Finite State Transducers Research Articles

Subword Dictionary Learning and Segmentation for Expanding the Vocabulary of Automatic Speech Recognition in Tamil and Kannada

We present automatic speech recognition (ASR) systems for Tamil and Kannada based on subword modeling to effectively handle unlimited vocabulary arising due to the highly agglutinative nature of the languages. We propose a variant of the byte pair encoding (BPE) algorithm named extended-BPE, and Morfessor tool to segment each word into its subwords. We have effectively incorporated maximum likelihood and Viterbi estimation techniques with weighted finite state transducers framework in these algorithms to learn the subword dictionary from a large text corpus. Using the learned subword dictionary, the words in the transcriptions of the training data are segmented into subwords. We train deep neural network ASR systems that recognize the subword sequence for any given test speech utterance. The output subword sequence is then post-processed using deterministic rules to get the final word sequence. Because of this subword design, the number of words that can be recognized is much larger than the number of words in the training corpus. For Tamil ASR, we use 152 hours of data for training and 65 hours for testing, whereas for Kannada ASR, we use 275 hours for training and 72 hours for testing. Upon experimenting with different combinations of segmentation and estimation techniques, we find that the word error rate (WER) reduces drastically when compared to the baseline word-level ASR, achieving a maximum absolute WER reduction of 6.24% and 6.63% for Tamil and Kannada, respectively. Further, comparing our results with that of an end-to-end ASR model available on Github, we have seen that our subword language models can perform comparable to or exceed that of recent end-to-end ASR models for Kannada and Tamil languages.

Translation from Tunisian Dialect to Modern Standard Arabic: Exploring Finite-State Transducers and Sequence-to-Sequence Transformer Approaches

Translation from the mother tongue, including the Tunisian dialect, to modern standard Arabic is a highly significant field in natural language processing due to its wide range of applications and associated benefits. Recently, researchers have shown increased interest in the Tunisian dialect, primarily driven by the massive volume of content generated spontaneously by Tunisians on social media following the revolution. This article presents two distinct translators for converting the Tunisian dialect into Modern Standard Arabic. The first translator utilizes a rule-based approach, employing a collection of finite state transducers and a bilingual dictionary derived from the study corpus. On the other hand, the second translator relies on deep learning models, specifically the sequence-to-sequence transformer model and a parallel corpus. To assess, evaluate, and compare the performance of the two translators, we conducted tests using a parallel corpus comprising 8,599 words. The results achieved by both translators are noteworthy. The translator based on finite state transducers achieved a BLEU score of 56.65, while the transformer model-based translator achieved a higher score of 66.07.

Bengali reduplication generation with finite-state transducers (FSTs)

Bengali reduplication generation with finite-state transducers (FSTs)

Review on Sanskrit Sandhi Splitting using Deep Learning Techniques

A procedure called sandhi is used in Sanskrit to join short words (morphemes) to create compound words. A composite word are broken down into their component morphemes by a process known as sandhi splitting. This study focuses on several performance technologies and methodologies used to perform the above operation on Sanskrit sentences. Various approaches were identified for the problem from the literature survey. Initial approaches involved use of Finite State Transducers. Earlier the approaches introduced to increase accuracy include use of mathematical models and various optimality theories. Graph based approaches and parser based techniques were introduced later. With the advancement of deep learning techniques Recurrent Neural Networks, Long-Short Term Memory models and Double decoder models were adopted which involved training machine learning models through neural networks and classifier algorithms. Bidirectional LSTM models with attention mechanism, transformer based models and large language models like BERT were the most recent methodologies adopted and proved to be of higher accuracy and performance.

Pebble-depth

In this paper we introduce a new feasible notion of Bennett's logical depth based on pebble transducers. This notion is defined based on the difference between the minimal length descriptional complexity of prefixes of infinite sequences from the perspective of finite-state transducers and pebble transducers. Our notion of pebble-depth satisfies the four fundamental properties of depth: i.e. deep sequences exist, trivial sequences are not deep, random sequences are not deep, and the existence of a slow growth law type result. We also compare pebble-depth to other depth notions based on finite-state transducers, pushdown compressors, and the Lempel-Ziv 78 compression algorithm. We first demonstrate how there exists a normal pebble-deep sequence even though there is no normal finite-state-deep sequence. We next build a sequence that has a pebble-depth level of roughly 1, a pushdown-depth level of roughly 1/2 and a finite-state-depth level of roughly 0. We then build a sequence that has a pebble-depth level of roughly 1/2 and a Lempel-Ziv-depth level of roughly 0.

Current-State Opacity Based on State Outputs

Cyber-physical systems (CPS) are formed of physical and computational components that interact to operate safely and efficiently according to a desired behavior. In this paper, we consider that the network used to transmit data from the plant to its monitoring and control device is vulnerable to cyber-attacks. The attacker eavesdrops on the communication channel to infer that the system is in a secret state. We assume that the information transmitted in the network is the status of system components, such as sensors, actuators, and local controllers’ memory variables. Therefore, the system can be modeled from the attacker's perspective as a finite state transducer, where each state of the model is associated with an output vector formed of signals associated with system components. Considering that the attacker has full knowledge of the system model and does not know the system's current state when they initiate to eavesdrop on the communication channel, the attacker tries to estimate the system's current state by observing the transmitted state outputs. Thus, in this case, it is necessary to verify if the attacker is never able to detect whenever the system is in a secret state. In the Discrete-Event System literature, the system property associated with the intruder's inability to discover the system secret based on the transmitted information is called opacity. Since in this paper, the attacker observes state outputs of a finite state transducer, a new definition of opacity, called current-state opacity based on state outputs (CSO-SO), is introduced. We also present a method for the verification of this property.

Semi-Automatic Building and Learning of a Multilingual Ontology

Most online platforms, applications, and Websites use a massive amount of heterogeneous evolving data. These data must be structured and normalized before integration to improve the search and increase the relevance of results. An ontology can address this critical task by efficiently managing data and providing structured formats through techniques such as the Web Ontology Language (OWL). However, building an ontology can be costly, primarily if conducted manually. In this context, we propose a new methodology for automatically building and learning a multilingual ontology using Arabic as the base language via a corpus collected from Wikipedia. Our proposed methodology relies on Finite-state transducers (FSTs). FSTs are regrouped into a cascade to reduce errors and minimize ambiguity. The produced ontology is extended to English and French and independent language images via a translator we developed using APIs. The rationale for starting with the Arabic corpus to extract terms is that entity linking is more convenient from Arabic to other languages. In addition, many Wikipedia articles in English and French (for instance) do not have associated Arabic articles, but the opposite is true. In addition, dealing with Arabic terms permits us to enrich the Arabic module of the free linguistic platform we use in dictionaries and graphs. To assess the efficiency of our proposed methodology, we conducted performance metrics. The reported results are encouraging and promising.

Design and Development of Morphological Analyzer for Tigrigna Verbs using Hybrid Approach

Morphological analyzer is the base for various high-level NLP applications such as information retrieval, spell checking, grammar checking, machine translation, speech recognition, POS tagging and automatic sentence construction. This paper is carefully designed for design and analysis of morphological analyzer Tigrigna verbs using hybrid of memory learning and rules based approaches. The experiment have conducted using Python 3 where TiMBL algorithms IB2 and TRIBL2, and Finite State Transducer rules are used. The performance of the system has been evaluated using 10 fold cross validation technique. Testing was conducted using optimized parameter settings for regular verbs and linguistic rules of the Tigrigna language allomorph and phonology for the irregular verbs. The accuracy of the memory based approach with optimized parameters of TiMBL algorithm IB2 and TRIBL2 was 93.24% and 92.31%, respectively. Finally, the hybrid approach had an actual performance of 95.6% using linguistic rules for handling irregular and copula verbs.

A Comparative Study on Selecting Acoustic Modeling Units for WFST-based Mongolian Speech Recognition

Traditional weighted finite-state transducer– (WFST) based Mongolian automatic speech recognition (ASR) systems use phonemes as pronunciation lexicon modeling units. However, Mongolian is an agglutinative, low-resource language, and building an ASR system based on the phoneme pronunciation lexicon remains a challenge for various reasons. First, the phoneme pronunciation lexicon manually constructed by Mongolian linguists is finite, which is usually used to build a grapheme-to-phoneme conversion (G2P) model to frequently expand new words. However, the data sparsity decreases the robustness of the G2P model and affects the performance of the final ASR system. Second, homophones and polysyllabic words are common in Mongolian, which has a certain impact on the construction of the Mongolian acoustic model. To address these problems, in this work, we first propose a grapheme-to-phoneme alignment model to obtain the mapping relationship between phonemes and subword units. Then, we construct an acoustic subword segmentation set to segment words directly instead of using the traditional G2P method to predict phoneme sequences to expand the pronunciation lexicon. Further, by analyzing the Mongolian encoding form, we also propose an acoustic subword modeling units construction method that removes control characters. Finally, we investigate various acoustic subword modeling units for pronunciation lexicon construction for the Mongolian ASR system. Experiments on a Mongolian dataset with 325 hours of training show that the pronunciation lexicon based on the acoustic subword modeling unit can effectively construct the WFST-based Mongolian ASR system. Further, removing the control characters when building the acoustic subword modeling unit can further improve the ASR system performance.

Iterated uniform finite-state transducers on unary languages

An iterated uniform finite-state transducer executes the same length-preserving transduction in iterative sweeps. The first sweep occurs on the input string, while any subsequent sweep works on the output of the previous one. All sweeps always start from the sole initial state. The device accepts upon halting in an accepting state at the end of a sweep.We consider devices with one-way sweep motion and two-way sweep motion, i.e., sweeps are either from left to right only, or strictly alternate from left to right and from right to left. In addition, devices may work deterministically or nondeterministically.We focus on iterated uniform finite-state transducers accepting unary languages, i.e., languages built over single-letter alphabets.We show that any unary regular language can be accepted by a deterministic iterated uniform finite-state transducer with at most max⁡{2⋅ϱ,p}+1 states, where ϱ and p are the greatest primes in the factorization of the, respectively, pre-periodic and periodic part of the language. Such a state cost cannot be improved by using two-way motion, and it turns out to greatly outperform in the worst case the state costs of equivalent classical models of finite-state automata.Next, we give a characterization of classes of unary languages accepted by non-constant sweep-bounded iterated uniform finite-state transducers in terms of time-bounded one-way cellular automata. This characterization enables both to exhibit interesting families of unary nonregular languages accepted by iterated uniform finite-state transducers, and to prove the undecidability of several questions related to iterated uniform finite-state transducers accepting unary languages with an amount of sweeps that is at least logarithmic.

Language Recovery in Discrete-Event Systems against Sensor Deception Attacks

Cyber-physical systems are characterized by the intrinsic combination of software and physical components that usually include (wired and wireless) communication devices, sensors, actuators, and control processing units. Some wireless devices communicate over insecure channels, rendering cyber-physical systems at risk of malicious attacks that might lead to catastrophic damage. This paper touches upon the problem of sensor deception attacks in supervisory control of discrete-event systems, where an attacker can insert, delete, or replace sensor readings to mislead the supervisor and induce system damage. We model potential attacks using nondeterministic finite-state transducers and then introduce a new defence strategy that utilizes insertion functions. Insertion functions are a type of monitoring interface that alters the system’s behaviour by adding extra observable events. Finally, we construct a nondeterministic finite-state transducer called a supervisor filter that recovers the original language generated by the plant by handling the altered language. The insertion function and the supervisor filter cooperate to control the system and confuse the intruder without confusing the supervisor.

Low Complexity, Low Probability Patterns and Consequences for Algorithmic Probability Applications

Developing new ways to estimate probabilities can be valuable for science, statistics, engineering, and other fields. By considering the information content of different output patterns, recent work invoking algorithmic information theory inspired arguments has shown that a priori probability predictions based on pattern complexities can be made in a broad class of input-output maps. These algorithmic probability predictions do not depend on a detailed knowledge of how output patterns were produced, or historical statistical data. Although quantitatively fairly accurate, a main weakness of these predictions is that they are given as an upper bound on the probability of a pattern, but many low complexity, low probability patterns occur, for which the upper bound has little predictive value. Here, we study this low complexity, low probability phenomenon by looking at example maps, namely a finite state transducer, natural time series data, RNA molecule structures, and polynomial curves. Some mechanisms causing low complexity, low probability behaviour are identified, and we argue this behaviour should be assumed as a default in the real-world algorithmic probability studies. Additionally, we examine some applications of algorithmic probability and discuss some implications of low complexity, low probability patterns for several research areas including simplicity in physics and biology, a priori probability predictions, Solomonoff induction and Occam’s razor, machine learning, and password guessing.

Pushdown and Lempel-Ziv depth

In previously published work (Jordon and Moser, 2020), notions of finite-state-depth and pushdown-depth were presented. These were based on finite-state transducers and information lossless pushdown compressors. Unfortunately, a complete separation between the two notions was not established. This paper introduces a new formulation of pushdown-depth based on restricting how fast a pushdown compressor's stack can grow. This allows us to do a full comparison by demonstrating the existence of sequences with high finite-state-depth and low pushdown-depth, and vice-versa. A new notion based on the Lempel-Ziv 78 algorithm is also presented. Its difference from finite-state-depth is shown by a Lempel-Ziv deep sequence that is not finite-state deep, and vice versa. Lempel-Ziv-depth's difference from pushdown-depth is shown by building sequences that have a pushdown-depth of roughly 1/2 but low Lempel-Ziv depth, and by a sequence with high Lempel-Ziv depth but low pushdown-depth. Properties of all three notions are also studied.

A Computational Analysis of Arabic Noun Morphology

This project looks at Arabic word generation from a computational angle. It focuses on the computational production and analysis of morphological Arabic nouns. The work begins with a stem-based descriptive analysis of Arabic noun morphology that fulfills both the computational formalization and the linguistic description. There includes a thorough discussion of both inflectional and derivational systems. The spelling of Arabic nouns is also covered, as well as morphotactics and morphophonemics. The work then offers a computer implementation of Arabic nouns built on a rule-based computational morphological methodology. The overall system is constructed using the NooJ toolkit, which supports both pushdown automata and finite-state automata (FSA) (PDA). Three elements make up the morphological generation and analysis system: a lexicon, morphotactics, and rules. The lexicon component catalogs lexical elements (indivisible words and affixes), the morphotactics component specifies ordering restrictions for morphemes, and the rules component converts lexical representations into surface representations and vice versa. Other rules, such as orthographic, morphophonemic, and morphological rules, are also stored as two-level rules. The core editable lexicon of lemmas used as input by the system is drawn from three sources: the Buckwalter Arabic morphological analyzer lexicon, the Arramooz machine-readable dictionary, and the Alghani Azzahir dictionary. A complete annotated vocabulary of inflected noun forms (combined into a single type of finite-state transducers (FSTs)) is the system's output. The lexicon that was developed is then put to use in morphological analysis. The study then offers the system's evaluation. Accuracy, precision, and recall are three widely used metrics to assess the system's performance. Two empirical experiments will be conducted as part of the evaluation task. The system analyzing Arabic words that have been discredited morphologically is evaluated in the first experiment. Accuracy, precision, and recall for the system when employing discredited Arabic words are (90.4%), (98.3%), and (88.9%), respectively. The technique is tested in a second experiment using undiacritical words. The achieved outcomes of this experiment were (94.7%) accuracy, (96.7%) precision, and (91.6% ) recall, respectively. Additionally, the measurement average for the two tests has been determined. The average performance values are respectively (92.55%), (97.5%), and (90.25) percent in terms of recall, precision, and accuracy. Overall, the results are encouraging and demonstrate the system's propensity for dealing with both diacritically and undiacritically written Arabic texts. This system can analyze Arabic text corpora in-depth and tag nouns according to their morphological characteristics. It breaks the word under analysis into three pieces (the stem, proclitics/prefixes, and suffixes/enclitics) and assigns each one a specific morphological feature tag or possibly many tags if the portion in question has numerous clitics or affixes. Many applications of natural language processing, including parsing, lemmatization, stemming, part-of-speech (POS) tagging, corpus annotations, word sense disambiguation, machine translation, information retrieval, text generation, spelling checkers, etc., depend on computational morphology. It is made up of morphological generation and analysis paradigms. According to a set of features, morphological generation attempts to construct every feasible derived and inflected form of a given lemma. On the other hand, morphological analysis is the process of dissecting a word into its component morphemes and giving each morpheme linguistic tags or qualities.

CKMorph: a comprehensive morphological analyzer for Central Kurdish

A morphological analyzer, a significant component of many natural language processing applications, especially for morphologically rich languages, divides an input word into all its composing morphemes and identifies their morphological roles. This paper introduces a comprehensive morphological analyzer for Central Kurdish (CK), also known as Sorani, a low-resourced language with rich morphology. Building upon the limited existing literature, we first assembled and systematically categorized an extensive collection of the morphological and morphophonological rules of the language. Additionally, we collected and manually labeled a generative lexicon containing nearly 10,000 verb, noun and adjective stems, named entities, and other types of word stems. We used these rule sets and resources to implement CKMorph Analyzer based on finite-state transducers. In order to provide a benchmark for future research, we collected, manually labeled, and publicly shared test sets for evaluating the accuracy and coverage of the analyzer. CKMorph was able to correctly analyze 95.9% of the first test set, containing 1000 CK words morphologically analyzed according to the context. Moreover, CKMorph gave at least one analysis for 95.5% of 4.22 M CK tokens of the second test set. The demonstration of the application and resources, including the CK verb database and test sets, are openly accessible at github.com/CKMorph.

Transfer and Triangulation Pivot Translation Approaches for Burmese Dialects

Parallel corpora for the languages of Myanmar (Beik, Burmese, Rakhine) are extremely scarce but a necessary requirement for machine translation R&D. Previous studies have proved that pivoting leads to better translation quality if the bridge language is closely related to the source and target language pair. The baseline study is conducted based on the three major approaches of machine translation; Weighted Finite State Transducer (WFST), Phrase-Based Statistical Machine Translation (PBSMT) and Deep Recurrent Neural Network (Deep-RNN). Based on the baseline results, this paper mainly investigated the pivot language technique for PBSMT with Burmese dialects. We employed two different pivot translation methods: transfer (sentence level) and triangulation (phrase level). We present the experimental results on Dawei-Beik, Beik-Dawei translations and Beik-Rakhine, Rakhine-Beik translation via Burmese. Both the transfer and triangulation approaches outperformed the baseline (direct translation), specifically for the Rakhine-Beik language pair. Moreover, the results of the average BiLingual Evaluation Understudy (BLEU), Character n-gram F-score (chrF), and Word Error Rate (WER) scores of the 10-fold cross-validation experiments proved that the triangulation pivot has significantly better acceleration than the transfer pivot. We plan to release the parallel corpora of Burmese dialects and present several avenues for further research.

Attack-Resilient Supervisory Control of Discrete-Event Systems: A Finite-State Transducer Approach

Attack-Resilient Supervisory Control of Discrete-Event Systems: A Finite-State Transducer Approach

A Fibonacci analogue of the two’s complement numeration system

Using the classic two’s complement notation of signed integers, the fundamental arithmetic operations of addition, subtraction, and multiplication are identical to those for unsigned binary numbers. We introduce a Fibonacci-equivalent of the two’s complement notation and we show that addition in this numeration system can be performed by a deterministic finite-state transducer. The result is based on the Berstel adder, which performs addition of the usual Fibonacci representations of nonnegative integers and for which we provide a new constructive proof. Moreover, we characterize the Fibonacci-equivalent of the two’s complement notation as an increasing bijection between ℤ and a particular language.

Variable Structure and Modeling Units for Chinese Lipreading

Lipreading is a type of Human–Computer Interaction (HCI) based on visual information. From a linguistic point of view, Chinese is a monosyllabic language with a much higher proportion of homophones than English. Identifying homophones in Chinese Mandarin lipreading is very challenging. Since the lip shape in the context can distinguish homophones, and smaller recognition units can reduce the types of recognition and alleviate data sparsity, we propose to improve the accuracy of lipreading by simultaneously exploiting the correlation of lip features at different distances and smaller modeling units. We implement a long short-term multi-feature space to represent lip features, and CTC–Attention to learn temporal correlations. We also introduce Weight Finite State Transducer (WFST) to enhance the semantic analysis capability of the model. Our model aims to distinguish homophones and improve the accuracy of lipreading. To reduce data sparsity, we use Tonal Initials and Finals (TIF) as the modeling units. We record a sentence-level Chinese lipreading dataset, ICSLR, and label Mandarin characters, syllables, and TIF. We demonstrate the effectiveness of the proposed approach compared to its counterparts through extensive experiments on Grid, ICSLR, and CMLR datasets.

Improving out of vocabulary words recognition accuracy for an end-to-end Russian speech recognition system

Automatic Speech Recognition (ASR) systems are experiencing an active introduction into our daily lives, simplifying the way we interact with electronic devices. The advent of end-to-end approaches has only accelerated this process. However, the constant evolution and a high degree of inflection of the Russian language lead to the problem of recognizing new words outside the vocabulary (Out Of Vocabulary, OOV) because they did not take part in the training process of the ASR system. In such a case, the ASR model tends to predict the most similar word from the training data which leads to a recognition error. This is especially true for ASR models that use decoding based on a Weighted Finite State Transducer (WFST), since they are obviously limited by the list of vocabulary words that may appear as a result of recognition. In this paper, this problem is investigated on the basis of an open data set of the Russian language (common voice) and an integrated ASR system using the WFST decoder. A method for retraining an integral ASR system based on the discriminative loss function MMI (maximum mutual information) and a method for decoding the integral model using a TG graph are proposed. Discriminative learning allows smoothing the probability distribution of acoustic class prediction, thus adding more variability in the recognition results. Decoding using the TG graph, in turn, is not limited to recognizing only vocabulary words and allows the use of a language model trained on a large amount of external text data. An eight-hour subset from the common voice base is used as a test set. The total number of OOV words in this test sample is 18.1 %. The results show that the use of the proposed methods allows to reduce the word recognition error (Word Error Rate, WER) by 3 % in absolute value relative to the standard method of decoding integral models (beam search), while maintaining the ability to recognize OOV words at a comparable level. The use of the proposed methods should improve the overall quality of recognition of ASR systems and make such systems more resistant to the recognition of new words that were not involved in the learning process.