The detection of emotional connotations in texts is a recent task in computational linguistics. Its economic stakes are promising; for example, a company could detect, by analyzing the blo-gosphere, people's opinion on its products.

The goal of the SemEval task is to annotate news headlines for emotions (using a predefined list: anger, disgust, fear, joy, sadness & surprise), and for valence (positive or negative). A specific difficulty here is related to the small number of words available for the analysis.

Our system is mainly rule-based and uses a lin­guistic approach. From a macroscopic point of view, we follow the hypothesis that, in a news title, all the words potentially carry emotions. If linguis­tic resources make it possible to detect these emo­tions individually, how can we deal with headlines where bad and good emotions appear at once?

Our objective is to identify the expression which carries the main topic of the title. One can consider that this expression has a primary importance.

We also seek to lay down rules for detecting specific emotions. For instance, surprise some­times comes from the contrast between good and bad news. And sometimes, simple lexical elements are characteristic of an emotion; a negation or a modal auxiliary in a title may be a relevant indica­tor of surprise.

We describe here the techniques we imple­mented to address all these points.

A preliminary problem that we had to solve was related to the Anglo-Saxon habit of putting initial capital letters in all the words of a title.

The first pass of our system thus detected news titles that were "improperly" capitalized, and "de-capitalizes" their common words.

For that, we used the SS-Tagger on the title; ac­cording to the part of speech of each word, infor­mation found in WordNet, and some hand-crafted rules, the system chooses or not to keep the initial.

The impact of this processing step is far from negligible, from the point of view of the Stanford Parser. Indeed, let us have a look at the difference between the parsing of the title, before (figure 1) and after (figure 2) this processing.

For the moment, we consider the output of the Stanford Parser as an array of PoS-tagged words. We use WordNet's morphology functions to find the possible base form of each word.

At this stage, an important question arose: was lexical disambiguation possible? We thought not, because with short sentences, few relevant heuris­tics apply. We chose another solution, by consider­ing that the emotion and valence values of a word were the linear combination of that of all its possi­ble meanings, balanced by the frequency of each lemma.

We detected emotion and valence values for each word, by using our enriched version of

WordNet-Affect and SentiWordNet.

In fact, we also detected some extra information:

• An additional 7th emotion, that looks like "compassion for people needing protection". Our assumption is that certain words express a subjacent need for protection. For exam­ple, there is "student" behind "school", and "child" behind "adoption". So, we built a list of words designating something that needs protection; we also include in this list words such as "troops", "shoppers".

• We tried to detect acronyms relating to technology; for this, we defined a list of high-tech companies and a very basic regu­lar expression rule saying that a word (not in WordNet) containing numbers, or capitals not in first position, should be something high-tech. (This very basic rule seems to

work nicely on PS3, iPod, NASA.). We use these high-tech indications to increase the "joy" emotion.

• We counted lexical elements that we think are good indicators of surprise: negations, modal auxiliaries, question marks.

At this stage, we begin some post-processing on individual words. Which factors cause anger rather that sadness? We believe that human intention (to harm) causes the former emotion, while natural factors such as disease or climatic catastrophes cause the latter. So, we used a few rules related to the WordNet noun hierarchy, based on the fact that when a noun is a hyponym of a given synset, we boost some emotions:

At this stage, our system tries to find the main subject of the news title. Again, we use the output of the Stanford Parser, but this time, we make use of the dependency graph. We consider that the main word is the root of the dependency graph, i.e. the word that is never a dependant word. (For instance, in figure 2, the main word is "predicts".)

We think that the contribution of this main word is much more important than that of the other words of the title. So, we multiply its individual valence and emotion by 6.

The last important part of linguistic processing is the detection of contrasts and accentuations between "good" or "bad" things. We search pat­terns like [noun->subject->verb] or [verb-^direct object-^noun] in the dependency graph, with verbs that increase or decrease a quantity. Using the valence of the given noun, this gives our system the ability to detect very good news ("boosts (brain) power") or good news where something bad gets less important ("reduces risk", "slows decline", "hurricane weakens".).