2,1 Conditional Model Approach

5In machine learning supervised problems are defined as a set of input called training examples (x⁽¹⁾, y⁽¹⁾) · ·· (x^(m),y^(m)), where each input x⁽ⁱ⁾ paired with a output label y⁽ⁱ⁾. In this task, our goal is to learn a function f : X → Y , where X and Y refers to the set of possible input and labels.

6For PoS tagging problem, each input represents a sequence of words , …, and labels be a sequence of tags , …, , where n_i refers to the length of i^th training example. In this machine learning each input x be a sentence of Italian language and each label be the possible PoS tag. We use conditional model to define the function f(x) and we define the conditional probability as

7for any x, y pair. We use training examples to estimate the parameters of the model and output of the model for a given test example x is measured as

8Thus we consider the most likely label y as the output of the trained model. If the model p(y|x) is close to the true conditional distribution of a labels given inputs, so the function f(x) will consider as an optimal.

2.2 Generative Model

9In this model, we use the Bayes’ rule to transform the Eq.1 into a set of other probabilities called generative model. Without estimating the conditional probability p(y|x), in generative model we use the Bayesian classification

10p (x, y)

11over (x, y) pairs. In this case, we further break down the probability p(x, y) as follows:

12p (x, y) = p (y) p (x|y)(2)

13and then we estimate the model p(y) and p(x, y) separately. We consider p(y) as a prior probability distribution over label y and p(x|y) is the probability of generating the input x, given that the underlying label is y.

14We use the Bayes rule to derive the conditional probability p(y|x) for any (x, y) pair:

15where

16We apply Bayes rule directly to a new test example x, so the output of the model f (x), can be estimated as follows:

17To simplify Eq.5, we use Hidden Markov Model (HMM) taggers with two simplifying assumptions. The first assumption is that the probability of word appearing depends only on its own PoS tag as follows:

18where p(wn1 tn1) means probably of tag ti with word wi. The second assumption is the probability of a tag appearing is dependent only on the previous tag, rather than entire tag sequence. This is known as bigram assumption and can be measured as follows:

19Further, we incorporate these two assumptions in Eq.5 by which a bigram tagger estimates the most probable tag as follows:

3.1 Dataset

20For the proposed task organizers re-uses the tweets being part of the EVALITA2014 SENTIPLOC corpus. Both the development and test set first annotated manually for a global amount of 4, 041 and 1, 749 tweets and distributed as the new development set. Then a new manually annotated test set, which is composed of 600 and 700 tweets were produced using texts from the same period of time. All the annotations are carried out by three different annotators. Further a tokenised version of the texts is also distributed in order to avoid tokenisation problems among participants and the boring problem of disappeared tweets.

3.2 Results

21For this task, total 13 runs were submitted 9 teams and among these runs 4 Unofficial runs also submitted. In Table 1 we list out all results for this task.

22Team 2, 3, 5 and 6 submitted one Un-Official run with compulsory one and these Un-Official submissions are ranked as 12^th, 3^rd, 7^th and 9^th respectively. We also listed these submissions in Table 1 with other runs. Our team NLP–NITMZ represent as Team8 and ranked as 11^th in this task.

3.3 Comparison with other submissions

23In this competition, a total of 4759 words were given for tagging purpose. These words were categories into 22 PoS tags and our team successfully tags 4091 words with 668 unsuccessful tags. The 1^st ranked team successfully tags 4435 words and the last positioned team i.e. Team9 successfully identified 3617 tags. In Table 2, we provide our system tag wise statistics.