Computer Speech and Language最新文献

Self-supervised learning (SSL) leverages large datasets of unlabeled speech to reach impressive performance with reduced amounts of annotated data. The high number of proposed approaches fostered the emergence of comprehensive benchmarks that evaluate their performance on a set of downstream tasks exploring various aspects of the speech signal. However, while the number of considered tasks has been growing, most proposals rely upon a single downstream architecture that maps the frozen SSL representations to the task labels. This study examines how benchmarking results are affected by changes in the probing head architecture. Interestingly, we found that altering the downstream architecture structure leads to significant fluctuations in the performance ranking of the evaluated models. Against common practices in speech SSL benchmarking, we evaluate larger-capacity probing heads, showing their impact on performance, inference costs, generalization, and multi-level feature exploitation.

A method for extracting relations from sentences by utilizing their dependency trees to identify key phrases is presented in this paper. Dependency trees are commonly used in natural language processing to represent the grammatical structure of a sentence, and this approach builds upon this representation to extract meaningful relations between phrases. Identifying key phrases is crucial in relation extraction as they often indicate the entities and actions involved in a relation. The method uses community detection algorithms on the dependency tree to identify groups of related words that form key phrases, such as subject-verb-object structures. The experiments on the Semeval-2010 task8 dataset and the TACRED dataset demonstrate that the proposed method outperforms existing baseline methods.

Our work explores the differences between GRU-based and transformer-based approaches in the context of sentiment analysis on text dialog. In addition to the overall performance on the downstream task, we assess the knowledge transfer capabilities of the models by applying a thorough zero-shot analysis at task level, and on the cross-lingual performance between five European languages. The ability to generalize over different tasks and languages is of high importance, as the data needed for a particular application may be scarce or non existent. We perform evaluations on both known benchmark datasets and a novel synthetic dataset for dialog data, containing Romanian call-center conversations. We study the most appropriate combination of synthetic and real data for fine-tuning on the downstream task, enabling our models to perform in low-resource environments. We leverage the informative power of the conversational context, showing that appending the previous four utterances of the same speaker to the input sequence has the greatest benefit on the inference performance. The cross-lingual and cross-task evaluations have shown that the transformer-based models possess superior transfer abilities to the GRU model, especially in the zero-shot setting. Considering its prior intensive fine-tuning on multiple labeled datasets for various tasks, FLAN-T5 excels in the zero-shot task experiments, obtaining a zero-shot accuracy of 51.27% on the IEMOCAP dataset, alongside the classical BERT that obtained the highest zero-shot accuracy on the MELD dataset with 55.08%.

Large volumes of data are constantly being published on the web; however, the majority of this data is often unstructured, making it difficult to comprehend and interpret. To extract meaningful and structured information from such data, researchers and practitioners have turned to Information Extraction (IE) methods. One of the most challenging IE tasks is Event Extraction (EE), which involves extracting information related to specific incidents and their associated actors from text. EE has broad applications, including building a knowledge base, information retrieval, summarization, and online monitoring systems. Over the past few decades, various event ontologies, such as ACE, CAMEO, and ICEWS, have been developed to define event forms, actors, and dimensions of events observed in text. However, these ontologies have some limitations, such as covering only a few topics like political events, having inflexible structures in defining argument roles, lacking analytical dimensions, and insufficient gold-standard data. To address these concerns, we propose a new event ontology, COfEE, which integrates expert domain knowledge, previous ontologies, and a data-driven approach for identifying events from text. COfEE comprises two hierarchy levels (event types and event sub-types) that include new categories related to environmental issues, cyberspace, criminal activity, and natural disasters that require real-time monitoring. In addition, dynamic roles are defined for each event sub-type to capture various dimensions of events. The proposed ontology is evaluated on Wikipedia events, and it is shown to be comprehensive and general. Furthermore, to facilitate the preparation of gold-standard data for event extraction, we present a language-independent online tool based on COfEE. A gold-standard dataset annotated by ten human experts consisting of 24,000 news articles in Persian according to the COfEE ontology is also prepared. To diversify the data, news articles from the Wikipedia event portal and the 100 most popular Persian news agencies between 2008 and 2021 were collected. Finally, we introduce a supervised method based on deep learning techniques to automatically extract relevant events and their corresponding actors.

The aim of conversational speech processing is to analyze human conversations in natural settings. It finds numerous applications in personality traits identification, speech therapy, speaker identification and verification, speech emotion detection, and speaker diarization. However, large-scale annotated datasets required for feature extraction and conversational model training only exist for a handful of languages (e.g. English, Mandarin, and French) as the gathering, cleaning, and annotation of such datasets is tedious, time-consuming, and expensive. We propose two scalable, language-agnostic algorithms for automatically generating multi-speaker, variable-length, spontaneous conversations. These algorithms synthesize conversations using existing non-conversational speech datasets. We also contribute the resulting datasets (283 hours, 50 speakers). As a comparison, we also gathered the first spontaneous conversational dataset for Urdu (24 hours, 212 speakers) from public talk shows. Using speaker diarization as an example, we evaluate our datasets and report the first baseline diarization error rates (DER) for Urdu (25% for synthetic dataset-based models, and 29% for natural conversations). Our conversational speech generation technique allows training speaker diarization pipelines without the need for preparing huge conversational repositories.

Large Language Models (LLMs) have gained widespread popularity due to their instruction-following abilities. In this study, we evaluate their ability in simulating user interactions for task-oriented dialogue (TOD) systems. Our findings demonstrate that prompting LLMs reveals their promising capabilities for training and testing dialogue policies, eliminating the need for domain expertise in crafting complex rules or relying on large annotated data, as required by traditional simulators. The results show that the dialogue system trained with the ChatGPT simulator achieves a success rate of 59%, comparable to a 62% success rate of the dialogue system trained with the manual-rules, agenda-based user simulator (ABUS). Furthermore, the dialogue system trained with the ChatGPT simulator demonstrates better generalization ability compared to the dialogue system trained with the ABUS. Its success rate outperforms that of the dialogue system trained with the ABUS by 4% on GenTUS, 5% on the ChatGPT Simulator, and 3% on the Llama simulator. Nevertheless, LLM-based user simulators provide challenging environment, lexically rich, diverse, and random responses. Llama simulator outperforms the human reference in all lexical diversity metrics with a margin of 0.66 in SE, 0.39 in CE, 0.01 in MSTTR, 0.04 in HDD, and 0.55 in MTLD, while the ChatGPT simulator achieves comparable results. This ultimately contributes to enhancing the system’s ability to generalize more effectively.

Evaluating students' textual response is a common and critical task in language research and education practice. However, manual assessment can be tedious and may lack consistency, posing challenges for both scientific discovery and frontline teaching. Leveraging state-of-the-art large language models (LLMs), we aim to define and operationalize LLM-Surprisal, a numeric representation of the interplay between lexical diversity and syntactic complexity, and to empirically and theoretically demonstrate its relevance for automatic writing assessment and Chinese L2 (second language) learners’ English writing development. We developed an LLM-based natural language processing pipeline that can automatically compute text Surprisal scores. By comparing Surprisal metrics with the widely used classic indices in L2 studies, we extended the usage of computational metrics in Chinese learners’ L2 English writing. Our analyses suggested that LLM-Surprisals can distinguish L2 from L1 (first language) writing, index L2 development stages, and predict scores provided by human professionals. This indicated that the Surprisal dimension may manifest itself as critical aspects in L2 development. The relative advantages and disadvantages of these approaches were discussed in depth. We concluded that LLMs are promising tools that can enhance L2 research. Our showcase paves the way for more nuanced approaches to computationally assessing and understanding L2 development. Our pipelines and findings will inspire language teachers, learners, and researchers to operationalize LLMs in an innovative and accessible manner.

The prevalence of conversational interfaces is rapidly rising, since improved algorithms allow for remarkable proficiency in understanding and generating natural language. This also holds for Conversational Recommender Systems (CRS), that benefit from information being provided by the user in the course of the dialogue to offer personalized recommendations. However, the challenge remains eliciting the user's characteristics and preferences in a way that leads to the most optimal user experience. Hence, the current research was aimed at investigating the effect of different Preference Elicitation (PE) methods on the user experience of a CRS. We introduce two axes across which PE methods can be classified, namely the degree of system prompt guidance and the level of user input restriction. We built three versions of a CRS to conduct a between-subjects experiment which compared three conditions: high guidance-high restriction, high guidance-low restriction and low guidance-low restriction. We tested their effect on ten constructs of user experience measures on 66 European participants, all working in agriculture or forestry.

The study did not find any significant effects of the three preference elicitation methods on all user experience constructs collected through questionnaires. However, we did find significant differences in terms of the objective measures chat duration (Speed), response time (Cognitive Demand) and recommendation performance (Accuracy of Recommended Items). Regarding the recommendation performance, it was found that the preference elicitation methods with high guidance led to a higher match score than the condition with low guidance. The certainty score was highest in the condition with high guidance and high input restriction. Finally, we found through a question at the end of the conversation that users who were satisfied with the recommendation responded more positively to six out of ten user experience constructs. This suggests that satisfaction with the recommendation performance is a crucial factor in the user experience of CRSs.

Theory of mind (ToM), understanding others’ mental states, is a defining skill belonging to humans. Research assessing LLMs’ ToM performance yields conflicting findings and leads to discussions about whether and how they could show ToM understanding. Psychological research indicates that the characteristics of a specific language can influence how mental states are represented and communicated. Thus, it is reasonable to expect language characteristics to influence how LLMs communicate with humans, especially when the conversation involves references to mental states. This study examines how these characteristics affect LLMs’ ToM performance by evaluating GPT 3.5 and 4 performances in English and Turkish. Turkish provides an excellent contrast to English since Turkish has a different syntactic structure and special verbs, san- and zannet-, meaning “falsely believe.” Using Open AI's Chat Completion API, we collected responses from GPT models for first- and second-order ToM scenarios in English and Turkish. Our innovative approach combined completion prompts and open-ended questions within the same chat session, offering deep insights into models’ reasoning processes. Our data showed that while GPT models can respond accurately to standard ToM tasks (100% accuracy), their performance deteriorates (below chance level) with slight modifications. This high sensitivity suggests a lack of robustness in ToM performance. GPT 4 outperformed its predecessor, GPT 3.5, showing improvement in ToM performance to some extent. The models generally performed better when tasks were presented in English than in Turkish. These findings indicate that GPT models cannot reliably pass first-order and second-order ToM tasks in either of the languages yet. The findings have significant implications for Explainability of LLMs by highlighting challenges and biases that they face when simulating human-like ToM understanding in different languages.

Video understanding technology has become increasingly important in various disciplines, yet current approaches have primarily focused on lower comprehension level of video content, posing challenges for providing comprehensive and professional insights at a higher comprehension level. Video analysis plays a crucial role in athlete training and strategy development in racket sports. This study aims to demonstrate an innovative and higher-level video comprehension framework (ChatMatch), which integrates computer vision technologies with the cutting-edge large language models (LLM) to enable intelligent analysis and inference of racket sports videos. To examine the feasibility of this framework, we deployed a prototype of ChatMatch in the badminton in this study. A vision-based encoder was first proposed to extract the meta-features included the locations, actions, gestures, and action results of players in each frame of racket match videos, followed by a rule-based decoding method to transform the extracted information in both structured knowledge and unstructured knowledge. A set of LLM-based agents included namely task identifier, coach agent, statistician agent, and video manager, was developed through a prompt engineering and driven by an automated mechanism. The automatic collaborative interaction among the agents enabled the provision of a comprehensive response to professional inquiries from users. The validation findings showed that our vision models had excellent performances in meta-feature extraction, achieving a location identification accuracy of 0.991, an action recognition accuracy of 0.902, and a gesture recognition accuracy of 0.950. Additionally, a total of 100 questions were gathered from four proficient badminton players and one coach to evaluate the performance of the LLM-based agents, and the outcomes obtained from ChatMatch exhibited commendable results across general inquiries, statistical queries, and video retrieval tasks. These findings highlight the potential of using this approach that can offer valuable insights for athletes and coaches while significantly improve the efficiency of sports video analysis.