Case Vectors: Spatial Representations of the Law Using Document Embeddings

Elliott Ash, Daniel L. ChenSSRN 2018 · DOI

TL;DR

This paper introduces "case vectors," a novel method using document embeddings to represent legal cases as points in a high-dimensional space. This allows for quantitative analysis of legal language, revealing how courts, time, and topics influence judicial writing, and offering a new lens for understanding judicial reasoning and relationships.

What Problem Does This Paper Address?

Traditional methods for analyzing legal language often rely on manual coding of small datasets, which is time-consuming and limits the scope of inquiry. While some computational approaches exist, they often struggle to capture the nuanced semantic relationships within legal texts. This paper aims to bridge this gap by leveraging recent advancements in natural language processing, specifically document embeddings, to create a spatially organized representation of legal cases. This spatial representation, or "case space," can then be analyzed to uncover patterns and relationships that were previously difficult to quantify.

Methods

The core of this research lies in the application of document embeddings, specifically the Doc2Vec model. This technique translates entire documents (in this case, legal opinions) into dense numerical vectors. The process involves training a model to predict words based on their context and the overall document, resulting in vectors where semantically similar documents are located closer together in a geometric space.

To analyze judicial behavior, the authors aggregate these document vectors. They create judge vectors by averaging the document vectors of all cases authored by a particular judge. To isolate specific influences, they employ centering techniques, removing variations attributed to court, topic, and year before aggregating judge vectors. This allows them to focus on the "ideological component" of judicial language. For visualization, they use t-SNE plots to project these high-dimensional vectors into a 2D space, preserving relative distances between points.

Datasets

The study utilizes a comprehensive corpus of U.S. Supreme Court cases and U.S. Circuit Court cases spanning from 1887 to 2013. This dataset is augmented with detailed metadata, including court, date, case topic (categorized using Donald Songer's 7-category "General Issue" designation), and authoring judge. Biographical information on judges, such as birth date, gender, and appointing president, is also incorporated.

Key Insights and Ideas

  • Legal Language as a Spatial Construct: The paper demonstrates that legal opinions can be effectively mapped into a geometric space, where proximity signifies linguistic similarity and potentially shared reasoning.
  • Uncovering Hidden Structures: Document embeddings reveal systematic differences in legal language across courts, time periods, and legal topics, even when controlling for other factors.
  • Judge-Specific Language Signatures: While not perfectly capturing political leanings, judge vectors do reflect generational differences and can distinguish judges based on their writing styles.
  • Limitations of Simple Embeddings: The study highlights that basic document embeddings, while powerful, may not fully capture complex ideological nuances or the impact of specific legal arguments (e.g., ruling for or against a plaintiff).

Key Results

  • Court Differentiation: Figure 1 shows that document vectors, when centered by topic and year and averaged by judge, effectively separate different courts, indicating distinct legal language patterns.
  • Temporal Trends: Figure 2 reveals a "steady linear development of case law" across decades when controlling for court and topic, suggesting that embeddings capture systematic temporal shifts in legal language.
  • Topic Discrimination: Figure 3 demonstrates that embeddings can discriminate between different legal topics, effectively capturing language variations across issue areas.
  • Political Party Inconclusiveness: Figure 4 shows that when controlling for court, topic, and year, judge vectors do not clearly separate judges by political party (Democrat vs. Republican), suggesting that judicial language might be less polarized than expected or that the embeddings aren't sensitive enough to capture this.
  • Generational Differences: Figure 5 reveals clear segmentation across birth cohorts, indicating that the vectorized language effectively captures differences in legal language used by judges from different generations.
  • Law School Inconclusiveness: Figure 6, similar to political party, shows apparent random distributions across the space in terms of law school, suggesting that this factor doesn't strongly influence the language style captured by the embeddings.
  • Court Similarities: Table 1 shows that the D.C. Circuit is most similar to the Supreme Court, and the 11th Circuit is similar to the 5th Circuit, aligning with intuitive geographical and historical connections.
  • Judge Similarities (Mixed Results): Tables 2 and 3 show some intuitive pairings (Scalia with Thomas) but also counter-intuitive ones, suggesting that judge vectors may not always reliably capture similarities between Supreme Court judges, possibly due to the limited number of opinions authored. However, analysis of circuit court judges (Table 5) shows more informative connections, particularly for judges known for economic analysis.

Limitations and Future Work

The authors acknowledge that the current approach, using a bag-of-words model, might not fully capture ideological differences that are encoded in phrases or grammatical structures. They suggest that richer representations, such as those based on dependency parsing, might be needed.

Future work could involve:

  • Structured Group Embeddings: Developing models that explicitly incorporate time and group components to better represent dynamic language evolution and judge relationships.
  • Citation Networks: Enriching case representations with citation information to capture the "precedent space" and identify complementary or substitutable legal arguments.
  • Language-Based Metrics of Implicit Bias: Applying these embedding techniques to detect sentiment and bias in judicial language, potentially revealing implicit biases related to race, gender, or policy preferences.
  • Judge Embeddings with Deeper Features: Implementing deep learning models that incorporate judge identity, ruling direction, and lower-court decision features to create more predictive and informative judge embeddings. This could enable counterfactual simulations to understand how judicial decisions might change with different judges.

Summary updated

BibTeX
@misc{ash2018case,
  author = {Elliott Ash and Daniel L. Chen},
  title = {{Case Vectors: Spatial Representations of the Law Using Document Embeddings}},
  year = {2018},
  doi = {10.2139/ssrn.3204926},
  howpublished = {SSRN},
}

Loading related content...

Figure