A lexical network approach to second language development

This research explores second language (L2) acquisition through the lens of Complex Dynamic System Theory (CDST), focusing on lexical development in L2 Chinese learners. CDST emphasizes the dynamic and complex nature of language learning, highlighting features like non-linearity and self-organization. Previous research using CDST has employed lexical and syntactic complexity indices (vocabulary richness, mean length of units, phrasal/clausal features), revealing dynamic interlanguage properties. However, challenges remain in consistently measuring linguistic features and exploring interactions between language subsystems. For instance, some studies found limited variation in word length and frequency across proficiency levels, while others highlighted the effectiveness of indices like the Guiraud index and chunks in predicting proficiency. The study proposes to address these limitations by using complex network science. Complex networks, composed of nodes (elements) and links (interactions), offer a powerful tool to model complex adaptive systems. Language, a complex adaptive system (CAS), can be represented as a network at various levels (phonological, morphological, syntactic, semantic). The complex network approach captures connections between language system properties and cognition, and offers methods for processing large-scale data. Previous research has applied network analysis to syntactic structures (dependency treebanks), successfully differentiating L2 proficiency levels. Yet, this approach relies on resource-intensive treebank creation and is affected by annotation schemes. This study, therefore, adopts a simpler word co-occurrence (bigram) network approach, leveraging the increasing recognition of bigrams' importance in distinguishing L2 proficiency levels. Bigrams capture syntagmatic relationships and are reliable predictors of proficiency. Moreover, the usage-based perspective of language development emphasizes the role of multi-word expressions/chunks and the efficiency of rapid chunking in language acquisition. Bigram networks, based on word co-occurrences, efficiently capture intrinsic word relationships and provide insights into lexical acquisition. This study compares the effectiveness of a bigram lexical network approach with traditional lexical complexity indices in differentiating L2 Chinese proficiency levels, aiming to address the limitations of previous research and contribute to a more nuanced understanding of L2 lexical development.

The paper reviews existing literature on second language acquisition (SLA) within the framework of Complex Dynamic System Theory (CDST) and the application of complex network analysis to language. It highlights the limitations of previous studies that used traditional measures of lexical complexity such as vocabulary richness, mean length of utterance, and type-token ratio, pointing out inconsistencies in the results obtained using these measures across different studies and datasets. The review underscores the need for more comprehensive and consistent measures of lexical complexity and the potential of complex network analysis to address these shortcomings. The literature on applying network analysis to syntactic structures in second language learning is discussed, noting the success of these approaches in distinguishing learners of different proficiency levels. However, the limitations of relying on dependency treebanks and the influence of annotation schemes are acknowledged. The paper then justifies the use of word co-occurrence networks (based on bigrams) as a simpler and more readily applicable alternative, citing research highlighting the significance of bigrams and multi-word expressions in capturing syntagmatic relationships and predicting L2 proficiency. Finally, the literature connecting language processing models, spreading activation, and the representation of language as a cognitive network is reviewed to establish the theoretical grounding for using a bigram network approach.

The study utilized the Guangwai-Lancaster Chinese Learner Corpus (GLCLC), a large corpus of Mandarin Chinese written texts produced by learners of varying proficiency levels and native language backgrounds. Three proficiency levels (beginner, intermediate, advanced) were represented, with sub-corpora selected to include approximately 10,000 Chinese characters each from learners with Thai and Romance language backgrounds, along with native Chinese texts for comparison. Narrative texts were chosen to minimize genre-related effects. The NIPIR-ICTCLAS tool segmented words and annotated parts of speech. Lexical networks were constructed using bigrams (word co-occurrences) and analyzed using Pajek software. Seven network parameters were calculated: network size (n), number of edges (m), network density (ND), network centralization (NC), clustering coefficient (C), average path length (<l>), and number of clusters (Cl). The S-Small-worldness parameter was calculated to assess the small-world property of the networks. Additionally, seven traditional lexical complexity indices were calculated using CTAP software: Uber index (U), Guiraud index, average word length (AWL), lexical density (LD), number of hapax legomena (HL), long word percentage (LWP), and type-token ratio of bigrams (TTR-bigram). Hierarchical clustering with SPSS 20 was used to assess the ability of both the network parameters and the lexical complexity indices to distinguish the different proficiency levels. Information gain was calculated to determine the predictive power of each index.

The results revealed that all seven L2 Chinese lexical networks exhibited characteristics of complex networks (small-world and scale-free properties), significantly different from the tree-like networks observed in young children's language. Higher proficiency levels were associated with more nodes, edges, and clusters. However, network density, centralization, and clustering coefficient decreased with increasing proficiency, possibly due to the fixed corpus size. Hierarchical clustering based on network parameters successfully grouped learners by proficiency level, with Native Chinese forming a separate cluster. Information gain analysis indicated that network size (n), number of edges (m), network density (ND), and clustering coefficient (C) were the strongest predictors of proficiency. In contrast, hierarchical clustering based on traditional lexical complexity indices successfully separated beginners and native speakers but failed to reliably distinguish intermediate and advanced learners. The Guiraud index and TTR-bigram showed the strongest predictive power among the lexical complexity indices. The comparison of clustering results showed that the lexical network-based method performed better than the lexical complexity-based method in distinguishing proficiency levels, and it was less influenced by native language background compared to the lexical complexity method.

The findings support the use of lexical network analysis as a powerful tool for understanding L2 lexical development. The superior performance of the network-based method over traditional lexical complexity indices highlights the importance of considering word associations and network structure in addition to individual word frequencies. The consistent complex network characteristics across proficiency levels suggest that L2 lexical development follows principles similar to those observed in other complex systems. The decrease in network density, centralization, and clustering coefficient with increasing proficiency could reflect the increasing specialization and organization of the learner's lexical knowledge. The study's limitations, primarily the reliance on a specific corpus and the potential influence of corpus size on network parameter patterns, are acknowledged. Future research should explore diverse corpora and investigate the effects of corpus size systematically. The study’s success in identifying proficiency levels suggests the potential of lexical network analysis for pedagogical applications, such as personalized feedback and targeted vocabulary instruction.

This study demonstrated that lexical networks provide a valuable approach to understanding L2 lexical development. The network-based method effectively distinguishes proficiency levels and offers a more comprehensive view than traditional lexical complexity indices. The study’s findings underscore the importance of incorporating both lexical frequency and network features within the CDST framework for a more complete understanding of L2 lexical richness and development. Future research could extend this work by applying the multi-layer complex network approach to capture interactions between language users and language units, and explore the role of generative AI in language learning.

The study's findings are based on a specific corpus (GLCLC) with a focus on written narratives from learners of Thai and Romance language backgrounds. The fixed corpus size of 10,000 characters might influence the observed patterns in network parameters. Future research should explore the generalizability of the findings across different corpora, genres, and learner populations. The study also relied on a specific bigram network approach; alternative network models or the incorporation of higher-order n-grams could yield additional insights.