TL;DR
This paper uses complex network analysis to explore the structural differences in the language produced by simultaneous interpreters (SI) and consecutive interpreters (CI). They found that CI outputs exhibit denser, more interconnected networks with a greater reliance on function words, suggesting a different cognitive load and processing strategy compared to SI.
What Problem Does This Paper Address?
Interpreting, especially simultaneous (SI) and consecutive (CI) interpreting, is a cognitively demanding task. While existing models describe the processing mechanisms, a macroscopic view of the outputs of these different interpreting types has been limited. The authors aim to bridge this gap by using complex network analysis to quantitatively characterize the structural properties of interpreted language, providing insights into the underlying cognitive mechanisms and how they manifest in the final output. They want to understand if and how the distinct cognitive demands of SI and CI lead to observable differences in the linguistic structure of their outputs.
Methods
The core of this research is the application of complex network analysis to linguistic data. Specifically, they constructed syntactic dependency networks from transcribed interpreted speeches. In these networks:
- Vertices represent individual words.
- Edges represent the syntactic dependencies between words (e.g., subject-verb, adjective-noun).
They then analyzed various network parameters, including:
- Degree: How many other words a given word is directly connected to.
- Clustering Coefficient: How interconnected a word's neighbors are.
- Density: The overall interconnectedness of the network.
- Average Path Length: The average shortest distance between any two words.
- Centrality Measures (Degree and Betweenness): Identifying "hub" words that are crucial for network structure.
They compared these network properties between SI and CI outputs, using ANCOVAs to control for input text characteristics. A novel aspect is the focus on the syntactic dependency network, which is less influenced by text content compared to co-occurrence or semantic networks. They also analyzed the proportion of function words within the central vertices (hubs) of these networks.
Datasets
The study used a corpus of transcribed real-world interpretations and their corresponding input texts.
- CI Corpus: Speeches by Chinese Premiers at annual press conferences (12 input texts, 12 output texts).
- SI Corpus: Keynote speeches delivered by Chinese government leaders at international forums (12 input texts, 12 output texts).
All interpretations were from Mandarin Chinese to English. The input texts for SI and CI were matched for formality, topic, and lexical diversity to ensure comparability.
Key Insights and Ideas
- Interpreting Outputs as Networks: The paper frames interpreted language as a complex network, allowing for a holistic, structural analysis beyond individual word or sentence features.
- Syntactic Structure Reflects Cognitive Load: Differences in network properties (like density and connectivity) are linked to the cognitive demands of SI and CI, suggesting that the way interpreters process and produce language leaves a distinct structural "fingerprint."
- Function Words as Cognitive Markers: The increased prominence of function words in CI network hubs suggests they play a more critical role in structuring the output when cognitive load is high and memory demands are significant.
- CI: Denser, More Connected, Less Lexically Rich: CI outputs show a more tightly knit syntactic structure, potentially to manage cognitive load by relying on established patterns and reducing the need for novel lexical choices.
- SI: More Lexically Diverse, Less Centralized Hubs: SI outputs, while also exhibiting complex network properties, lean towards greater vocabulary richness and a different distribution of key structural words.
Key Results
- Scale-Free and Small-World Properties: Both SI and CI output networks exhibit scale-free and small-world structures, confirming these are general properties of human language networks.
- CI Networks are Denser and More Connected: CI output networks had significantly higher average degrees, clustering coefficients, and density compared to SI networks. This indicates better connectivity and transitivity.
- SI Networks Have More Vertices (Higher Vocabulary Richness): SI output networks generally had more vertices, suggesting a greater degree of vocabulary richness or lexical diversity compared to CI.
- Function Words are More Central in CI: The percentage of function words among the top central vertices (hubs) was significantly higher in CI outputs than in SI outputs, for both degree and betweenness centrality.
- No Significant Difference in Path Length or Diameter: The average shortest path length and diameter were similar between SI and CI networks, suggesting macro-level connectivity is comparable.
Limitations and Future Work
- Focus on Syntactic Dependency: While this choice minimizes content bias, it means other linguistic features (e.g., semantic, pragmatic) are not directly captured.
- Limited Corpus Size: Although efforts were made to ensure comparability, the corpus size, while substantial, could be expanded for broader generalizability.
- Future Work: The authors suggest that this complex network approach can be integrated with other quantitative methods (like dependency distance and lexical analysis) to provide a more comprehensive understanding of interpreting. They also propose that these findings could inform the development of AI systems for machine translation by incorporating cognitive factors.
Summary updated