeline.
The paper is organized as follows. Section 2 brie y describes the Blogmeter platform. Section 3 presents the annotation pipeline. In section 4 we touch upon two interesting topics in SA and in section 5 we presents case studies coming from real-life application of sentiment analysis. 2
BlogMeter4 is a social media monitoring service operating since 2009 and used by
private and public companies in order to collect consumer and market insights
from social media and conversations taking places through them ([
In this paper we will focus on the Understanding phase, which includes automatic classi cation and SA. In detail it consists of: { creation of a domain-based taxonomy (i.e. an ontology of brands, products, people, topics); { identi cation and automatic classi cation of relevant documents (according to the taxonomy); { sentiment evaluation and opinion mining (automatic or supervised).
The monitored sources are typically user-generated media, such as blogs, forums, social networks, news groups, content sharing sites, sites of questions and answers (Q&A), reviews of products / services, which are active in many countries and in di erent languages. The overall number of sources is more than 500,000 blogs (of which approximately 70,000 active, with a post in the last three months) and 700 gathering places (forums, newsgroups, Q&A sites, content sharing platforms, social networks). This computation considers Facebook as a single source, but in fact, it is the largest collector of conversations (the system monitors the public status updates and the production of over 4,000 Italian o cial pages). We also consider web services like Instagram, Google+, Tumblr, Twitter or sectoral services like Foursquare or TripAdvisor. On the average, every day the system analyzes the following number of "documents": 4 www.blogmeter.eu { 3.7 million post retrieved from web sources; { over 2 million interactions from 1,000 Twitter business pro les and 4,000
Facebook business pages. 3
Documents extracted from the web in the form of unstructured information
are made available to the semantic annotation pipeline which analyze and
classify them according to the domain-based taxonomies de ned for the client. The
annotation pipeline uses the UIMA framework (the Unstructured Information
Management Architecture originally developed by IBM and now by the Apache
Software Foundation [
In this section we describe the "sentiment lexicon" used by the semantic
annotator, i.e. the repository containing terms, concepts and patterns used in the SA
annotation. Researchers have been building sentiment lexica for many years, in
particular for the english language, and a review on recent results can be found
for example in [
Our sentiment lexicon for Italian contains about 10.000 entries (6.200
single words and 3.400 multi-word expressions); each entry has information about
sentiment, i.e. polarity, emotions, and domain application. It has been created
and updated during the past three years, performing social media monitoring
and SA in di erent application domains. Recently, an italian lexicon for
sentiment analysis (Sentix) has been developed by [
The semantic annotator is a pattern matching component, which uses the sentiment lexicon, operates on the previous linguistic annotations and creates the corresponding sentiment concepts. The annotator can therefore recognize multi-word expressions that don't explicitly convey polarity and emotions but are related to concepts that do. 4
Social media and users' opinions and mood states are increasingly linked. Social networks were born as a means of interaction and places for sharing contents; now it is widespread the desire to share emotions and opinions quickly and with as many people as possible. An example of a highly chatted domain on the web is Social TV, as people love expressing their opinions about TV hosts and participants. 4.1
"Tre tweet a tuo favore su diecimila? Troiano sei un ottimista!" ("Three tweets out of ten thousand in your favor? Troiano you are an optimist!") "E come ogni gioved il solito interrogativo: Troiano, perche?" ("As every Thursday the same question: Troiano, why?") "Troiano migliorato? Non ho piu parole." ("Troiano improved? I have no more words.")
In the same application domain we detected a high percentage of correspondence between ironic tweets and questions: actually, since this TV program is not a cultural show in which questions and answers are the fundamental part, the ironic nature of questions co-occurring with a NE could be taken for granted.
Before proceeding with the identi cation of algorithms for the automatic
recognition of irony, we chose to focus on the in-depth knowledge of speci c
domains through the research of recurring elements in order to understand how
those domains work. In future developments we will test the validity and
representativeness of examples like those reported above.
The interest for emotion detection in social media monitoring grew in 2011 after
the publication of the paper [
The de nition of a set of basic (or primary) emotions is a debated topic, and
the study and analysis of emotions and their expression in texts obviously has
a long tradition in philosophy and psychology (see for example [
An interesting task we are investigating is trying to understand which kind
of relationship does exist between emotions and irony ([
The manual annotation of emotions in a reference italian corpus would be a useful advance for testing the accuracy of the automatic system. 5
In this section we present some case studies and charts that visualize mood and opinion trends generated in di erent contexts. 5.1
As seen before, one dimension of the mood analysis is the main polarity expressed in a text. Blogmeter mood analysis has been used: { as a gauge for the common feelings expressed through a peculiar social network and/or in a given span of time. The gure 1 for example shows the mood expressed in italian Twitter in the period January-April 2013. { as a marker of the general mood during speci c events. This kind of indicator was very useful during the tracking of live TV shows, due to its capacity to highlight positive and negative peaks on the social network in relation with a show's progression. 5.2
When the goal becomes more speci c and the need is linking a speci c subject (i.e. a target that could be a brand, role model or personality) with its related opinions throughout the post, sentiment analysis allows you to automatically examine thousands of messages in depth. Blogmeter's opinion mining has been applied for di erent industries, such as politics, banking and telecom, where the buzz has quite high volumes and at the same time very polarized opinions.
Another advanced application is near real-time semantic alerting. When sentiment analysis uncovers critical messages they are automatically labelled as negative and sent by email to those who can promptly intervene. This is important for instance in the transportation industry, where users need frequently updated information and feedback. 5.3
Recently there has been a growing interest for emotion analysis. This kind of investigation can be very useful when the two poles, negative and positive moods, have results that are too streamlined to explain more complex feelings. For instance, this analysis has been used to explore the emotions behind social, political and natural events like the Italian earthquake in 2012 (Fig. 3).
Semantic analysis can also be powerful in order to detect additional meanings, which are not covered in the mood/opinion/emotion dimension, expressing other kinds of people attitudes. In one of these applications, Blogmeter worked on clear voting intentions expressed on the social web, searching for declarations like "Ill vote X" or "I'll choose Y" (and not just "I like Z"). During the nal days of the last Italian political campaign, the analysis revealed the striking rise of Beppe Grillo's party. 6
We presented Blogmeter, a social media listening service that provides interesting insights about common feelings expressed in social media, opinions about speci c subjects and declared attitudes towards real actions or events.
We showed how, in order to achieve those results, it is important to exploit the potential of a well structured linguistic annotation pipeline, but also a domainspeci c concept-level sentiment lexicon (also called "contextualized sentiment lexicon" in the literature).
We also presented some case studies, as examples of mood, opinion and emotion recognition in real life use cases. We leave the issue of automatic recognition of irony for further investigation. we hope to have the opportunity to compare the accuracy of Blogmeter system with other ones using an o cial italian corpus for sentiment analysis (such as SentiTUT).
We would like to thank Sacha Monotti Graziadei, Vittorio Di Tomaso and Vincenzo Cosenza for always stimulating and leading new researches. Eugenia Burchi and Meghan White for their fundamental supervision of the paper; Matteo Casu for the essential help in its preparation. Last but not least, all Blogmeter colleagues for always giving their daily contributions.