This paper details our submission to the HAHA 2019 (5) group task on humor recognition. We propose a novel humor recognition system based on tensor embeddings, capable of being trained without the need for any external training corpora. Our model achieves an F1-score of 0.736 on a binary humor classi cation task and a root-mean-squarederror of 0.963 on a humor scoring task, both using a Spanish-language Twitter corpus (2). While our experiments are performed on Spanish documents, our approach is truly language agnostic and can be applied to any language with minimal adaptation.
Humor is an integral part of human interaction. It can be used to defuse tense
situations, increase likeability, or even for pure entertainment. As such, the
automatic recognition of humor represents an important step for natural
humancomputer interaction (
This paper details our submission to the HAHA 2019 (
Although humor recognition has typically been framed as a binary classi cation
task (1; 11; 21), recent works have moved beyond mere humor detection and
toward humor evaluation by framing humor recognition as a relative ranking
task (3; 14; 18). Generally, these works have focused on documents generated
around a common prompt such as cartoon captions (
Humor is a complex phenomenon, incorporating aspects of phonology, style,
semantics, and word-choice. As such, existing humor recognition works have
tended to use a variety of features designed to capture this complexity. (
The majority of humor recognition works focus on English language
humor. Given the dependence of many models on language dependant resources,
adapting these models for non-English contexts can be challenging. (
Tensor Embeddings
State-of-the-art document embedding approaches like doc2vec (
amounts of training data. By comparison, tensor decomposition is capable of
generating low-rank embeddings of sentences that capture the similarity of
contextual patterns without the need for large training corpora (
Such \tensor embeddings", combined with a label propagation technique,
have been shown e ective in semi-supervised fake news detection (
For each document in the corpus we compute a tensor embedding based on word co-occurrence. That is, for a corpus D = {s1; s2; : : : ; sD} with D sentences, we rst extract a vocabulary w1; w2; : : : ; wV , where V is the number of words. For each sentence s in D, we count the word-word co-occurrences within a small window H. This results in a frequency matrix Ws ∈ ZV ×V , where Z denotes the set of integers. In particular, Ws(i; j) indicates the frequency that word wi and wj co-occur in s within the window H. This allows us to encode the lexical patterns of s in Ws. All Ws are then stacked, creating a three-dimensional tensor W ∈ ZV ×V ×D. The objective of tensor decomposition is to nd an approximation W^ of W such that:
R
^
W = Q vr ⊗ vr ⊗ dr;
r=1
(
In particular, C = [d1; d2; : : : ; dR] ∈ RD×R, where the s-row of C is the embedding vector of sentence s.
The low-rank sentence embeddings are calculated using the alternating least
squares method of the CANDECOMP/PARAFAC tensor decomposition
implementation in Matlab tensor toolbox 4 (
Humor Recognition
After extracting tensor embeddings for each document, we performed a simple
baseline experiment using the (
then used to predict the labels and scores of documents in the test set as part
of HAHA 2019's (
Second, inspired by (
Third, inspired by (
Except where otherwise noted, all results reported in this paper were obtained using the Random Forest classi er and regressor implementations in scikit-learn5 with default parameters. These results are comparable to results we obtained using Support Vector Machine and Extra Trees models. Label propagation was also performed using the implementation in scikit-learn with default parameters. 4
The performance of our humor recognition models is shown in Table 1. Despite the simplicity of our models, we are able to achieve reasonable performance. Our tensor embedding only models outperforms both our label propagation and enhanced models across the board. Despite the performance of a similar model
on fake news detection (
It is important to note at while the (
Another advantage of our approach is its relative simplicity. The most
computationally expensive aspect of our model is computing the tensor decomposition.
This makes approach quite scalable when compared with complex
sequencedbased neural models such as those used by (
Part of the reason for the poor label propagation performance may be the
(
The semi-supervised nature of label propagation may further explain this
lack of performance. Given the di culty in obtaining reliable humor judgments,
it should come as no surprise that most humor datasets tend to be relatively
small. Some datasets, like Pun of the Day (
The performance of label propagation may also be a ecting the performance
of the enhanced model. Another potential limitation the enhanced model is our
lexical centrality feature. One major di erence between the (
Another potential area for improvement is related to the hyper parameters
used to generate our tensor embeddings. The window size of 5 and tensor rank
of 50 was chosen empirically due to their high performance on smaller,
Englishlanguage datasets. The larger size of the (
In this paper we have shown that tensor embeddings are capable of producing reasonable performance for both binary humor classi cation and humor scoring. Furthermore, we identify some key advantages of this tensor embedding approach including its simplicity and language agnosticism. Finally, we o er several potential avenues for improving the performance of our models including better hyperparamter tuning.