Simple dynamic word embeddings for mapping perceptions in the public sphere

TL;DR

This paper introduces a unified dynamic embedding model to analyze social biases in language over time, applied to historical and radio transcript corpora, revealing when dynamic models are more effective than static ones.

Contribution

The paper presents a novel dynamic embedding approach for attribute-specific word representations, enabling better analysis of evolving social biases in large-scale corpora.

Findings

Dynamic embeddings reveal temporal bias patterns.

Static models may be less effective for certain bias analyses.

Application to diverse corpora demonstrates model versatility.

Abstract

Word embeddings trained on large-scale historical corpora can illuminate human biases and stereotypes that perpetuate social inequalities. These embeddings are often trained in separate vector space models defined according to different attributes of interest. In this paper, we develop a unified dynamic embedding model that learns attribute-specific word embeddings. We apply our model to investigate i) 20th century gender and ethnic occupation biases embedded in the Corpus of Historical American English (COHA), and ii) biases against refugees embedded in a novel corpus of talk radio transcripts containing 119 million words produced over one month across 83 stations and 64 cities. Our results shed preliminary light on scenarios when dynamic embedding models may be more suitable for representing linguistic biases than individual vector space models, and vice-versa.

Equations

$$E(w,A)={\gamma }_w+{\Sigma }_{a\in A}{\beta }_w^a$$

$$bia{s}_A=\frac{{\Sigma }_{r\in R}cos(E(r,A),\overline{a})-cos(E(r,A),\overline{o})}{|R|}$$

Full text

This is a modified version of a paper that originally appeared in the 2019 NAACL workshop on NLP+CSS. Since its initial publication, we discovered implementation errors that invalidate the results presented in the original version of the paper. These errors are corrected in this version. A corrigendum at the end of this paper summarizes the original errors and how we corrected them.

1 Introduction

Language has long been described as both a cause and reflection of our psycho-social contexts Lewis and Lupyan (2018). Recent work using word embeddings—low-dimensional vector representations of words trained on large datasets to capture key semantic information—has demonstrated that language encodes several gender, racial, and other biases that correlate with both implicit biases Caliskan et al. (2017) and historical trends Garg et al. (2018).

These studies have validated the use of word embeddings to measure a range of psychological and social contexts, yet in most cases, they do not leverage the full power of available datasets. For example, the historical biases presented in Garg et al. (2018) are computed using decade-specific word embeddings produced by training different Word2Vec Mikolov et al. (2013) models on a large corpus of historical text from that decade. The authors then use a Procrustes alignment to project embeddings from different models into the same vector space so they can be compared across decades Hamilton et al. (2016). While this approach is reasonable when there are large-scale datasets available for a given attribute of interest (e.g. decade), it requires an additional optimization step and disregards valuable training data that could be pooled and leveraged across attribute values to help with both training and regularization—especially when data is sparse.

In this paper, we present a unified dynamic word embedding model that jointly trains linguistic information alongside any categorical variable of interest describing its context (-e.g. year, geography, income bracket, etc.). We apply this model to two datasets: i) the Corpus of Historical American English (COHA Davies (2010)) to analyze gender and ethnic occupation biases, and ii) a novel data corpus of 119 million words spoken on talk radio Beeferman and Roy (2018) during a one-month period in late 2018 across 64 US cities to explore perceptions about refugees. Our results shed preliminary light on scenarios when dynamic embedding models may be more suitable for representing linguistic biases than individual vector space models, and vice-versa.

2 Model

We describe our model and implementation below.

2.1 Overview

Our dynamic embedding for word $w$ is defined as

[TABLE]

where $\gamma_{w}$ is an attribute-invariant embedding of $w$ computed across the entire corpus, $\beta_{w}^{a}$ is the offset for $w$ with respect to attribute $a$ across the set of attributes $A$ we are interested in computing the word embedding with respect to. For example, if we wish to compute the embedding for the word “refugee” as it was used on the 25th day of a particular 30-day corpus of talk radio transcripts, we would set $w=refugee$ and $A=\{25\}$. This approach, as formalized in Equation 1 above, is identical to one introduced by Bamman et al. (2014), though finer details of our model and training differ slightly, as described below.

To learn $\gamma_{w}$ and $\beta_{w}^{a}$, we train a neural network. Our model is a simple extension to the distributed memory (DM) model for learning paragraph vectors originally introduced in Le and Mikolov (2014). The DM model uses a continuous bag-of-words architecture to jointly train a paragraph ID with a sequence of words sampled from that paragraph to predict a particular word given the words that surround it. The output of this model includes a semantic vector representation of a) each paragraph, and b) each word in the vocabulary.

Our model extends the DM model by adding an additional dimension to the paragraph vector to learn specific paragraph-by-word—or, in our context, attribute-by-word—embeddings (i.e., $\beta_{w}^{a}$). The penultimate layer (before word prediction) is computed as an average of the dynamic embeddings for each context word, i.e., $X=\frac{1}{N}\Sigma_{i=1}^{N}E(w_{i},S,A)$, where $N$ is the size of our context window. This average embedding is then multiplied by the output layer parameters and fed through the final layer for word prediction. Figure 1 depicts our model architecture.

2.2 Implementation

We build on an existing PyTorch implementation of paragraph vectors111Available at: https://github.com/inejc/paragraph-vectors. to implement our model, setting the dimensionality of $\gamma_{w}$ and $\beta_{w}^{a}$ to be 100. We use the Adam optimization algorithm with a batch size of 128, word context window size of 8 (sampling four words to the left and right of a target prediction word), learning rate of 0.001, and L2 penalty to regularize all model parameters (where $\lambda$=1e-5). We only train embeddings for words that occur at least 10 times in the corpus. For training, we use the negative sampling loss function, which Mikolov et al. (2013) show is more efficient than the hierarchical softmax and yields competitive results222We include three noise words when computing the loss..

3 Case study 1: gender and ethnic occupation biases in COHA

We first train our model on the Corpus of Historical American English Davies (2010) for 2 epochs (each epoch takes approximately 40 hours to train) in order to compare its outputs to those produced via the individual decade-by-decade word embedding models used in Garg et al. (2018). We use the same metric and word lists as the authors to compute bias scores, substituting in the cosine distances between vectors for the norm of their difference (both approaches yield nearly identical results). In particular, we compute linguistic bias scores for two of their analyses: i) the extent to which female versus male words are semantically similar to occupation-related words, and ii) the extent to which Asian vs. White last names are semantically similar to the same, from 1910 through 1990. We then qualitatively analyze relationships between changes in these scores and the actual changes in female and Asian workforce participation rates (relative to men and Whites, respectively) over the same time period.

Figure 2 depicts these results. The results from the dynamic model depicted in (a) qualitatively appear to match the ground truth workforce participation rates, except in the earlier decades of the 20th century. The results depicted in (b), however, do not appear to resemble those in Garg et al. (2018) or ground truth. One hypothesis for these trends, especially those depicted in (b), is that the attribute value-specific offset word vectors might not be encoding the right, or enough, information about those words when compared to the word’s attribute-invariant embedding. This seems plausible especially since the norms of the attribute-invariant word vectors are often one to two orders of magnitude larger than the norms of their corresponding attribute value-specific offsets. Given research that has identified that words occurring more frequently in consistent semantic contexts tend to have larger norms Schakel and Wilson (2015), it is possible that some of the Asian last names used as input words into the computations for (b) simply do not occur frequently enough within the individual decades for their decade-specific representations to retain explanatory power when compared to their corpus-wide, attribute-invariant representations. An analogous reason might explain why the bias scores for the earlier decades in the 20th century are so high: it is possible that the corpus-wide embeddings for certain words used in those calculations are simply “overpowering” their decade-specific offsets.

Running some additional tests starts to confirm some of these hypotheses. For example, “war” was a prominent feature of the 20th century, and so, we might expect the most semantically-similar words per decade to the word “war” to relate to the wars that were particularly salient during those (or adjacent) decades. Figure 3(a) shows (in order of cosine similarity) the words per decade that are most similar to that decade-specific embedding for “war”. We can see that in some cases, our intuition holds: e.g., the 1950s contain words like “holocaust” and “world” (presumably for“world war”), both of which seem to be reasonable for that time period. However, “holocaust” also appears in the list for 1930—even though the term was first mentioned in the New York Times in the context of the second World War in the early 1940s333https://en.wikipedia.org/wiki/Names_of_the_Holocaust.. Possible explanations for these observations include: i) there is a strong corpus-wide association between “war” and “holocaust” that makes them similar across decades (even those preceding the Holocaust), and ii) the 1930s-specific offset for “holocaust” was perhaps simply not updated enough during training (due to a low number of occurrences) to move away from its initialized (random values). If those initialized values were small enough (e.g. had a small norm), it’s quite possible that “holocaust” could still show up as a semantically similar term to “war” even before it became a more commonly-used term. These appear to be plausible explanations when we look at the decade-specific associations with “war” computed using the non-dynamic model of Garg et al. (2018) and shown in Figure 3(b): the associations appear to depict historical realities much more accurately than those inferred by our dynamic model.

We see a similar pattern when looking at words that are most similar to “computer” per decade. Figure 3(c) shows words that are most similar according to our dynamic model, and (d) shows the most similar words according to the non-dynamic model. According to our dynamic model, “internet” is a nearest neighbor as early as 1910, which is clearly historically inaccurate. On the other hand, the non-dynamic model does not learn an embedding for computer until the 1960s (due to limited use of the word in the first half of the 20th century), after which it produces similar words that appear to accurately reflect decade-specific trends.

One opportunity for future work involves trying different (smaller) values of $\lambda$ in our l2 regularization term, as doing so might encourage attribute value-specific embeddings for a given word to retain more “influence” in its overall representation. In any case, our results reveal some of the possible shortcomings of using our dynamic embedding model, particularly when there is enough data to simply learn and align individual attribute-specific word embedding models.

4 Case study 2: bias against refugees expressed on talk radio

The earlier section revealed some of the shortcomings of our model when applied to a large historical corpus. To explore its performance in a smaller-scale corpus captured over a much shorter time horizon, we apply our model to the transcriptions of talk radio shows in order to identify biases against refugees. Talk radio is a significant source of news for a large fraction of Americans: In 2017, over 90% of Americans over the age of 12 listened to some type of broadcast radio during the course of a given week, with news/talk radio serving as one of the most popular types Pew (2018). With listener call-ins and live dialog, talk radio provides an interesting source of information, commentary, and discussion that distinguishes it from discourse found in both print and social media. Given the proliferation of refugees and displaced peoples in recent years (totalling nearly 66 million individuals in 2016 UNHCR (2017))—coupled with the rise of talk radio as a particularly popular media channel for conservative political discourse Mort (2012)—analyzing bias towards refugees across talk radio stations may provide a unique window into how Americans perceive this important issue.

4.1 Dataset and analyses

Our data is sourced from talk radio audio data collected and automatically transcribed by the media analytics nonprofit Cortico444http://cortico.ai.. The data is further processed to identify different speaker turns into “snippets”; infer the gender of the speaker; and compute other useful metrics (more details on the radio data pipeline can be found in Beeferman and Roy (2018)).

We train our dynamic embedding model on a talk radio datasets sourced from 83 stations located in 64 cities across the US. The dataset includes over 4.8 million snippets comprised of 119 million total words produced by 433 shows between August 15, and September 15, 2018555As a rough proxy for removing syndicated content, we include only those snippets produced by a talk radio shows that air on one station..

Bias against refugees is defined similarly to how the authors of Garg et al. (2018) define bias against Asians during the 20th century, measuring to what extent radio shows associate “outsider” adjectives like “aggressive”, “frightening”, “illegal”, etc. with refugee and immigrant-related terms in comparison to all other adjectives. To compute refugee bias scores with respect to the attribute set $A$, we use a modified version of the relative norm distance metric from Garg et al. (2018):

[TABLE]

Where $E(r,A)$ is the dynamic embedding for a given refugee-related word $r$ (e.g. “refugee”, ”immigrant”, ”asylum”, etc); $\overline{a}$ is the average dynamic embedding computed for each $w$ in the set of all adjectives with respect to $A$; $\overline{o}$ is analogously defined for outsider adjectives; and $cos(\cdot)$ is cosine distance. A positive value for $bias_{A}$ indicates discourse about refugees that is biased against them as “outsiders”. We normalize by the total number of refugee-related words, $R$, to provide a relative indication of the amount of linguistic bias present in the data (since $bias_{A}$ is bounded by $\pm|R|$).

We use our model to analyze how bias on talk radio against refugees varies by day between August 15 and September 15, 2018. The median number of words for each day in the talk radio corpus is 4 million—over 5x fewer than a median of 22 million words per decade used to train each decade-specific model in Garg et al. (2018). As a comparison, we also compute bias scores by training one Word2Vec model per day and projecting all day-by-day models into the same vector space using orthogonal Procrustes alignment666We use the Gensim implementations of Word2Vec and orthogonal Procrustes alignment, aligning hyperparameters as closely as possible to our dynamic model. similar to Hamilton et al. (2016). The results from these analyses are depicted in Figure 4. Unlike the earlier section where we used actual workforce participation rates as a “ground truth” to compare linguistic biases against, it is unclear what ground truth is in this case in order to be able to evaluate model performance. Still, one key difference in the results is that most of the daily bias scores computed using the non-dynamic model are negative—suggesting, in general, that talk radio participants do not discuss refugees as “outsiders”—whereas all of the scores computing using the dynamic model are positive. One reason to believe the latter is a more accurate depiction of reality—that discourse about refugees on talk radio would tend to cast them as “outsiders”—is that talk radio has been identified as a popular media source for political conservatives in the US Mort (2012). In turn, political conservatives—and particularly their more extreme right-wing factions—have historically been unwelcoming towards refugees Bencek and Strasheim (2016); Klaus (2017).

A possible analogue to using longer-term historical changes in workforce participation participation rates as a proxy for “ground truth” is to use shorter-term “historical” changes in refugee-related news events. Upon qualitative inspection, it appears that the results from our dynamic embedding model might be tracking the news cycle. For example, the lowest bias scores in the chart—August 17 and 24—could perhaps correspond to stories about how “U.S. Will Not Spend $230 Million Allocated to Repair Devastated Syrian Cities” and “For Rohingya, Years of Torture at the Hands of a Neighbor”, respectively (as reported by the New York Times). Intuitively, public discourse that is relatively less-biased against (and perhaps even more empathetic towards) refugees in response to these stories appears to make sense. Conversely, the highest bias scores on August 15 and 25 could be related to other New York Times stories titled “ISIS Member Arrested in Sacramento, U.S. Says”777in which case, the arrested individual had actually applied for refugee status in the US, and “Year After Rohingya Massacres, Top Generals Unrepentant and Unpunished”, respectively. In the case of the latter story, it seems much of the radio discussion describes “angry protests” by refugees affected by the Rohingya crisis—which may be contributing to elevated negative/“outsider” bias scores. Of course, there are several seemingly prominent refugee-related events that do not correspond to extreme daily bias scores, e.g. the September 14 announcement that the “U.S. Is Ending Final Source of Aid for Palestinian Civilians”.

While the non-dynamic model’s mostly-negative absolute scores might suggest that it fails to capture generally unwelcoming discourse about refugees, its extreme peaks and valleys, too, correspond to news stories in ways we might expect. For example, the two days with the lowest bias scores—August 19 and September 3—coincide with stories titled ”‘All of Africa Is Here’: Where Europe’s Southern Border Is Just a Fence” and ”Mediterranean Death Rate Is Highest Since 2015 Migration Crisis”, respectively. Both of these articles describe the treacherous journeys many refugees embark on as they flee their countries. One of the days with the highest bias scores (September 9) coincides with a story titled ”Sweden’s Centrists Prevail Even as Far Right Has Its Best Showing Ever” (which also references the right-wing party’s anti-immigrant positions); the other (September 11) does not appear to correspond to a specific refugee-related story.

Together, these results suggest that our dynamic model may produce embeddings that more accurately capture linguistic biases towards refugees, but also that using the news cycle as “ground truth” is not sufficient for evaluation purposes. Future research should include i) more thoughtful selections of “ground truth” for model comparisons and validation, ii) comparisons to other dynamic word embedding models that treat time as a continuously-valued attribute, e.g. Bamler and Mandt (2017); Rudolph and Blei (2018); Yao et al. (2018), and iii) an exploration of if/how these findings hold for topics other than refugees, across different time slices and discourse corpora.888In the original version of this paper, we also included a case study where we analyzed how bias against refugees differs by geography. This was a largely speculative analysis that we included after validating our model using COHA and the time-series of radio data discussed in this section. Since our model’s validity is much weaker in this corrected version of the paper, we have removed this speculative geographic analysis. Additional details on the original analyses can be found in the corrigendum.

5 Conclusion

In this paper, we present a dynamic word embedding model mirroring the earlier work of Bamman et al. (2014) to learn attribute-specific embeddings. We observe several limitations of our model in comparison to the non-dynamic models used in Garg et al. (2018) to measure gender and ethnic occupation biases embedded in COHA. However, preliminary results suggest our model might more accurately capture linguistic biases expressed against refugees on talk radio, where data per attribute value is much sparser than in COHA. Our results are highly preliminary and require further investigation to determine under which conditions dynamic embedding models like ours might be more suitable to use than individual attribute-specific models. Some directions for investigation include: i) training our model with different hyperparameter configurations—especially changes in $\lambda$ for l2 regularization to assess its impact on the influence of words’ attribute-specific offsets on their overall embeddings; ii) exploring applications to different datasets and topics; iii) more thoughtful selections for “ground truth” to evaluate results, and iv) comparing the results of our model to other temporal embedding methods. We invite researchers to build on these efforts999Our code can be found here: https://github.com/ngillani/dynamic-word-emb. to further shed light on which language models are most appropriate for capturing changing perceptions and attitudes in our digital public sphere.

Acknowledgments

We thank Doug Beeferman, Prashanth Vijayaraghavan and David McClure for their valuable input on this project. We also thank Nikhil Garg for sharing prior results from his work.