Identifying Intensity of the Structure and Content in Tweets and the Discriminative Power of Attributes in Context with Referential Translation Machines

TL;DR

The paper investigates using referential translation machines (RTMs) to assess the discriminative power of lexical attributes and to predict emotion intensity in tweets across English, Arabic, and Spanish. By framing both tasks as machine translation performance prediction (MTPP), the authors develop stacked RTM architectures (combined and separate predictors) and leverage translation-similarity features derived from WordNet Affect emotion lists. They evaluate on SemEval-2018 Task 10 (discriminative attributes) and Task 1 (affect in tweets), using cross-language lexicons and best-worst scaling to derive intensity scores, and report metrics such as Pearson's correlation $r$ and $F_1$. Post-challenge experiments introduce symbolic grounding of thresholds and statistics to calibrate predictions, yielding modest but consistent improvements. Overall, the approach provides a novel, multilingual framework for measuring semantic similarity and emotion intensity, with potential applicability to broader NLP tasks and cross-language comparisons.

Abstract

We use referential translation machines (RTMs) to identify the similarity between an attribute and two words in English by casting the task as machine translation performance prediction (MTPP) between the words and the attribute word and the distance between their similarities for Task 10 with stacked RTM models. RTMs are also used to predict the intensity of the structure and content in tweets in English, Arabic, and Spanish in Task 1 where MTPP is between the tweets and the set of words for the emotion selected from WordNet affect emotion lists. Stacked RTM models obtain encouraging results in both.

Figures (3)

• Figure 1: RTM depiction: parfda selects interpretants close to the data using corpora; two MTPPS use interpretants, training data, and test data to generate features in the same space; learning and prediction use these features as input. Spheres are for feature spaces.
• Figure 2: RTM with stacked combined prediction use a combined model to obtain feature representations and predictions for both $w_1 \rightarrow a$ and $w_2 \rightarrow a$, which are processed before additional learning and prediction.
• Figure 3: RTM with stacked separate predictions use two different learning steps to obtain feature representations and predictions for either $w_1 \rightarrow a$ or $w_2 \rightarrow a$, which are processed before additional learning and prediction.