Exploring Semantic Differences Between Japanese and English Emotional Terms Using Word Embeddings

What this article does

I assume that basic emotions are Gaussian-distributed in an arbitrary semantic space, regardless of the language. By considering the vector representations of basic emotions within fastText models trained separately in English and Japanese, and comparing the structures created by the inter-distribution distances (Bhattacharyya distance) of the six basic emotions, I will reflect on the differences between English and Japanese emotional words.

Motivation

I believe your "ureshii" (happy/glad) and my "ureshii" are likely quite different in meaning. Well, of course, they are different because the things we feel "ureshii" about are different, but I suspect that even in situations described by the same word "ureshii," we often experience qualitatively different emotions at the level of the qualia we hold when using that word.

Whether the emotions we feel are universal to humans is a grand question. On one hand, some studies propose several types of "basic emotions," claiming that there are universal emotions regardless of culture or language. On the other hand, there is research from a social-constructionist perspective that focuses on individual differences and socio-cultural variations in emotion, arguing that different parts are indeed different. It is an area of endless debate, but if the meanings of emotions differ between individuals or languages, the question arises as to how to evaluate those differences.

In fields like cultural anthropology, researchers might use photos or audio to investigate whether there are differences in how groups perceive emotions. Instead of that, I wanted to see if I could more easily compare something like the "difference in the nuance of words expressing emotion."

Disclaimer

The author of this article is not an expert in anything related to this topic, so all the following content is based on my own vague understanding. Please do not take it too seriously.

Also, this seems to be an article written around March 2019, so various parts are likely outdated. Furthermore, it is now known that the Japanese fastText model from that time was a "broken" model with some corrupted vectors (I learned this after writing the article), so I am not sure if this is a meaningful experiment.

Vague Discussion

Here, the goal is to compare the "differences in the nuances of words expressing emotions" between languages.

To achieve this, I first consider capturing the "nuance of words expressing emotions" within an appropriate word embedding vector space. The nuance of a word is, after all, its relationship with other words. Therefore, I quantify the "nuance of words expressing emotions" within the vector space by considering the similarity (distance) between emotional words. In short, I create a distance matrix.

While a distance matrix can quantify the "nuance of words expressing emotions" within a single language, a distance matrix obtained using common vector distances cannot be used for cross-linguistic comparison because the underlying basis vectors differ. Therefore, I assume that several basic emotions are Gaussian-distributed in an arbitrary semantic space regardless of the language, and consider a distance matrix created by the inter-distribution distances of these basic emotions.

According to my vague understanding, translation of words between arbitrary word embeddings (assuming models created with the same parameters) is equivalent to a linear transformation of vectors by considering the mapping from one word embedding to another. Since inter-distribution distances such as the Hellinger distance or Bhattacharyya distance have the property that the distance structure does not change even when an arbitrary linear transformation is applied, it should be acceptable to perform cross-linguistic comparisons if we consider the distance structure created by the inter-distribution distances of basic emotions.

Concrete Discussion

Models Used

I use the pre-trained fastText models released by facebookresearch. This time, I chose the English and Japanese pre-trained models. Both are 300-dimensional models trained with the same parameters, and the vocabulary size is 2,000,000.

Since the size is too large for my weak machine to handle, I will only use the first 150,000 words (the vocabulary is sorted by frequency of occurrence) from each model.

Basic Emotions

I adopted the six basic emotions proposed by Ekman ("anger", "fear", "disgust", "joy", "sadness", "surprise") and calculated the inter-distribution distances of basic emotions within each of the English and Japanese languages.

That said, since it is naturally unknown which vocabulary originates from which basic emotion, I decided to roughly select words that seem to represent basic emotions and use 50 nearest neighbor vectors for each word as samples for each basic emotion.

The following are the words I chose:

	anger	fear	disgust	joy	sadness	surprise
English	^angry$	^fearing$	^disgusted$	^joyful$	^sad$	^surprised$
Japanese	^怒っ$	^怖い$	^嫌$	^嬉しい$	^悲しい$	^驚い$

Below are plots of the neighborhood vectors reduced to two dimensions using UMAP. Although I took 50 neighborhood vectors for each of the six basic emotions, only around 260 points are plotted because there are duplicate words in both English and Japanese. The colors were assigned by re-clustering with k-means after removing duplicates.

Nearest Neighbor Search

The author of this article is an R user and used {rflann} (note that it has already been removed from CRAN as of March 2019). For nearest neighbor search in R, there is also {RcppAnnoy}.

Bhattacharyya Distance

I calculated the Bhattacharyya coefficient using bio3d::bhattacharyya() and then applied -log(). I used it based on the general vibe, so the usage might be wrong, but I won't think about it too deeply.

Visualization

Since it might be hard to understand if I just post the distance matrix directly, I decided to draw a graph.

It looks something like this. I'm not sure if it's appropriate to use MDS when using Bhattacharyya distance for the distance matrix (whether it's meaningful to approximate a space spanned by Bhattacharyya distance with a 2D Euclidean space), but I won't think about it too deeply.

g <- make_full_graph(n = 6)
V(g)$name <- emotions[[i]]$rownames
ggraph(g, layout = "igraph", algorithm = "mds", dist = distances[[1]][[2]], dim = 2) +
    geom_edge_link() +
    geom_node_point() +
    geom_node_label(aes(label = name), repel = TRUE) +
    theme_light()

Results

It looks like this.

Distance Structure with Bhattacharyya Distance

English

Japanese

Distance Structure with Euclidean Distance

For reference, I am also posting graphs of the distance structure created by calculating the Euclidean distance between "basic emotion vectors," where each basic emotion sample was fed into kmeans(centers = 1) to obtain the center.

English

Japanese

Reflecting on the Differences

There was less correspondence than I expected, so I'm not quite sure what to write.

I also drew dendrograms, but my only real impression is that the nuance of "fear" seems quite different. However, based on the logic of fastText, this isn't necessarily about emotions being qualitatively similar because they are close in inter-distribution distance; rather, it means the contexts in which they appear tend to be similar. So, I think it suggests that the contexts in which Japanese people feel "fear" are quite different from the contexts in which they feel the other five basic emotions.

Dendrograms

English

Japanese