Modeling human affect is non-trivial. To undertake this challenge, a novel shared task focusing on happiness is organized at an AAAI workshop. The CL-AFF Shared Task, titled \In Pursuit of Happiness", consists of two sub-tasks on a dataset of descriptions of happy moments (taken from the HappyDB dataset), each annotated with individuals' demographics, recollection time and relevant labels. We focus on the rst sub-task, which is a semi-supervised task to determine a happy moment's agency and social label. We present a deep learning system for this task based on Recurrent Convolutional Neural Networks (RCNN). The presented system (which we call IoH-RCNN) is trained and tested on the available dataset using 10-fold cross-validation. For predicting the agency label, the average accuracy, f1 and AUC are 85.5, 90.3 and 80.0, respectively. As for predicting the social label, the average accuracy, f1 and AUC are 91.8, 92.2 and 91.2, respectively.
Understanding user expression is very challenging. To tackle this issue, the
organizers of the Affective Content Analysis (AffCon) workshop at AAAI 2019
propose a shared task with the goal of providing better modeling for human
affect. The task, which is the rst of its kind, focuses on happiness and is titled
\In Pursuit of Happiness" [
The task offers two datasets (a small labeled one and a large unlabeled one)
taken from the HappyDB dataset [
We focus on the rst sub-task and address it using Neural Networks (NN).
Speci cally, we propose a system based on the Recurrent Convolutional NN
(RCNN) architecture of Lai et al. [
It is worth mentioning that the \In Pursuit of Happiness" tasks are quite
unique. In fact, we could not nd prior work addressing the semi-supervised
task, which is the task under consideration in this work. Kindly, remember that
the focus here is on predicting the agency and social labels of happy moments
each described in a single sentence written in English and is accompanied with
information related to the author (age, gender, etc.) and the happy moment
(e.g., duration, concept, etc.). Thus, works focusing on modalities other than
text or addressing affective analysis from a different perspective or with different
objectives are outside the scope of this paper. Examples include [
The rest of this paper is organized as follows. In the following section, the \In Pursuit of Happiness" task is described. The proposed IoH-RCNN approach is presented in Section 3 along with its experimental evaluation. Finally, the paper is concluded in Section 4 with main ndings and thoughts on future directions of this work. 2
The task at hand, which is titled \In Pursuit of Happiness" is presented in this
section in a brief way. Interested readers are referred to the task description
paper [
The common practices when it comes to emotion analysis do not necessarily
capture the experiential, contextual and agentic attributes of happy moments.
Since human affect, in general, is context-driven, labeled datasets must account
for these factors in generating predictive models of affect. This was the main
motivation behind offering the CL-Aff shared task at the AffCon workshop of
AAAI 2019. Based on the HappyDB dataset [
According to the organizers, the labeling is done by annotating each record with labels that identify the `agency' of the author and the `social' characteristic of the moment, as well as concept labels describing its theme. The two labels that will be used for classi cation are both binary ones. They are agency and social. The agency label is whether the author is in control or not, whereas, the social label is whether the described moment involves other people other than the author or not. 1 https://sites.google.com/view/affcon2019/cl-aff-shared-task
The information made available with each happy moment include information about the author as well as information about the happy moment itself. The former include the age, gender, country, marital status, parenthood status and demographics of the author, while the latter include the re ection period information and the duration of the happy moment. Finally, each moment is annotated with 1-4 concepts, where the available concepts are 15. One example of concepts assigned to a single happy moment is: \familyjeducationjparty".
The shared task has two components: a semi-supervised one and an unsupervised one. For the former component, the participating teams can use both the labeled and unlabeled sets to predict the agency and social labels of each happy moment in the test set. As for the unsupervised task, it is concerned with modeling happiness. I.e., it asks for new characterizations and insights for the testing happy moments (e.g., in terms of affect, emotion, participants and content.)
The evaluation is done in a simple way. For the semi-supervised task, the evaluation metrics are simple accuracy, f1 measure and AUC (Area Under the ROC Curve). Each metric is applied to each of the two binary labels under consideration (agency and social). On the other hand, due to the open-ended nature of the unsupervised task, the evaluation metric is not de ned. We note that we are only interested in addressing the semi-supervised task in this paper.
A wide variety of text classi cation tasks in the eld of natural language
processing (NLP), such as topic identi cation, spam ltering and sentiment analysis can
be considered supervised learning problems and, thus, can be addressed with a
text classi er. Many semi-supervised algorithms that use labeled and unlabeled
data have been proposed to enhance supervised classi cation and improve the
model robustness by more precise decision boundary. One of the most
successful ones is Lai et al. [
The structure of the IoH-RCNN model is presented in Figure 1. The input for the model is a sentence consisting of a sequence of words S = w1; : : : ; wn describing the happy moment along with the set of features provided by the task organizers (author gender, moment duration, context, etc.) and the output is the label value f0; 1g. Note that we built different RCNN for each problem under consideration (i.e., one of predicting agency and another one to predict social). The textual part of the input is handled through RCNN, which produces an output vector called y(3). As for the other parts of the input, which are the supplied (external) features, they are fed into a Feed Forward NN (FFNN), that produces an output vector called f f . The two output vectors are then concatenated and fed into the output layer. Finally, the loss function we use is softmax cross entropy.
In order to get a more precise word embedding, words are combined with their contexts. The contexts are captured by using a Bi-RNN, which allows for better disambiguation of each word's meaning. This gives RCNN an advantage over traditional NN which captures context information only within a xed window. The recurrent structure captures the left and right context information by performing a forward scan/pass followed by a backward scna/pass over the input.
To be more speci c, we explain how RCNN work using the same notation
used in the original paper [
xi = [cl(wi); e(wi); cr(wi)] The word representations go through a linear transformation before being moved to the next layer.
y(2) = tanh(W (2)xi + b(2)) i The latent semantic vector y(2) is where each semantic factor will be analyzed i to determine the most useful factor for representing the text.
The previous paragraph explained the recurrent structure of RCNN. Now, we explain the CNN part, which views recurrent structure as the convolutional layer. After calculating word representations and latent semantic vector, a maxpooling operation is applied to capture the information throughout the entire text. Here, the max operation is performed in an element-wise fashion. 3.1
Empirical Evaluation The proposed system (IoH-RCNN) is implemented on Google's TensorFlow. It is evaluated and tested on the provided data using 10-fold cross validation. The results for the agency label are as follows.
{ Average accuracy= 85.5 { Average f1= 90.3 { Average AUC= 80.0 { Average accuracy= 91.8 { Average f1= 92.2 { Average AUC= 91.2 As for the Social label, the results are as follows.
As mentioned earlier, we compare IoH-RCNN with three baseline systems: BiLSTM, Att-BLSTM and CNN. We experiment with different con gurations for the parameters of these systems and report the best results. The con gurations that produce the best results are reported in Table 1.
A comparison of the four models' accuracies is shown in Table 2. The table shows the superiority of IoH-RCNN over baseline systems in predicting both the agency and social labels.
We note that we also experiment with other innovative approaches such as the Transformer model. However, the results we obtain are not high enough compared with those of the IoH-RCNN system. 4
In this paper, we addressed the semi-supervised sub-task of the CL-AFF Shared
Task, titled \In Pursuit of Happiness". Speci cally, the problem is: given a small
set of happy moments descriptions labeled with agency and social labels and a
large unlabeled set of happy moments descriptions, the goal is to build a model
to predict the agency and social labels of the test set. The happy moments are
annotated with information related to the author (age, gender, etc.) and the
happy moment (e.g., duration, concept, etc.). The model we presented to solve
this problem is based on the exciting RCNN. The obtained results were high
(with accuracies of 85.5 and 91.8 for the agency the social labels, respectively)
compared with other methods such as Bi-LSTM, Attn-BLSTM, CNN and
Transformer (whose best accuracies for the agency the social labels were 85.2 and 89.9,
respectively). In the future, we plan on exploring more cutting edge techniques
such as Ensemble Classi cation, Transfer Learning and Bidirectional Encoder
Representations from Transformers (BERT) [
We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research.