Version classiqueVersion mobile

EVALITA Evaluation of NLP and Speech Tools for Italian - December 17th, 2020

 | 
Valerio Basile
, 
Danilo Croce
, 
Maria Maro
, 
et al.

SardiStance: Stance Detection

ghostwriter19 @ SardiStance: Generating new tweets to classify SardiStance EVALITA 2020 political tweets

Mauro Bennici

Résumé

Understanding the events and the dominant thought is of great help to convey the desired message to our potential audience, be it marketing or political propaganda.
Succeeding while the event is still ongoing is of vital importance to prepare alerts that require immediate action.
A micro message platform like Twitter is the ideal place to be able to read a large amount of data linked to a theme and self-categorized by its users using hashtags and mentions.
In this research, I will show how a simple translator can be used to bring styles, vocabulary, grammar, and other characteristics to a common factor that leads each of us to be unique in the way we express ourselves.

Note de l’éditeur

Copyright ©️2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).

Texte intégral

1. Introduction

1Each of us has a unique way of writing. However, the fewer options we have to experience ourselves to express our concept, the more the necessary synthesis leads to the loss of precious information to accurately assess our real intentions.

2Furthermore, the more the subject is debated, the more changes in style and tone occur. The conversation becomes full of irony or aggressive. Extrapolating a single line is dangerous without context. The same sentence can have different interpretations depending on the moment in which it is pronounced, the audience it is intended for, the place where you are, in the historical period in which it was composed.

3My hypothesis is that we can translate all these different styles into a single "language style" that fully expresses the real intentions of the writer. The challenge is to understand when a user has expressed a comment in favor, against, or neutral towards the Sardines' Italian political movement.

4The research was carried out for the SardiStance (Cignarella et al., 2020) task in the EVALITA 2020 (Basile et al., 2020). Two models were created for the Task 1, but they also performed well on the Task 2.

2. Description of the system

5The two tasks are similar. In Task A, it is necessary to classify the stance of a tweet based only on the text of the tweet. Task A is divided into two subtasks:

  • Constrained. It is allowed to use additional resources such as a Lexicon but no other resources (such as labeled tweets) to help the training process.

  • Unconstrained. Where each resource used must be reported in the final report.

6In Task B, you can use the context information provided by the post author. Additional information refers:

  • to post statistics (favors, retweets, reply, source)

  • to the author's information (number of posts, number of followers, emoji in the bio)

  • to the author's circle of relationships (friends, replies, retweets, and quotes)

7The research focuses on Task A Constrained.

8Considering the constraints of Task A, it is not possible to access any additional information other than the text of the tweet, I concentrated on understanding how to clean it up.

9The Training dataset contains:

  • the tweet ID

  • the user ID

  • the text

10the label The labels options are:

  • Against

  • Favor

  • Neutral / None

11To be sure to do not use any data except the text, the user id, useful for Task B, was discarded.

12In order to validate my hypotheses, I used the AlBERTo model, created from tweets, (Polignano at al., 2019) and an auto training system such as Ktrain1, a framework that wrap TensorFlow2, to classify the tweets. To avoid manual error and involuntary optimization, I used the autofit option.

13First, I wrote a series of algorithms to make the texts to be compared homogeneous.

14The first one was to break up the composed hashtags into sentences and words.

15For example, using capital letters as a separator:

  • #IoStoConLeSardine has become "io sto con le sardine" ["I'm with sardines"].

  • #NessunoTocchiLeSardine has become "nessuno tocchi le sardine"["nobody touches the sardines"].

16As a second step, I made sure to remove repeated vowels in a sentence, such as:

  • "Svegliaaaa" to get the word "Sveglia" [Wake up!].

17I also replaced the word sardines with "PartitoPoliticoS" ["PoliticalPartyS"] to prevent the entity from being mistaken for the fish that is its symbol.

18I did not remove any stop words because it is useful to create the translation system.

19At this point, I made a copy of the dataset to translate it. I used the spaCy3 language functions of POS tagging, Dependency Parse, and Entity Recognition to have all the essential components of my translator.

20The translator is a simple text representation. It is a matter of rewriting the sentence following the scheme:

  • subject adjectives

  • subjects

  • verb in the infinitive form

  • adjectives objects

  • objects

  • exclamations / other words

21At this stage, the words are not modified to make the sentence grammatically correct. Words are exchanged places, only the verb are modified to the infinitive form. The entities of type person [PER] take precedence over others.

22The translator concentrates its attention on the aspect inside the sentences to be sure to do not remove valid sentiment polarity words (Barbieri et al, 2016). And to avoid to lose them in a round-trip translation activity on translation services (Marivate & Sefara, 2020). The attempt to represent the text in a more recognizable and identifiable form for an algorithm passes from the fact that it can still recognize the entities described and the polarity expressed for each of them. For this purpose, the translator makes several attempts to fit words into their suggested position.

23Finally, I trained two models with the Ktrain framework. The model 1, which use the translated tweets, was submitted as ghostwriter19_Task_A_1_c. The model 2, trained with the only cleaned tweets, was submitted as ghostwriter19_Task_A_2_c.

2.1 First results

24The model will be evaluated with the F1-score. The main score is the average of the F1-score of the Favor tweets and the F1-score of the Against tweets.

25When comparing the two models, the first result is that the translated tweets performed worse, albeit by a few percentage points (table 1).

Table 1: First results

Model

F1-Score

ghostwriter19_Task_A_1_c

0.5613

ghostwriter19_Task_A_2_c

0.6004

Estimated Baseline

0.5386

26Analyzing the results of both the models in detail (table 2 and 3), we have that:

Table 2: F1-score details of model 1

ghostwriter19_Task_A_1_c

F1-Score

Against

0.69

Favor

0.43

Neutral

0.42

Table 3: F1-score details of model 2

ghostwriter19_Task_A_2_c

F1-Score

Against

0.70

Favor

0.50

Neutral

0.32

27The problem is evident. Model 1 has a more challenging time distinguishing the favor tweets from neutral ones. The good news is that both the models overcame the estimated baseline.

2.2 Hashtags and Mentions

28Thinking that on Twitter the hashtags are also used for classification purposes, the operation that replaces them was modified. Now the hashtags are added at the end of the new tweets. Also, the mentions are considered and processed as hashtags (table 4).

Table 4: Model 1 with hashtags and mentions in the translated tweets

Model

F1-Score

ghostwriter19_Task_A_1_c

0.5822

ghostwriter19_Task_A_2_c

0.6004

Estimated Baseline

0.5386

29Analyzing the results in detail (table 5), we can see that:

Table 5: F1-score details of model 1 with hashtags and mentions in the translated tweets

ghostwriter19_Task_A_1_c

F1-Score

Against

0.71

Favor

0.45

Neutral

0.41

30The model gained two percentage points for both Against and Favor, compared with a one-point loss in Neutral. Unfortunately, it still remains two points below the model 2, with the only cleaned tweets.

2.3 Passive verbs

31Analyzing the new texts generated, I noticed that essential information was lost by putting all the verbs in the infinitive. If the verb was in the passive form, the subject and object of the sentence were reversed. At the same time, I noticed that very long tweets contained more than one sentence.

32I modified the translator to consider passive and active verbs, swapping the sentence's subject and object if necessary. The hashtags inserted at the end of the tweet only left at the end of the new tweet generated (table 6).

Table 6: Model 1 with hashtags and mentions in the translated tweets, plus active / passive verbs

Model

F1-Score

ghostwriter19_Task_A_1_c

0.6306

ghostwriter19_Task_A_2_c

0.6004

Estimated Baseline

0.5386

33Analyzing the results in detail (table 7), we can see that:

Table 7: F1-score details of model 1 with hashtags and mentions in the translated tweets, plus active / passive verbs

ghostwriter19_Task_A_1_c

F1-Score

Against

0.76

Favor

0.50

Neutral

0.40

34The model gained five percentage points for Against and Favor tweets, compared with a one-point more loss for Neutral ones. Now the translation model is the best model.

3. Results

35Model 1 was ultimately 3 percentage points better than Model 2 with the Training dataset. The best performance of the model was also confirmed with Test datasets, with 2.5 percentage points of advantage.

3.1 Results for Task A

36The final results with the Test dataset are:

Table 8: Test dataset results for Task A

Model

F1-score

ghostwriter19_Task_A_1_c

0.6257

ghostwriter19_Task_A_2_c

0.6004

Baseline

0.5784

37The model 1 is about 7.5% better than the baseline (table 8).

38I remember that both models were trained with the autofit option, so without any particular study, to validate whether a "translation" of the original text could bring apparent advantages.

3.2 Results for Task B

39Although no context information was used, I still proposed the predictions for Task A to Task B.

40The final results with the Test dataset are:

Table 9: Test dataset results for Task B

Model

F1-score

ghostwriter19_Task_A_1_c

0.6257

ghostwriter19_Task_A_2_c

0.6004

Baseline

0.6284

41Even if model 1 was not able to reach the proposed baseline, the difference between the two systems is 0.4% (table 9). The detailed results of the models are showed in the tables 10 and 11.

3.3 Detailed results for Task A

Table 10: TASK A detailed results of the proposed models compared to the baseline model.

model

f-avg

prec_a

prec_f

prec_n

recall_a

recall_f

recall_n

f_a

f_f

f_n

1_c

0.6257

0.8106

0.4709

0.3226

0.6981

0.5357

0.4651

0.7502

0.5012

0.3810

2_c

0.6004

0.8094

0.4772

0.2921

0.6523

0.4796

0.5349

0.7224

0.4784

0.3778

baseline

0.5784

0.7549

0.3975

0.2589

0.6806

0.4949

0.2965

0.7158

0.4409

0.2764

3.4 Detailed results for Task B

Table 11: TASK B detailed results of the proposed models compared to the baseline model.

model

f-avg

prec_a

prec_f

prec_n

recall_a

recall_f

recall_n

f_a

f_f

f_n

1_c

0.6257

0.8106

0.4709

0.3226

0.6981

0.5357

0.4651

0.7502

0.5012

0.3810

2_c

0.6004

0.8094

0.4772

0.2921

0.6523

0.4796

0.5349

0.7224

0.4784

0.3778

baseline

0.6284

0.7845

0.4506

0.3054

0.7507

0.5357

0.2965

0.7672

0.4895

0.3009

4. Conclusion

42In a preliminary way, the final results demonstrate that it is possible to obtain an improvement of the predictions by reducing the differences of expression to a predetermined structure.

43The system is, however, right now, more efficient in terms of training times and final scores than ensemble systems of Bi-LSTM, which were used successfully up to 2 years ago (Bennici & Portocarrero, 2018).

44The next step is also to optimize the model's training to ascertain that the performance gain is maintained and in what percentage. At the same time, the translator can be improved by switching to a sequence-to-sequence system for a meaningful and efficient text representation that will include, among other things, the change of every words forms accordingly with the grammar and the original intention of the writers (Lewis et al., 2019).

Bibliographie

Barbieri, F., Basile, V., Croce, D., Nissim, M., Novielli, N., & Patti, V. (2016). Overview of the Evalita 2016 SENTIment POLarity Classification Task. In Proceedings of Third Italian Conference on Computational Linguistics (CLiC-it 2016) & Fifth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2016), Naples, Italy. CEUR-WS.org.

Basile, V., Croce, D., Di Maro, M., & Passaro, L. (2020). EVALITA 2020: Overview of the 7th Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. In Proceedings of Seventh Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2020), CEUR-WS.org.

Bennici, M., & Portocarrero, X. S. (2018). Ensemble for aspect-based sentiment analysis. In Tommaso Caselli, Nicole Novielli, Viviana Patti, and Paolo Rosso, editors, Proceedings of the 6th evaluation campaign of Natural Language Processing and Speech tools for Italian (EVALITA’18), Turin, Italy. CEUR-WS.org.

Cignarella, A. T., Lai, M., Bosco, C., Patti, V., & Rosso, P. (2020). SardiStance@EVALITA2020: Overview of the Task on Stance Detection in Italian Tweets. In Proceedings of the 7th Evaluation Campaign of Natural Language Processing and Speech Tools for Italian (EVALITA 2020). CEUR-WS.org.

Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., . . . Zettlemoyer, L. (2019, October 29). BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. https://arxiv.org/abs/1910.13461

Marivate, V., & Sefara, T. (2020). Improving Short Text Classification Through Global Augmentation Methods. Lecture Notes in Computer Science Machine Learning and Knowledge Extraction, 385-399. doi:10.1007/978-3-030-57321-8_21

Polignano, M., Basile, P., de Gemmis, M., Semeraro, G., & Basile, V. (2019). Alberto: Italian bert language understanding model for nlp challenging tasks based on tweets. In Proceedings of the Sixth Italian Conference on Computational Linguistics (CLiC-it 2019). CEUR-WS.org.

CC-BY-NC-ND-4.0

Le texte seul est utilisable sous licence CC BY-NC-ND 4.0. Les autres éléments (illustrations, fichiers annexes importés) sont « Tous droits réservés », sauf mention contraire.

Lire

Open access

Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search