Sentim at SemEval-2019 Task 3: Convolutional Neural Networks For Sentiment in Conversations

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 302–306 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 302 Sentim at SemEval-2019 Task 3: Convolutional Neural Networks For Sentiment in Conversations Jacob Anderson Sentim LLC, USA papers@sentimllc.com Abstract In this work convolutional neural networks were used in order to determine the sentiment in a conversational setting. This paper’s con- tributions include a method for handling any sized input and a method for breaking down the conversation into separate parts for easier processing. Finally, clustering was shown to improve results and that such a model for han- dling sentiment in conversations is both fast and accurate. 1 Introduction The model for this paper was created for Task 3 of SemEval 2019, EmoContext (Chatterjee et al., 2019). The basic idea of the task is to classify the emotion (as “angry”, “sad”, “happy”, or “oth- ers”) that someone is expressing in the third turn of a three part conversation, given the previous two turns as context. The dominant approach in many natural lan- guage tasks is to use recurrent neural networks or convolutional neural networks (CNN) (Conneau et al., 2016). For classification tasks, recurrent neural networks have a natural advantage because of their ability to take in any size input and out- put a fixed size output. This ability allows for greater generalization as no data is removed nor added in order for the inputs to match in length. While CNN’s can also support input of any size, they lack the ability to generate a fixed size output from any sized input. In text classification tasks, this often means that the input is fixed in size in order for the output to also have a fixed size. This work expands upon a previous work (An- derson, 2018), a way of using CNN’s for classifi- cation to allow for any sized input length without adding or removing data. That work was expanded upon in this paper by making it simpler, using it in a different setting, and applying a new method to compensate for the previous model’s deficiencies. 2 Model Description The overall architecture of this model can be bro- ken into two parts: the language understanding model and the emotion prediction model. The lan- guage understanding model takes in each part of the conversation and processes it separately in or- der to produce a latent vector representing the net- work’s understanding of that part of the conversa- tion. Then, the emotion prediction model takes all of the latent vectors from the language understand- ing model and combines them in order to make a prediction for what the emotion is. More specifically, the language understanding model takes in the first and third turns of each con- versation1, processes them separately, and returns a respective understanding vector of size 128 for each input. The emotion model then concatenates those vectors to make one vector of size 256 and then processes that output through a small fully connected network in order …

Content truncated to 500 words