English. In this paper we present the FBNEWS15 corpus, a new Italian resource for sentiment analysis and emotion detection. The corpus has been built by crawling the Facebook pages of the most important newspapers in Italy and it has been organized into topics using LDA. In this work we provide a preliminary analysis of the corpus, including the most debated news in 2015.
The use of Social Networks (SN) platforms like
Facebook and Twitter has developed
overwhelmingly in recent years. SN are exploited for
different purposes ranging from the sharing of contents
among friends and useful contacts to the
newsgathering about different domains such as politics
and sports
Several recent studies provide insights on how
the popularity of blogs and other user generated
content impacted the way in which news are
consumed and
Mass information is gradually moving towards
general platforms, and official websites are losing
their lead position in providing informatio
Facebook is one of the most popular online SN
in the world with 1 billion active users per month
and it offers the possibility to collect data from
people of different ages, educational levels and
cultures. From a linguistic point of view, previous
studies
In this paper we present the FB-NEWS15 corpus, a new Italian resource for sentiment analysis and emotion detection. The FBNEWS15 corpus can be freely downloaded at colinglab.humnet.unipi.it/resources/under the Creative Commons Attribution License creativecommons.org/licenses/by/2.0.1
The debate among users in commenting news and posts on Facebook offers a lot of subjective material to study the way in which people express their own opinions and emotions about a target event. In fact, in FB-NEWS15 we find linguistic items expressing the whole range of positive and negative emotions. In analyzing a news corpus, however, it is not simple to aggregate the posts on the basis of a certain fact, since several posts relate to the same event. For this reason, we decided to organize the corpus into clusters of topically related news identified with Latent Dirichlet Allocation (LDA: Blei et al. (2003)). This approach allow us to infer the most debated news in the corpus, and, in a second step, to discover the readers’ sentiment about a particular topic.
The paper is organized as follows: Section 2 describes the creation of the corpus, from crawling (2.1) to linguistic annotation (2.2), and finally provides basic corpus statistics (2.3). Section 3 reports on the automatic topic extraction with LDA. 2
For the creation of the corpus we followed the most important Italian newspapers. Since we were interested in building a corpus as heterogeneous as possible, we decided to focus on major newspapers with different political orientations, and which have in general heterogeneous readers.
Facebook allow users to post states, links, photos and videos on their own wall. In general, users can be divided into two macro-categories: People and Pages. People are often individuals, and the interaction with them is usually bidirectional (user A can read what user B publishes if A and B have a friendship relation). Conversely, Pages are typically used to represent organizations, public figures (web stars), companies or, as in our case, newspapers. In this case, the relationship is unidirectional, in the sense that user A can access the timeline of the page P by putting a ”Like” on P. Unlike a single-user, who usually publishes photos, videos and links about his private life, the timeline of a newspaper Facebook page, in general contains news titles with a link to the official website of the newspaper, where the user can read the 1All data collected have been processed anonymously for scientific purposes, without storing personal information. full article. The corpus keeps tracks of the threefold hierarchical structure of Facebook, which includes the news posts by the newspaper, the users’ comments to the posts and the replies to the comments. In this context, it is clear that the emotive content of the post is often neutral, but this post can inspire long discussions among readers, which can become useful material for sentiment analysis and emotion detection. Figure 1 shows a post, with some of its comments and replies.
In order to create the FB-NEWS15, we decided to download the timeline of the following newspapers, from 1 January 1 to 31 December 2015: La Repubblica, Il Giornale, L’Avvenire, Libero, Il Fatto Quotidiano, Rainews24, Corriere Della Sera, Huffington Post Italia. Facebook offers developers Application Programming Interfaces (APIs) for creating apps with Facebook’s native functionalities. In order to develop the crawler, we exploited the Graph API, which provides a simple view of the Facebook social graph by showing the objects in the graph and the connections between them. The Graph API allows us to navigate through the graph of the social network, which is organized into nodes <doc user="<newspaper(string)>" id="<id_post(string)>" type="post" parent_post="" parent_comment="" date="AAAA-MM-DD HH:MM:SS" location="" likes="662" comments="54" shares="322"> Un business truffaldino [E ora finitela con l’eco-balla dei controlli sulle emissioni] </doc> (Users, Pages, Photos and Comments) and Edges (Connections such as Friendship or Likes). The graph is navigated by exploiting HTTP requests, that may be implemented using any programming language. The native APIs offered by Facebook has some drawbacks: i) the maintenance of the app, since the APIs change over time, making it necessary to update the code of the crawler; ii) only public data can be accessed without requiring the user’s consent; iii) Facebook places limitations on the number of requests through a given period of time. For each post, comment and reply, we stored the message (text), the story (presence of photos and links tags), its timestamp, the type (post, comment, reply), the parent post/comment, the number of likes, shares and replies (Figure 2). 2.2
A very basic preprocessing phase has been
applied to the corpus before linguistic annotation,
to replace urls with the tag URL . The text has
been subsequently feed to a pipeline of
generalpurpose NLP tools. In particular, it has been
POS-tagged with the Part-Of-Speech tagger
described in
Except for Avvenire and Rainews24, for which we downloaded very few data, the other newspapers are attested in the corpus in a balanced way. In general, the number of posts is very low compared to the number of comments and replies. The average number of posts for each newspaper is 27,341.25, while for comments and replies is respectively 2,016,243.38 and 576,498.5. Table 1shows the number of texts (including posts, comments and replies) in FB-NEWS15 for each Newspaper and Figure 2.3 shows their cumulative distribution for each Newspaper.
NEWSPAPER
La Repubblica
Avvenire
Il Giornale
Libero
Il Fatto Quotidiano
Rainews24
Huffington Post
Corriere della Sera
OVERALL
FB-NEWS15 contains texts referring to a large
variety of events. In order to organize the
corpus into clusters of thematically related news, we
used LDA
In order to determine the most debated topics in 2015, we used LDA to assign 50 topics to the posts in FB-NEWS15 posts and we navigated the graph to assign the topics to the comments and the replies. Later, we restricted the topics associated to a post P to the topics T having a probability higher than the 90th percentile of the topic distribution of P . In this way, each post has been assigned, on average, to 3.06 topics. Finally, comments and replies have inherited the probability of belonging to the topic T from their parent post. Among the extracted topics ranked according to the sum of these probabilities we can find national and foreign politics, terrorism and church but also food, football, cinema and weather forecast. We report some topics below, with the number of texts and the relative ranking (i.e., rank 1 is given to the topic with the higher number of texts).
NATIONAL POLITICS (2,516,640 TEXTS, RANK 1): fRenzi, presidente, premier, Mattarella, riforma, Alfano, senato, camera, Boschi, aulag (Renzi, president, Mattarella, reform, Alfano, senate, chamber, Boschi, hall) SCHOOL (1,707,145 TEXTS, RANK 2):fscuola, giovane, studente, protesta, corso, mancare, sospendere, inglese, spiegare, lezioneg (school, young, protest, class, lack, suspend, English, explain, lesson) CRIME (1,543,735 TEXTS, RANK 7): fuccidere, polizia, arrestare, fermare, sparare, uomo, poliziotto, colpo, ferire, agenteg (kill, police, detain, stop, open fire, man, policeman, bump, wound, police officer) ISIS (1,267,749 TEXTS, RANK 16): fIsis, guerra, siria, minaccia, U.S.A., Libia, colpire, islamico, usare, jihadisti g (Isis, war, Syria, threat, U.S.A., Libya, damage, islamic, use, jihadist) FOOD (949,520 TEXTS, RANK 40): fmangiare, ricetta, cibo, preparare, consiglio, evitare, perfetto, trucco, salute, sempliceg (eat, FOOTBALL (606,560
TEXTS, RANK 50): fseguire la diretta, guardare il video, campo, calcio, serie, Napoli, Milan, segnare, battere, partitag (follow the live, look at the video, football field, football, league, Naples, Milan) 4
As one of the most widespread social networks, Facebook offers the possibility to collect opinionated pieces of texts from people of different ages, cultures and education. The composition of FBNEWS15, in which each comment is explicitly associated with a particular post, allows us to study the differences in terms of readers’ perceptions about a particular topic. Differently from other social media like Twitter, Facebook contains larger texts including lot of subjective expressions that are very useful for the construction of sentiment and emotive lexicons.
Starting from previous works