Narrow Results

In this paper, we address the task of solving proportional analogies between sentences in a generative manner. To this end, we use pre-trained word or sentence embedding models or fine-tune several pre-trained language models. Experiments on analogies between short sentences show that fine-tuning GPT-2 achieves the best performance, while combining a word embedding model with its vector-to-sequence decoder delivers competitive accuracy. However, the increased sizes of large-scale language models might not be worth the slight increase in performance observed when compared with lightweight models to simple analogy tasks. To extend this, we develop a data pipeline for collecting semantico-formal analogies between long sentences, spotting parallel associations beyond individual words. We create a complex set of analogies, each validated with evidence of relational matches on underlying concepts. Comparative analyses of three language models demonstrate that the autoregressive framework is particularly effective at learning analogy structures when sequences are long.

We present an approach for the development of Language Understanding systems from a Transduction point of view. We describe the use of two types of automatically inferred transducers as the appropriate models for the understanding phase in dialog systems.

People across the world habitually turn to online social media to share their experiences, thoughts, ideas, and opinions as they go about their daily lives. These posts collectively contain a wealth of insights into how masses perceive their surroundings. Therefore, extracting people's perceptions from social media posts can provide valuable information about pertinent issues such as public transportation, emergency conditions, and even reactions to political actions or other activities. This paper proposes a novel approach to extract such perceptions from a corpus of social media posts originating from a given broad geographical region. The approach divides the broad region into a number of sub-regions, and trains language models over social media conversations within these sub-regions. Using Bayesian and geo-smoothing methods, the ensemble of language models can be queried with phrases embodying a perception. Discrete and continuous visualization methods represent the extent to which social media posts within the sub-regions express the query. The capabilities of the perception mining approach are illustrated using transportation-themed scenarios.

This paper presents the ideas, experiments and specifications related to the Supervised TextRank (STR) technique, a word tagging method based on the TextRank algorithm. The main innovation of STR technique is the use of a graph-based ranking algorithm similar to PageRank in a supervised fashion, gathering the information needed to build the graph representations of the text from a tagged corpus. We also propose a flexible graph specification language that allows to easily experiment with multiple configurations for the topology of the graph and for the information associated to the nodes and the edges. We have carried experiments in the Part-Of-Speech task, a common tagging problem in Natural Language Processing. In our best result we have achieved a precision of 96.16%, at the same level of the best tagging tools.

A vast amount of textual information on the internet has amplified the importance of text summarization models. Abstractive summarization generates original words and sentences that may not exist in the source document to be summarized. Such abstractive models may suffer from shortcomings such as linguistic acceptability and hallucinations. Recall-Oriented Understudy for Gisting Evaluation (ROUGE) is a metric commonly used to evaluate abstractive summarization models. However, due to its n-gram-based approach, it ignores several critical linguistic aspects. In this work, we propose Similarity, Entailment, and Acceptability Score (SEAScore), an automatic evaluation metric for evaluating abstractive text summarization models using the power of state-of-the-art pre-trained language models. SEAScore comprises three language models (LMs) that extract meaningful linguistic features from candidate and reference summaries and a weighted sum aggregator that computes an evaluation score. Experimental results show that our LM-based SEAScore metric correlates better with human judgment than standard evaluation metrics such as ROUGE-N and BERTScore.

Language Models constitute an effective framework for text retrieval tasks. Recently, it has been extended to various collaborative filtering tasks. In particular, relevance-based language models can be used for generating highly accurate recommendations using a memory-based approach. On the other hand, the query likelihood model has proven to be a successful strategy for neighbourhood computation. Since relevance-based language models rely on user neighbourhoods for producing recommendations, we propose to use the query likelihood model for computing those neighbourhoods instead of cosine similarity. The combination of both techniques results in a formal probabilistic recommender system which has not been used before in collaborative filtering. A thorough evaluation on three datasets shows that the query likelihood model provides better results than cosine similarity. To understand this improvement, we devise two properties that a good neighbourhood algorithm should satisfy. Our axiomatic analysis shows that the query likelihood model always enforces those constraints while cosine similarity does not.

Research interest in Chinese character recognition in Taiwan in recent years has been intense, due in part to cultural considerations, and in part to advances in computer hardware development. This chapter addresses coarse character classification, candidate selection, statistical character recognition, recognition based on structural character primitives such as line segments, strokes and radicals, as well as postprocessing and model development.

Coarse character classification and candidate selection are used to reduce matching complexity; statistical methods of character recognition are shown to be effective feature-matching which shows good performance is reported; and, structural-based methods able to distinguish between similar characters are investigated thoroughly. Since no temporal information is available for off-line recognition systems, the character test base is still limited. Methods used to extract structural primitives are also investigated.

Language models based on syntactical or semantic considerations are used to select the most probable characters from sets of candidates, and are applied in postprocessing in input sentence images. These models generally employ the dynamic programming methods. To increase identification capacity, various ways of grouping Chinese words into a reasonable number of classes are also proposed.

This article presents a review of some techniques that have been found to be useful for measuring the similarity between documents. After defining measures of similarity, the article describes some of the standard preprocessing steps that are useful. The binary search models and vector space models are then examined. The sources of shortcomings in these models are identified and the manner in which latent semantic analysis overcomes some of them is described. Further improvement, using the probabilistic semantic analysis, is then described. Avenues for further improvements, using language models are also suggested.