Subjects: Computer Science >> Computer Software Subjects: Linguistics and Applied Linguistics >> Linguistics and Applied Linguistics submitted time 2024-04-21
Abstract： This slide presentation describes the research on Guiding Large Language Models to Generate Computer-Parsable Content in terms of Background, Motivation, Method, Effect, Prospect and Acknowledgements. For the full paper, please refer to: https://arxiv.org/abs/2404.05499
Peer Review Status:
Awaiting Review
Subjects: Computer Science >> Computer Software Subjects: Linguistics and Applied Linguistics >> Linguistics and Applied Linguistics submitted time 2024-04-07
Abstract： Large language models (LLMs) have demonstrated remarkable capabilities in learning patterns from massive text corpora, including word relationships, sentence structures, and even complex semantic and pragmatic information. However, it remains challenging to induce pre-trained language models to generate structured content that strictly follows specific conventions.We propose a scheme for guiding LLMs to generate highly usable content for computers without the need for fine-tuning and additional neural network inference, by introducing coroutine-based content generation constraints through a pre-agreed context-free grammar (CFG), which guides the autoregressive model Transformer to sample the correct tokens during its decoding phase to form a program-compliant form in the decoding phase of the autoregressive model Transformer to form a formal language that conforms to the program conventions. This will effectively improve the stability and consistency of LLMs in generating target data structures, types or instructions, and reduce the difficulty of application development and integration.We first verified that the error rate of models such as GPT-2 and Gemma reaches 95% when the length of the generated DSLs are greater than 36 and 282, respectively, through the experiment of matching bracket pairs , which illustrates the performance problem of some current LLMs in the generation of specific DSLs. We also present YieldLang, a coroutine-based DSL generation framework, and conduct experiments using LLMs on multiple task datasets, including tasks such as JSON, Mermaid flowchart, and function call expression generation. These experiments show that the approach in this paper improves its accuracy by a factor of 1.09 to 11.6 compared to the benchmarks, and in the best case is able to reduce the number of samples used by the LLMs to generate JSON to about 16.5% of the benchmarks, which will effectively improve the usability of the content generated by the LLMs for computer programs.
Peer Review Status:Awaiting Review
Subjects: Linguistics and Applied Linguistics >> Linguistics and Applied Linguistics submitted time 2024-01-26
Abstract： Rhetoric in language, like air, is ubiquitous. It is not only presented in the form of narrow rhetoric (rhetorical devices), but from a broad rhetorical perspective, rhetoric is also implicit in all sentences, inherently encompassing the domain of narrow rhetoric. This article starts with rhetorical devices, explains the source language and target language in a broad sense of rhetoric, explores the connection between the two, analyzes the process of sentence translation from a broad rhetorical perspective, and proposes dynamic principles for measuring the quality of sentence translation.
Peer Review Status:Awaiting Review
Subjects: Linguistics and Applied Linguistics >> Linguistics and Applied Linguistics Subjects: Computer Science >> Natural Language Understanding and Machine Translation submitted time 2024-01-11
Abstract： The research and engineering paradigm of natural language processing has been shifted with the rapid development of large languages models represented by the GPT series. It makes a significant impact on the related fields such as healthcare, education, judiciary and finance. At the same time, it also brings new possibilities for linguistics, the study of language itself. In this paper, we employ GPT4, Baichuan2 as well as ChatGLM3 and investigate their abilities of analyzing complex linguistic phenomena, taking ambiguity as an example. The experimental results show that GPT4 can effectively perceive and understand complex linguistic phenomena by integrating ambiguity resolution and syntactic analysis. For Baichuan2, if it is guided properly via prompt engineering, its analytical ability can be improved without parameter optimization. In addition, the relationship between linguistic phenomena and large language models can be visually demonstrated by monitoring the internal features and neuron activities of the models when processing ambiguous sentences in different context. In general, our experiments indicate that large language models are beneficial to better understanding the analyzing complex linguistic phenomena, hence providing new alternatives for linguistic research.
Subjects: Linguistics and Applied Linguistics >> Linguistics and Applied Linguistics submitted time 2024-01-05
Abstract： Generally, a word’s meaning consists of at least two components. The first is denotative meaning, representing the definitional meaning found in dictionaries and serving as the word’s fundamental meaning. The second component involves semantics that a word “absorbs” from its linguistic context, not constrained by definitions; this is known as semantic prosody, described as “a consistent aura of meaning with which a form is imbued by its collocates” (Louw, 1993, p. 157). While theories and empirical studies have shed light on mechanisms supporting the acquisition of the first word meaning component, the acquisition of the connotative meaning engendered by semantic prosody has been overlooked. It remains unclear whether readers can unconsciously acquire the semantic prosody (or emotional connotations) of a novel word after encountering it consistently in a context with a strong emotional polarity.
Against this backdrop, we conducted a word learning experiment, manipulating context emotionality (negative vs. neutral vs. positive) and context variability (same-repeated vs. varied contexts) as crucial contextual variables. This aimed to address two understudied questions in vocabulary acquisition: (1) Does transfer of affect to a word from its linguistic context take place through reading exposures, facilitating the acquisition of semantic prosody for the word? If so, is such transfer influenced by context variability? (2) Does the acquired semantic prosody for words affect the acquisition of word forms and meanings, and is this acquisition modulated by context variability? This experiment involved two sessions: a reading-and-learning phase and a testing phase. During the reading-and-learning session, participants read emotionally charged passages, simultaneously learning embedded target words. The testing session included an immediate posttest, incorporating four vocabulary tests—valence rating, orthographic choice, definition matching, and definition generation. A total of 196 Chinese speakers participated in the experiment.
Mixed-effects models were utilized to analyze data from the valence rating task and the other three vocabulary knowledge tests. The findings revealed that, within the same-repeated context, manipulating context emotionality (positive vs. neutral vs. negative) significantly influenced valence ratings, showing significantly higher ratings in the positive condition compared to neutral and negative conditions. Conversely, in the varied context, no significant differences in valence ratings were observed. This result supports the hypothesis of the “Double-Jujube Tree” effect, emphasizing the effect of repetitive texts compared to multiple texts. However, in the varied context, valence ratings played a role in influencing participants’ performances in the vocabulary tests, leading to better outcomes as valence ratings increased. In the same-repeated context, valence ratings had minimal effect on accuracy in the orthographic choice test and the definition prompting test.
We posit that the effective mechanism for learning the semantic-prosody-engendered connotations of words involves the transfer of affect from their collocations. However, this transfer seems to be contingent on context variability, occurring only in the same-repeated context and not in the varied context. Furthermore, we illustrate that the emotionality of context influences the quality of both orthographic and semantic word learning, with words being better learned in positive contexts as opposed to negative or neutral ones.
Hits
Hits
Downloads
Comment
YES