ACM Transactions on Asian and Low-Resource Language Information Processing最新文献

Binarization of Tamizhi (Tamil-Brahmi) inscription images are highly challenging as it is captured from very old stone inscriptions that exists around 3rd century BCE in India. The difficulty is due to the degradation of these inscriptions by environmental factors and human negligence over ages. Though many works have been carried out in the binarization of inscription images, very few research was performed for inscription images and no work has been reported for binarization of inscriptions inscribed on irregular medium. The findings of the analysis hold true to all writings that are carved in irregular background. This paper reviews the performance of various binarization techniques on Tamizhi inscription images. Since no previous work was performed, we have applied the existing binarization algorithms on Tamizhi inscription images and analyzed the performance of these algorithms with proper reasoning. In future, we believe that this reasoning on the results will help a new researcher, to adapt or combine or devise new binarization techniques.

Automatic text summarization (ATS) provides a summary of distinct categories of information using natural language processing (NLP). Low-resource languages like Hindi have restricted applications of these techniques. This study proposes a method for automatically generating summaries of Hindi documents using extractive technique. The approach retrieves pertinent sentences from the source documents by employing multiple linguistic features and machine learning (ML) using maximum likelihood estimation (MLE) and maximum entropy (ME). We conducted pre-processing on the input documents, such as eliminating Hindi stop words and stemming. We have obtained 15 linguistic feature scores from each document to identify the phrases with high scores for summary generation. We have performed experiments over BBC News articles, CNN News, DUC 2004, Hindi Text Short Summarization Corpus, Indian Language News Text Summarization Corpus, and Wikipedia Articles for the proposed text summarizer. The Hindi Text Short Summarization Corpus and Indian Language News Text Summarization Corpus datasets are in Hindi, whereas BBC News articles, CNN News, and the DUC 2004 datasets have been translated into Hindi using Google, Microsoft Bing, and Systran translators for experiments. The summarization results have been calculated and shown for Hindi as well as for English to compare the performance of a low and rich-resource language. Multiple ROUGE metrics, along with precision, recall, and F-measure, have been used for the evaluation, which shows the better performance of the proposed method with multiple ROUGE scores. We compare the proposed method with the supervised and unsupervised machine learning methodologies, including support vector machine (SVM), Naive Bayes (NB), decision tree (DT), latent semantic analysis (LSA), latent Dirichlet allocation (LDA), and K-means clustering, and it was found that the proposed method outperforms these methods.

Current Sundanese stemmers either ignore reduplication words or define rules to handle only affixes. There is a significant amount of reduplication words in the Sundanese language. Because of that, it is impossible to achieve superior stemming precision in the Sundanese language without addressing reduplication words. This paper presents an improved stemmer for the Sundanese language, which handles affixed and reduplicated words. With a Sundanese root word list, we use a rules-based stemming technique. In our approach, all stems produced by the affixes removal or normalization processes are added to the stem list. Using a stem list can help increase stemmer accuracy by reducing stemming errors caused by affix removal sequence errors or morphological issues. The current Sundanese language stemmer, RBSS, was used as a comparison. Two datasets with 8218 unique affixed words and reduplication words were evaluated. The results show that our stemmer's strength and accuracy have improved noticeably. The use of stem list and word reduplication rules improved our stemmer's affixed type recognition and allowed us to achieve up to 99.30% accuracy.

Utterance rewriting aims to identify and supply the omitted information in human conversation, which further enables the downstream task to understand conversations more comprehensively. Recently, sequence edit methods, which leverage the overlap between two sentences, have been widely applied to narrow the search space confronted by the previous linear generation methods. However, these methods ignore the relationship between linguistic elements in the conversation, which reflects how the knowledge and thoughts are organized in human communication. In this case, although most of the content in rewritten sentences can be found in the context, we found that some connecting words expressing relationships are often missing, which results in the out-of-context problem for the previous sentence edit method. To that end, in this paper, we propose a new semantic Graph-based Incomplete Utterance Rewriting (Graph4IUR) framework, which takes the semantic graph to depict the relationship between linguistic elements and captures out-of-context words. Specifically, we adopt the Abstract Meaning Representation (AMR) [4] graph as the basic sentence-to-graph method to depict the dialogue from the graph perspective, which could well represent the high-level semantics relationships of sentences. Along this line, we further adapt the sentence editing models to rewrite without changing the sentence architecture, which brings a restriction to exploring the overlap part of the current and rewritten sentences in the IUR task. Extensive experimental results indicate that our Graph4IUR framework can effectively alleviate the out-of-context problem and improve the performance of the previous edit-based methods in the IUR task.

Over the years, social media has emerged as one of the most popular platforms where people express their views and share thoughts about various aspects. The social media content now includes a variety of components such as text, images, videos etc. One type of interest is memes, which often combine text and images. It is relevant to mention here that, social media being an unregulated platform, sometimes also has instances of discriminatory, offensive and hateful content being posted. Such content adversely affects the online well-being of the users. Therefore, it is very important to develop computational models to automatically detect such content so that appropriate corrective action can be taken. Accordingly, there have been research efforts on automatic detection of such content focused mainly on the texts. However, the fusion of multimodal data (as in memes) creates various challenges in developing computational models that can handle such data, more so in the case of low-resource languages. Among such challenges, the lack of suitable datasets for developing computational models for handling memes in low-resource languages is a major problem. This work attempts to bridge the research gap by providing a large-sized curated dataset comprising 5,054 memes in Hindi-English code-mixed language, which are manually annotated by three independent annotators. It comprises two subtasks: (i) Subtask-1 (Binary classification involving tagging a meme as misogynous or non-misogynous), and (ii) Subtask-2 (multi-label classification of memes into different categories). The data quality is evaluated by computing Krippendorff's alpha. Different computational models are then applied on the data in three settings: text-only, image-only, and multimodal models using fusion techniques. The results show that the proposed multimodal method using the fusion technique may be the preferred choice for the identification of misogyny in multimodal Internet content and that the dataset is suitable for advancing research and development in the area.

The goal of emotion detection is to find and recognise emotions in text, speech, gestures, facial expressions, and more. This paper proposes an effective multimodal emotion recognition system based on facial expressions, sentence-level text, and voice. Using public datasets, we examine face expression image classification and feature extraction. The Tri-modal fusion is used to integrate the findings and to provide the final emotion. The proposed method has been verified in classroom students, and the feelings correlate with their performance. This method categorizes students' expressions into seven emotions: happy, surprise, sad, fear, disgust, anger, and contempt. Compared to the unimodal models, the suggested multimodal network design may reach up to 65% accuracy. The proposed method can detect negative feelings such as boredom or loss of interest in the learning environment.

Scholars in the humanities heavily rely on ancient manuscripts to study history, religion, and socio-political structures of the past. Significant efforts have been devoted to digitizing these precious manuscripts using OCR technology. However, most manuscripts have been blemished over the centuries, making it unrealistic for OCR programs to accurately capture faded characters. This work presents the Transformer + Confidence Score mechanism architecture for post-processing Google’s Tibetan OCR-ed outputs. According to the Loss and Character Error Rate metrics, our Transformer + Confidence Score mechanism architecture proves superior to the Transformer, LSTM-to-LSTM, and GRU-to-GRU architectures. Our method can be adapted to any language dealing with post-processing OCR outputs.