There has been an increasing level of attention to and discussion of gender bias in sports, ranging from differences in pay and prize money111“U.S. Women, Fighting for Equal Pay, Win Easily as Fans Show Support”, The New York Times, April 6, 2016. http://www.nytimes.com/2016/04/07/sports/soccer/uswnt-colombia-friendly-equal-pay-complaint.html to different levels of focus on off-court topics in interviews by journalists. With respect to the latter, Cover the Athlete,222http://covertheathlete.com/ an initiative that urges the media to focus on sport performance, suggests that female athletes tend to get more “sexist commentary” and “inappropriate interview questions” than males do; the organization put out an attention-getting video in 2015 purportedly showing male athletes’ awkward reactions to receiving questions like those asked of female athletes. However, it is not universally acknowledged that female athletes attract more attention for off-court activities. For instance, a manual analysis by Kian:JournalOfBroadcastingElectronicMedia:2009 [Kian:JournalOfBroadcastingElectronicMedia:2009] of online articles revealed significantly more descriptors associated with the physical appearance and personal lives of male basketball players in comparison to female ones.
Transcripts of pre- or post-game press conferences offer an opportunity to determine quantitatively and in a data-driven manner how different are the questions which journalists pose to male players from those they pose to female players. Here are examples of a game-related and a non-game-relevant question, respectively, drawn from actual tennis interviews:
What happened in that fifth set, the first three games?
After practice, can you put tennis a little bit behind you and have dinner, shopping, have a little bit of fun?
To quantify gender discrepancies in questions, we propose a statistical language-model-based approach to measure how game-related questions are. In order to make such an approach effective, we restrict our attention in this study to a single sport—tennis—so that mere variations in the lingo of different sports do not introduce extra noise in our language models. Tennis is also useful for our investigation because, as kian2011 [kian2011] noted, it “marks the only professional sports where male and female athletes generally receive similar amounts of overall broadcast media coverage during the major tournaments.”
Using our methodology, we are able to quantify gender bias with respect to how game-related interview questions are. We also provide a more fine-grained analysis of how gender differences in journalistic questioning are displayed under various scenarios. To help with further analysis of interview questions and answers, we introduce a dataset of tennis post-match interview transcripts along with corresponding match information.333Dataset available at http://www.cs.cornell.edu/~liye/tennis.html
2 Related Work
In contrast with our work, prior investigations of bias in sport journalism rely on manual coding or are based on simple lists of manually defined keywords. These focus on bias with respect to race, nationality, and gender [Rainville and McCormick1977, Sabo et al.1996, Eastman and Billings2001, Bruce2004, Billings2008, Kian and Clavio2011, Ličen and Billings2013]; see van2010race [van2010race] for a review.
Much of the work on gender bias in sports reporting has focused on “air-time” [Eastman and Billings2000, Higgs et al.2003]. Other studies looked at stereotypical descriptions and framing [Messner et al.1993, Jones2004, Angelini and Billings2010, Kian et al.2009]. For surveys, see Knight:SexRoles [Knight:SexRoles] or Kaskan:SexRoles:2014 [Kaskan:SexRoles:2014], inter alia. Several studies have focused on the particular case of gender-correlated differences in tennis coverage [Hilliard1984, Vincent et al.2007, Kian and Clavio2011]. We extend this line of work by proposing an automatic way to quantify gender bias in sport journalism.
3 Dataset Description
We collect tennis press-conference transcripts from ASAP Sport’s website (http://www.asapsports.com/), whose tennis collection dates back to 1992 and is still updated for current tournaments. For our study, we take post- game interviews for tennis singles matches played between Jan, 2000 to Oct 18, 2015. We also obtain easily-extractable match information from a dataset provided by Tennis-Data,444 http://www.tennis-data.co.uk/ which covers the majority of the matches played on the men’s side from 2000-2015 and on the women’s side from 2007-2015.
We match interview transcripts with game statistics by date and player name, keeping only the question and answer pairs from games where the statistics are successfully merged. This gives us a dataset consisting of 6467 interview transcripts and a total of 81906 question snippets555Each snippet represents one turn from one journalist. Most question snippets contain at least one question, although some could be merely clarifications or comments. Note that reporter information (who asked which question) is not available in the transcript. posed to 167 female players and 191 male players.
To model tennis-game-specific language, we use live text play-by-play commentaries collected from the website Sports Mole (http://www.sportsmole.co.uk/). These tend to be short, averaging around 40 words. Here is a sample, taken from the Federer-Murray match at the 2015 Wimbledon semi-final:666 http://www.sportsmole.co.uk/tennis/wimbledon/live-commentary/live-commentary-roger-federer-vs-andy-murray-as-it-happened_232822.html
“The serve-and-volley is being used frequently by Federer and it’s enabling him to take control behind his own serve. Three game points are earned before an ace down the middle seal [sic] the love hold.”
For our analysis, we create a gender-balanced set of commentaries consisting of descriptions for 1981 games played for each gender.
As a preliminary step, we apply a word-level analysis to understand if there appear to be differences in word usage when journalists interview male players compared to female players. We then introduce our method for quantifying the degree to which a question is game-related, which we will use to explore gender differences.
4.1 Preliminary Analysis
To compare word usage in questions, we consider, for each word , the percentage of players who have ever been asked a question containing . We then consider words with the greatest difference in percentage between male and female players.777Words that are gender-specific (like ‘her’) are manually discarded. The top distinguishing words, which are listed below in descending order of percentage difference, seem to suggest that questions journalists pose to male players are more game-related:
- Male players:
clay, challenger(s), tie, sets, practiced, tiebreaker, maybe, see, impression, serve, history, volley, chance, height, support, shots, server(s), greatest, way, tiebreaks, tiebreakers, era, lucky, luck;888
It is interesting, but beyond the scope of this paper, to speculate on reasons why “luck” and “lucky” skew so strongly male.
- Female players:
yet, new, nervous, improve, seed, friends, nerves, mom, every, matter, become, meet, winning, type, won, draw, found, champion, stop, fight, wind, though, father, thing, love.
4.2 Game Language Model
To quantify how game-related a question is
in a data-driven fashion,
we train a bigram language model using KenLM999KenLM (https://kheafield.com/code/kenlm/ )
estimates language models using modified Kneser-Ney smoothing without pruning.
) estimates language models using modified Kneser-Ney smoothing without pruning.[Heafield et al.2013] on the gender-balanced set of live-text play-by-play commentaries introduced in Section 3.
For an individual question , we measure its perplexity with respect to this game language model as an indication of how game-related the question is: the higher the perplexity value, the less game-related the question. Perplexity, a standard measure of language-model fit [Jelinek et al.1977], is defined as follows for an -word sequence :
Below are some sample questions of low-perplexity and high-perplexity values:
|Low||What about your serve, Rafa?|
|The tiebreak, was that the key to the match?|
|High||Who designed your clothes today?|
|Do you normally watch horror films to relax?|
In this section we use the game language model to quantify gender-based bias in questions. We then compare the extent to which this difference depends of various factors, such as question type, game outcome, or player rank.
5.1 Main Result: Males vs. Females
We first compute perplexities for each individual question101010We identify individual
questions simply by looking for ‘?’. and then group the question instances according to the interviewee’s gender class.
Throughout we use the Mann-Whitney statistical significance test,111111
We used this non-parametric significance test instead of the -test because it doesn’t assume the samples to be normally distributed.
-test because it doesn’t assume the samples to be normally distributed.unless otherwise noted.
Comparing perplexity values between the two groups, we find that the mean perplexity of questions posed to male players is significantly smaller (-value 0.001) than that of questions posed to female players. This suggests that the questions male athletes receive are more game-related.
However, the number of interviews each player participates in varies greatly, with highly interviewed players answering as many as thousands of questions while some lesser-known players have fewer than 10 interview questions in the dataset. Thus it is conceivable that the difference is simply explained by questions asked to a few prolific players. To test whether this is the case, or whether the observation is more general, we micro-average the perplexities by player: for each of the 167 male players and 143 females who have at least 10 questions in our dataset, we consider the average perplexities of the questions they receive. Comparing these micro-averages, we find that it is still the case that questions posed to male players are significantly closer to game language (-value 0.05), indicating that the observed gender difference is not simply explained by a few highly interviewed players.
5.2 Relation to Other Factors
We further investigate how the level of gender bias is tied to different factors: how typical the question is (section 5.2.1), the ranking of the player (section 5.2.2), and whether the player won or lost the match (section 5.2.3). For all the following experiments, we use per-question perplexity for comparisons: per-player perplexity is not used due to limited sample size.
5.2.1 Typical vs. Atypical Questions
One might wonder whether the perplexity disparities we see in questions asked of female vs. male players are due to “off-the-wall” queries, rather than to those that are more typical in post-match interviews. We therefore use a data-driven approach to distinguish between typical and atypical questions.
For any given question, we consider how frequently its words appear in post-match press conferences in general. Specifically, we take the set of all questions as the set of documents, . We compute the inverse document frequency for each word (after stemming) that has appeared in our dataset, excluding the set consisting of stop words and a special token for entity names.121212We replace capitalized words and phrases with “NOUN”; for each word at the beginning of a sentence (which is always capitalized), we check whether it is a dictionary word. For a question that contains the set of unique words , we compute its atypicality score as:
We use the overall mean atypicality score of the entire question dataset as the cutoff point: questions with scores above the overall mean are considered atypical and the rest are considered typical.131313Questions consisting only of stop words and player or tournament names are still considered typical questions, even though they do not have an atypicality score. Below are some examples:
|Typical||Have you played each other before?|
|How do you feel playing here?|
|Atypical||What about your haircut?|
|Are you a vodka drinker?|
Figure 1 shows that a gender bias with respect to whether game-related language is used exists for both typical and atypical questions. However, additional analysis reveals that the difference in mean perplexity values between genders is highly statistically significantly larger for atypical questions, suggesting that gender bias is more salient among the more unusual queries.
5.2.2 Player Ranking
Higher ranked players generally attract more media attention, and therefore may be targeted differently by journalists. To understand the effect of player ranking, we divide players into two groups: top 10 players and the rest. For our analysis, we use the ranking of the player at the time the interview was conducted. (It is therefore possible that questions posed to the same player but at different times could fall into different ranking groups due to ranking fluctuations over time.) We find that questions to male players are significantly closer to game language regardless of player ranking (-value 0.001, Figure 2).
Furthermore, if we focus only on players who have ranked both in and outside the top 10 in our dataset, and pair the questions asked to them when they were higher-ranked to the questions asked when their ranking was lower, we find that there is no significant difference between questions asked to male athletes when they were in different ranking groups (Wilcoxon signed-rank -value 0.05). However, the difference is significant for females (Wilcoxon signed-rank -value 0.01), suggesting that gender bias may be more salient for lower ranked players as questions to lower-ranked female athletes tend to be less game-related.
While one might expect that star players would receive more off-court questions (yielding higher perplexities), the perplexity values for questions posed to top 10 players are actually lower regardless of gender. This may be because the training data for our language model is more focused on specific points played in matches, and may not be representative of tennis-related questions that are more general (e.g., longer-term career goals, personal records, injuries). In other words, our result suggests that journalists may attend more to the specifics of the games of higher ranked players, posing more specific questions about points played in the match during interviews.
5.2.3 Winning vs. Losing
While it is reasonable to expect that whether the interviewee won or lost would affect how game-related the questions are, the difference in mean perplexity for males and females conditioned on win/loss game outcome are comparable. In addition, for both male players and female players, there is no significant difference observed between the paired set of questions asked in winning interviews and the losing ones (Wilcoxon signed-rank -value 0.05), controlling for both player and season.141414We pair each question asked to a given player when winning to one question posed to the same player in the same calendar year when losing to construct the paired set of winning and losing questions for each gender. This suggests that that game result may not be a factor affecting how game-related the interview questions are.
6 Concluding discussion
In this work we propose a language-model based approach to quantify gender bias in the interview questions tennis players receive. We find that questions to male athletes are generally more game-related. The difference is more salient among the unusual questions in press conferences, and for lower-ranked players.
However, this preliminary study has a number of limitations. We have considered only a single sport. In addition, our dataset does not contain any information about who asked which question, which makes us unable to control for any idiosyncrasies of specific journalists. For example, it is conceivable that the disparities we observe are explained by differences in the journalists that are assigned to conduct the respective interviews.
In this work, we limit our scope to bias in terms of game-related language, not considering differences (or similarities) that may exist in other dimensions. Further studies may use a similar approach to quantify and explore differences in other dimensions, by using language models specifically trained to model other domains of interests, which may provide a more comprehensive view of how questions differ when targeting different groups.
Furthermore, our main focus is on questions asked during press conferences; we have not looked at the players’ responses. The transcripts data, which we release publicly, may provide opportunities for further studies.
We thank the anonymous reviewers and the participants in the Fall 2015 edition of the course “Natural Language Processing and Social Interaction” for helpful comments and discussion. This research was supported in part by a Discovery and Innovation Research Seed award from the Office of the Vice Provost for Research at Cornell.
- [Angelini and Billings2010] James R. Angelini and Andrew C. Billings. An agenda that sets the frames: Gender, language, and NBC’s americanized olympic telecast. Journal of Language and Social Psychology, 2010.
- [Billings2008] Andrew C. Billings. Olympic media: Inside the biggest show on television. Routledge, 2008.
- [Bruce2004] Toni Bruce. Marking the boundaries of the ‘normal’ in televised sports: The play-by-play of race. Media, Culture & Society, 26(6):861–879, 2004.
- [Eastman and Billings2000] Susan Tyler Eastman and Andrew C. Billings. Sportscasting and sports reporting: The power of gender bias. Journal of Sport & Social Issues, 24(2):192–213, 2000.
- [Eastman and Billings2001] Susan Tyler Eastman and Andrew C. Billings. Biased voices of sports: Racial and gender stereotyping in college basketball announcing. Howard Journal of Communication, 12(4):183––201, 2001.
- [Heafield et al.2013] Kenneth Heafield, Ivan Pouzyrevsky, Jonathan H. Clark, and Philipp Koehn. Scalable modified Kneser-Ney language model estimation. In Proceedings of the ACL, pages 690–696, August 2013.
- [Higgs et al.2003] Catriona T. Higgs, Karen H. Weiller, and Scott B Martin. Gender bias in the 1996 olympic games: A comparative analysis. Journal of Sport & Social Issues, 27(1):52–64, 2003.
- [Hilliard1984] Dan C. Hilliard. Media images of male and female professional athletes: An interpretive analysis of magazine articles. Sociology of Sport Journal, 1:251–262, 1984.
- [Jelinek et al.1977] Fred Jelinek, Robert L. Mercer, Lalit R. Bahl, and James K. Baker. Perplexity — a measure of the difficulty of speech recognition tasks. The Journal of the Acoustical Society of America, 62(S1):S63–S63, 1977.
- [Jones2004] Dianne Jones. Half the story? Olympic women on ABC news online. Media International Australia, 110(1):132–146, 2004.
- [Kaskan and Ho2014] Emily R. Kaskan and Ivy K. Ho. Microaggressions and female athletes. Sex Roles, 74(7-8):275–287, 11 2014.
- [Kian and Clavio2011] Edward M. Kian and Galen Clavio. A comparison of online media and traditional newspaper coverage of the men’s and women’s U.S. open tennis tournaments. Journal of Sports Media, 2011.
- [Kian et al.2009] Edward (Ted) M. Kian, Mondello Michael, and Vincent John. ESPN — the women’s sports network? a content analysis of internet coverage of March madness. Journal of Broadcasting & Electronic Media, 53(3):477–495, 9 2009.
- [Knight and Giuliano2001] Jennifer L. Knight and Traci A. Giuliano. He’s a Laker; she’s a “looker”: The consequences of gender-stereotypical portrayals of male and female athletes by the print media. Sex Roles, 45(3-4):217–229, 2001.
- [Ličen and Billings2013] Simon Ličen and Andrew C. Billings. Cheering for ‘our’ champs by watching ‘sexy’ female throwers: Representation of nationality and gender in Slovenian 2008 Summer Olympic television coverage. European Journal of Communication, 28(4):379–396, 2013.
- [Messner et al.1993] Michael A. Messner, Margaret Carlisle Duncan, and Kerry Jensen. Separating the men from the girls: The gendered language of televised sports. Gender & Society, 7(1):121–137, 1993.
- [Rainville and McCormick1977] Raymond E. Rainville and Edward McCormick. Extent of covert racial prejudice in pro football announcers’ speech. Journalism & Mass Communication Quarterly, 54(1):20–26, 1977.
- [Sabo et al.1996] Don Sabo, Sue Curry Jansen, Danny Tate, Margaret Carlisle Duncan, and Susan Leggett. Televising international sport: Race, ethnicity, and nationalistic bias. Journal of Sport & Social Issues, 20(1):7–21, 1996.
- [Van Sterkenburg et al.2010] Jacco Van Sterkenburg, Annelies Knoppers, and Sonja De Leeuw. Race, ethnicity, and content analysis of the sports media: A critical reflection. Media, Culture & Society, 32(5):819–839, 2010.
- [Vincent et al.2007] John Vincent, Paul M. Pedersen, Warren A. Whisenant, and Dwayne Massey. Analysing the print media coverage of professional tennis players: British newspaper narratives about female competitors in the Wimbledon championships. International Journal of Sport Management and Marketing, 2(3):281–300, 2007.