Getting on the same page: Comparing Community and Police Perceptions

The work of Serve and Connect strives to foster positive relationships between police and community members. There are many reasons — historical, social, racial, and economic — that have perpetuated poor interactions between first responders and communities, especially communities of color.

One of Serve and Connect’s flagship initiatives is the North Columbia Youth Empowerment Initiative started in November 2018 and supported by funding from the Robert Wood Johnson Foundation. During the Spring and Summer of 2019, the team conducted sixteen interviews (twelve with community members, and four with police representatives) to get a sense of how the initiative was progressing and their impressions of the work and future priorities.

So much qualitative data goes unanalyzed because it can be time-consuming. To analyze this set of data, we used three natural language processing methods: bag-of-words, topic modeling, and sentiment analysis.

BAG OF WORDS-Unigrams. One of the most simple ways to look at the dat was by simple words counts. First, we removed common words like “and” and “the” and then lemmatized the remaining words. Lemmatizing reduces words down to their underlying stem. We compared the frequency of words by role, community and police.

Although simple, there are some interesting insights here. The top four words between groups are identical, which may suggest some initial agreement. We then used the term-frequency, inverse document frequency methods to help correct for words the commonly occur across respondent. This method highlight “important” words that occur frequently.

We can see larger differences here. The police talked about relationships, but also some of the challenges that they encountered, while the community respondents highlighted assets and strengths.

Note: The “Bi” and “lo” phrases refer to a local supermarket.

BAG OF WORDS-Bigrams. Of course, multi-word phrases may be more informative because the sequence in which words occur matters. Therefore, we looked at bigrams, which are two-words sequences. We replicated the above analysis, first, using the simple count method, then again using TF-IDF

Like before, there are a lot of similarities between common phrases. “Safety” is clearly stressed between groups, especially youth safety.

The TF-IDF is much messier. The police respondents talk broadly about strategies, while the community-data appears to much more about the specifics (often citing specific stakeholders). We also used a network plot to see how the phrases were used. We can see the central importance of safety in this work.

TOPIC MODELING. We then wanted to see what the different groups were talking about. Topic modeling presumes that documents (in this case, the interviews) are composed of topics, and topics are composed of closely occurring words.

We use the LDA algorithm to sort occurrences of words into groups. To determine the optimal number of topics, we ran several different model combinations. We used the perplexity score to identify the optimal number of topics to sort the terms into. Looking first at the community respondents, we saw a pretty clear optimization at ten topics.

Like with the bigram analysis, the community topics seems to cluster around specific activities with specific stakeholders.

We took out topic 2 because there were too many terms for it to be interpretable.

The police data showed about eight topics, with the activities of the CRT applying across topics.

Like above, we took out topics three and seven because they had too many terms to be interpretable.

SENTIMENT ANALYSIS. We know that police-community relations can be emotionally volatile, eliciting strong feelings from stakeholders on both sides. We used sentiment analysis to examine the frequency of particular words that had either a positive or negative valence. The graph below identified positive or negative terms that occurred more than four times within the interviews. Generally, the community interviews had a positive sentiment than the police interviews.

CONCLUSIONS. So these interviews are one part of a much larger dataset that Serve and Connect are assembling, included other forms of quantitative and qualitative data. In addition, this data does not reflect the unnumerable connections and conversations that occur “off the record.” With that in mind, this analysis should not be taken representative of the work of Serve and Connect and the state of police-community relations as a whole. However, there are some interesting insights that can inform ongoing work.

What potentially stands our here is that the community members may be better informed about the specifics of the program, including the who, what, and where of programming. The informational pathway may need to go from community → police, as opposed to from the police → community.

There seems to be broad agreement about the goal of community safety, which suggests that the initial visioning work has been successful in getting people aligned in the same direction. Despite that, the police respondents used more negative words that community members, which may point the way to future conversations and acknowledgments around challenges police face. Serve and Connect has already been leading the way in this domain