AI Chat Character Apps

AI Chat Character Apps — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

Large language model

A large language model (LLM) is a neural network trained on a vast amount of text for natural language processing tasks, especially language generation. LLMs can typically generate, summarize, translate and analyze text in many contexts, and are a foundational technology behind modern chatbots. Biased or inaccurate training data can make an LLM's output less reliable. As of 2026, the most capable LLMs are based on transformer architectures, which, according to the 2017 paper "Attention Is All You Need", can be more efficient and parallelizable than earlier statistical and recurrent neural network models. Benchmark evaluations for LLMs attempt to measure model reasoning, factual accuracy, alignment, and safety. == History == Before the emergence of transformer-based models in 2017, some language models were considered large relative to the computational and data constraints of their time. In the early 1990s, IBM's statistical models pioneered word alignment techniques for machine translation, laying the groundwork for corpus-based language modeling. In 2001, a smoothed n-gram model, such as those employing Kneser–Ney smoothing, trained on 300 million words, achieved state-of-the-art perplexity on benchmark tests. During the 2000s, with the rise of widespread internet access, researchers began compiling massive text datasets from the web ("web as corpus") to train statistical language models. Moving beyond n-gram models, researchers started in 2000 to use neural networks as language models. Following the breakthrough of deep neural networks in image classification around 2012, similar architectures were adapted for language tasks. This shift was marked by the development of word embeddings (e.g., Word2Vec by Mikolov in 2013) and sequence-to-sequence (seq2seq) models using LSTM. In 2016, Google transitioned its translation service to neural machine translation (NMT), replacing statistical phrase-based models with deep recurrent neural networks. These early NMT systems used LSTM-based encoder-decoder architectures, as they preceded the invention of transformers. At the 2017 NeurIPS conference, Google researchers introduced the transformer architecture in their landmark paper "Attention Is All You Need". This paper's goal was to improve upon 2014 seq2seq technology, and was based mainly on the attention mechanism developed by Bahdanau et al. in 2014. The following year in 2018, BERT was introduced and quickly became "ubiquitous". Though the original transformer has both encoder and decoder blocks, BERT is an encoder-only model. Academic and research usage of BERT began to decline in 2023, following rapid improvements in the abilities of decoder-only models (such as GPT) to solve tasks via prompting. Although decoder-only GPT-1 was introduced in 2018, it was GPT-2 in 2019 that caught widespread attention because OpenAI claimed to have initially deemed it too powerful to release publicly, out of fear of malicious use. GPT-3 in 2020 went a step further and as of 2025 is available only via API with no offering of downloading the model to execute locally. But it was the consumer-facing chatbot ChatGPT in late 2022 that received extensive media coverage and public attention by 2023. The 2023 GPT-4 was praised for its increased accuracy and as a "holy grail" for its multimodal capabilities. OpenAI did not reveal the high-level architecture and the number of parameters of GPT-4. The release of ChatGPT led to an uptick in LLM usage across several research subfields of computer science, including robotics, software engineering, and societal impact work. In 2024, OpenAI released the reasoning model OpenAI o1, which generates long chains of thought before returning a final answer. Many LLMs with parameter counts comparable to those of OpenAI's GPT series have been developed. Since 2022, weights-available models have been gaining popularity, especially at first with BLOOM and LLaMA, though both have restrictions on usage and deployment. Mistral AI's open-weight models Mistral 7B and Mixtral 8x7B have a more permissive Apache License. In January 2025, DeepSeek released DeepSeek R1, a 671-billion-parameter open-weight model that performs comparably to OpenAI o1 but at a much lower price per token for users. Since 2023, many LLMs have been trained to be multimodal, having the ability to also process or generate other types of data, such as images, audio, or 3D meshes. Open-weight LLMs have become more influential since 2023. Per Vake et al. (2025), community-driven contributions to open-weight models improve their efficiency and performance via collaborative platforms such as Hugging Face. == Dataset preprocessing == === Tokenization === As machine learning algorithms process numbers rather than text, the text must be converted to numbers. In the first step, a vocabulary is decided upon, then integer indices are arbitrarily but uniquely assigned to each vocabulary entry, and finally, an embedding is associated with the integer index. Algorithms include byte-pair encoding (BPE) and WordPiece. There are also special tokens serving as control characters, such as [MASK] for masked-out token (as used in BERT), and [UNK] ("unknown") for characters not appearing in the vocabulary. Also, some special symbols are used to denote special text formatting. For example, "Ġ" denotes a preceding whitespace in RoBERTa and GPT and "##" denotes continuation of a preceding word in BERT. For example, the BPE tokenizer used by the legacy version of GPT-3 would split tokenizer: texts -> series of numerical "tokens" as Tokenization also compresses the datasets. Because LLMs generally require input to be an array that is not jagged, the shorter texts must be "padded" until they match the length of the longest one. ==== Byte-pair encoding ==== As an example, consider a tokenizer based on byte-pair encoding. In the first step, all unique characters (including blanks and punctuation marks) are treated as an initial set of n-grams (i.e. initial set of uni-grams). Successively the most frequent pair of adjacent characters is merged into a bi-gram and all instances of the pair are replaced by it. All occurrences of adjacent pairs of (previously merged) n-grams that most frequently occur together are then again merged into even lengthier n-gram, until a vocabulary of prescribed size is obtained. After a tokenizer is trained, any text can be tokenized by it, as long as it does not contain characters not appearing in the initial-set of uni-grams. === Dataset cleaning === In the context of training LLMs, datasets are typically cleaned by removing low-quality, duplicated, or toxic data. Cleaned datasets can increase training efficiency and lead to improved downstream performance. A trained LLM can be used to clean datasets for training a further LLM. With the increasing proportion of LLM-generated content on the web, data cleaning in the future may include filtering out such content. LLM-generated content can pose a problem if the content is similar to human text (making filtering difficult) but of lower quality (degrading performance of models trained on it). === Synthetic data === Training of largest language models might need more linguistic data than naturally available, or that the naturally occurring data is of insufficient quality. In these cases, synthetic data might be used. == Training == An LLM is a type of foundation model (large X model) trained on language. LLMs can be trained in different ways. In particular, GPT models are first pretrained to predict the next word on a large amount of data, before being fine-tuned. === Cost === Substantial infrastructure is necessary for training the largest models. The tendency towards larger models is visible in the list of large language models. For example, the training of GPT-2 (i.e. a 1.5-billion-parameter model) in 2019 cost $50,000, while training of the PaLM (i.e. a 540-billion-parameter model) in 2022 cost $8 million, and Megatron-Turing NLG 530B (in 2021) cost around $11 million. The qualifier "large" in "large language model" is inherently vague, as there is no definitive threshold for the number of parameters required to qualify as "large". === Fine-tuning === Before being fine-tuned, most LLMs are next-token predictors. The fine-tuning shapes the LLM's behavior via techniques like reinforcement learning from human feedback (RLHF) or constitutional AI. Instruction fine-tuning is a form of supervised learning used to teach LLMs to follow user instructions. In 2022, OpenAI demonstrated InstructGPT, a version of GPT-3 similarly fine-tuned to follow instructions. Reinforcement learning from human feedback (RLHF) involves training a reward model to predict which text humans prefer. Then, the LLM can be fine-tuned through reinforcement learning to better satisfy this reward model. Since humans typically prefer truthful, helpful and harmless answers, RLHF favors such answers. == Architecture == LLMs are generally based on the tra
Read more →
Jun'ichi Tsujii

Jun'ichi Tsujii (辻井潤一, Tsujii Jun'ichi; born 7 February 1949) is a Japanese computer scientist specializing in natural language processing and text mining, particularly in the field of biology and bioinformatics. == Education == Tsujii received his Bachelor of Engineering, Master of Engineering and PhD degrees in electrical engineering from Kyoto University in 1971, 1973, and 1978 respectively. He was Assistant Professor and Associate Professor at Kyoto University, before accepting a position as Professor of Computational Linguistics at the University of Manchester Institute of Science and Technology (UMIST) in 1988. He was President of the Association for Computational Linguistics (ACL) in 2006, and has been a permanent member of the International Committee on Computational Linguistics (ICCL) since 1992, and the chair of the committee since 2014. == Research == Since May 2015, Tsujii has been the director of the Artificial Intelligence Research Center at the National Institute of Advanced Industrial Science and Technology, Japan. Tsujii was previously a Principal Researcher at Microsoft Research Asia (MSRA). Before joining MSRA, he was a professor at the University of Tokyo, where he belonged to both the School of Inter-faculty Initiative on Informatics and the Graduate School of Information Science and Technology. Tsujii is also a Visiting Professor and Scientific Advisor at the National Centre for Text Mining (NaCTeM) at the University of Manchester in the United Kingdom. == Awards == On 14 May 2010, Tsujii was awarded the Medals of Honor with Purple Ribbon, one of Japan's highest awards, presented to influential contributors in the fields of art, academics or sports. In September 2014, Tsujii was awarded the FUNAI Achievement Award at the Forum on Information Technology (FIT), which took place at the University of Tsukuba. The award is presented to distinguished individuals engaged in research or related business activities in the field of Information Technology who have produced excellent achievements in the field, are still active in leading positions and have strong impact on young students and researchers. In December 2014, Tsujii was named as an ACL Fellow, in recognition of his significant contributions to MT, parsing by unification-based grammar and text mining for biology. In March 2016, Tsujii was awarded Okawa Prize for his contribution to the field of Natural Language Processing, Machine Translation and Text Mining, together with Professor Jaime Carbonnel of CMU. In August 2021, Tsujii received ACL Lifetime Achievement Award, which is considered the most prestigious award in the field of Computational Linguistics and Natural Language Processing. In May 2022, Tsujii received the Order of the Sacred Treasure, Gold Rays and Neck Ribbon, from the Japanese government. In October 2024, Tsujii was designated a Person of Cultural Merit. == Selected publications == Oiwa, Hidekazu; Tsujii, Jun'ichi (2014). Common Space Embedding of Primal-Dual Relation Semantic Spaces. COLING 2014. Dublin. pp. 1579–1590. Taura, K.; Matsuzaki, T.; Miwa, M.; Kamoshida, Y.; Yokoyama, D.; Dun, N.; Shibata, T.; Jun, C. S.; Tsujii, J. (2013). "Design and implementation of GXP make – A workflow system based on make". Future Generation Computer Systems. 29 (2): 662–672. doi:10.1016/j.future.2011.05.026. S2CID 31627886. Sun, X.; Zhang, Y.; Matsuzaki, T.; Tsuruoka, Y.; Tsujii, J. (2013). "Probabilistic Chinese word segmentation with non-local information and stochastic training". Information Processing & Management. 49 (3): 626–636. doi:10.1016/j.ipm.2012.12.003. Mu, T.; Goulermas, J. Y.; Tsujii, J.; Ananiadou, S. (2012). "Proximity-Based Frameworks for Generating Embeddings from Multi-Output Data". IEEE Transactions on Pattern Analysis and Machine Intelligence. 34 (11): 2216–2232. Bibcode:2012ITPAM..34.2216M. doi:10.1109/TPAMI.2012.20. PMID 23289130. S2CID 711467. Miwa, M.; Sætre, R.; Kim, J. D.; Tsujii, J. (2010). "Event Extraction with Complex Event Classification Using Rich Features". Journal of Bioinformatics and Computational Biology. 08 (1): 131–146. doi:10.1142/S0219720010004586. PMID 20183879. Kim, J. D.; Ohta, T.; Tsujii, J. (2008). "Corpus annotation for mining biomedical events from literature". BMC Bioinformatics. 9 10. doi:10.1186/1471-2105-9-10. PMC 2267702. PMID 18182099. Miyao, Y.; Tsujii, J. (2008). "Feature Forest Models for Probabilistic HPSG Parsing". Computational Linguistics. 34: 35–80. doi:10.1162/coli.2008.34.1.35. S2CID 885002. Sagae, Kenji; Tsujii, Jun'ichi (2007). Dependency Parsing and Domain Adaptation with LR Models and Parser Ensembles. EMNLP-CoNLL. pp. 1044–1050. Ananiadou, S; Pyysalo, S; Tsujii, J; Kell, D. B. (2010). "Event extraction for systems biology by text mining the literature". Trends in Biotechnology. 28 (7): 381–90. doi:10.1016/j.tibtech.2010.04.005. PMID 20570001. Tsuruoka, Y.; Tateishi, Y.; Kim, J. D.; Ohta, T.; McNaught, J.; Ananiadou, S.; Tsujii, J. (2005). "Developing a Robust Part-of-Speech Tagger for Biomedical Text". Advances in Informatics. Lecture Notes in Computer Science. Vol. 3746. p. 382. doi:10.1007/11573036_36. ISBN 978-3-540-29673-7. S2CID 206592413. Tsuruoka, Y.; Tsujii, J. (2005). Bidirectional inference with the easiest-first strategy for tagging sequence data. Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing - HLT '05. pp. 467–474. doi:10.3115/1220575.1220634. Tsujii, J.; Ananiadou, S. (2005). "Thesaurus or Logical Ontology, Which One Do We Need for Text Mining?". Language Resources and Evaluation. 39: 77–90. doi:10.1007/s10579-005-2697-0. S2CID 3204827. Kazama, J. I.; Tsujii, J. I. (2005). "Maximum Entropy Models with Inequality Constraints: A Case Study on Text Categorization". Machine Learning. 60 (1–3): 159–194. doi:10.1007/s10994-005-0911-3. hdl:10119/3305. Matsuzaki, T.; Miyao, Y.; Tsujii, J. I. (2005). Probabilistic CFG with latent annotations. 43rd Annual Meeting on Association for Computational Linguistics - ACL '05. p. 75. doi:10.3115/1219840.1219850. Kim, J. -D.; Ohta, T.; Tateisi, Y.; Tsujii, J. (2003). "GENIA corpus--a semantically annotated corpus for bio-textmining". Bioinformatics. 19: i180–i182. doi:10.1093/bioinformatics/btg1023. PMID 12855455. Hirschman, L.; Park, J. C.; Tsujii, J.; Wong, L.; Wu, C. H. (2002). "Accomplishments and challenges in literature data mining for biology". Bioinformatics. 18 (12): 1553–1561. doi:10.1093/bioinformatics/18.12.1553. PMID 12490438. Torisawa, K.; Tsujii, J. I. (1996). Computing phrasal-signs in HPSG prior to parsing. 16th conference on Computational linguistics -. Vol. 2. p. 949. doi:10.3115/993268.993332.
Read more →
Quantum finite automaton

In quantum computing, quantum finite automata (QFA) or quantum state machines are a quantum analog of probabilistic automata or a Markov decision process. They provide a mathematical abstraction of real-world quantum computers. Several types of automata may be defined, including measure-once and measure-many automata. Quantum finite automata can also be understood as the quantization of subshifts of finite type, or as a quantization of Markov chains. QFAs are, in turn, special cases of geometric finite automata or topological finite automata. The automata work by receiving a finite-length string σ = ( σ 0 , σ 1 , … , σ k ) {\displaystyle \sigma =(\sigma _{0},\sigma _{1},\dots ,\sigma _{k})} of letters σ i {\displaystyle \sigma _{i}} from a finite alphabet Σ {\displaystyle \Sigma } , and assigning to each such string a probability Pr ⁡ ( σ ) {\displaystyle \operatorname {Pr} (\sigma )} indicating the probability of the automaton being in an accept state; that is, indicating whether the automaton accepted or rejected the string. The languages accepted by QFAs are not the regular languages of deterministic finite automata, nor are they the stochastic languages of probabilistic finite automata. Study of these quantum languages remains an active area of research. == Informal description == There is a simple, intuitive way of understanding quantum finite automata. One begins with a graph-theoretic interpretation of deterministic finite automata (DFA). A DFA can be represented as a labelled directed graph, with states as nodes in the graph, and arrows representing state transitions. Each arrow is labelled with a possible input symbol, so that, given a specific state and an input symbol, the arrow points at the next state. One way of representing such a graph is by means of a set of adjacency matrices, with one matrix for each input symbol. In this case, a list of possible DFA states is written as a column vector. For a given input symbol, the adjacency matrix indicates how any given state (row in the state vector) will transition to the next state; a state transition is given by matrix multiplication. One needs a distinct adjacency matrix for each possible input symbol, since each input symbol can result in a different transition. The entries in the adjacency matrix must be zero's and one's. For any given column in the matrix, only one entry can be non-zero: this is the entry that indicates the next (unique) state transition. Similarly, the state of the system is a column vector, in which only one entry is non-zero: this entry corresponds to the current state of the system. Let Σ {\displaystyle \Sigma } denote the set of input symbols. For a given input symbol α ∈ Σ {\displaystyle \alpha \in \Sigma } , write U α {\displaystyle U_{\alpha }} as the adjacency matrix that describes the evolution of the DFA to its next state. The set { U α | α ∈ Σ } {\displaystyle \{U_{\alpha }|\alpha \in \Sigma \}} then completely describes the state transition function of the DFA. Let Q represent the set of possible states of the DFA. If there are N states in Q, then each matrix U α {\displaystyle U_{\alpha }} is N by N-dimensional. The initial state q 0 ∈ Q {\displaystyle q_{0}\in Q} corresponds to a column vector with a one in the q0'th row. A general state q is then a column vector with a one in the q'th row. By abuse of notation, let q0 and q also denote these two vectors. Then, after reading input symbols α β γ ⋯ {\displaystyle \alpha \beta \gamma \cdots } from the input tape, the state of the DFA will be given by q = ⋯ U γ U β U α q 0 . {\displaystyle q=\cdots U_{\gamma }U_{\beta }U_{\alpha }q_{0}.} The state transitions are given by ordinary matrix multiplication (that is, multiply q0 by U α {\displaystyle U_{\alpha }} , etc.); the order of application is 'reversed' only because we follow the standard notation of linear algebra. The above description of a DFA, in terms of linear operators and vectors, almost begs for generalization, by replacing the state-vector q by some general vector, and the matrices { U α } {\displaystyle \{U_{\alpha }\}} by some general operators. This is essentially what a QFA does: it replaces q by a unit vector, and the { U α } {\displaystyle \{U_{\alpha }\}} by unitary matrices. Other, similar generalizations also become obvious: the vector q can be some distribution on a manifold; the set of transition matrices become automorphisms of the manifold; this defines a topological finite automaton. Similarly, the matrices could be taken as automorphisms of a homogeneous space; this defines a geometric finite automaton. Before moving on to the formal description of a QFA, there are two noteworthy generalizations that should be mentioned and understood. The first is the non-deterministic finite automaton (NFA). In this case, the vector q is replaced by a vector that can have more than one entry that is non-zero. Such a vector then represents an element of the power set of Q; it’s just an indicator function on Q. Likewise, the state transition matrices { U α } {\displaystyle \{U_{\alpha }\}} are defined in such a way that a given column can have several non-zero entries in it. Equivalently, the multiply-add operations performed during component-wise matrix multiplication should be replaced by Boolean and-or operations so that the semantics are kept intact. A well-known theorem states that, for each DFA, there is an equivalent NFA, and vice versa. This implies that the set of languages that can be recognized by DFA's and NFA's are the same; these are the regular languages. In the generalization to QFAs, the set of recognized languages will be different to the regular languages. Describing that set is one of the outstanding research problems in QFA theory. Another generalization that should be immediately apparent is to use a stochastic matrix for the transition matrices, and a probability vector for the state; this gives a probabilistic finite automaton. The entries in the state vector must be real numbers, positive, and sum to one, in order for the state vector to be interpreted as a probability. The transition matrices must preserve this property: this is why they must be stochastic. Each state vector should be imagined as specifying a point in a simplex; thus, this is a topological automaton, with the simplex being the manifold, and the stochastic matrices being linear automorphisms of the simplex onto itself. Since each transition is (essentially) independent of the previous (if we disregard the distinction between accepted and rejected languages), the PFA essentially becomes a kind of Markov chain. By contrast, in a QFA, the manifold is complex projective space C P N {\displaystyle \mathbb {C} P^{N}} , and the transition matrices are unitary matrices. Each point in C P N {\displaystyle \mathbb {C} P^{N}} corresponds to a (pure) quantum-mechanical state; the unitary matrices can be thought of as governing the time evolution of the system (viz in the Schrödinger picture). The generalization from pure states to mixed states should be straightforward: A mixed state is simply a measure-theoretic probability distribution on C P N {\displaystyle \mathbb {C} P^{N}} . A worthy point to contemplate is the distributions that result on the manifold during the input of a language. In order for an automaton to be 'efficient' in recognizing a language, that distribution should be 'as uniform as possible'. This need for uniformity is the underlying principle behind maximum entropy methods: these simply guarantee crisp, compact operation of the automaton. Put in other words, the machine learning methods used to train hidden Markov models generalize to QFAs as well: the Viterbi algorithm and the forward–backward algorithm generalize readily to the QFA. Although the study of QFA was popularized in the work of Kondacs and Watrous in 1997 and later by Moore and Crutchfeld, they were described as early as 1971, by Ion Baianu. == Measure-once automata == Measure-once automata were introduced by Cris Moore and James P. Crutchfield. They may be defined formally as follows. As with an ordinary finite automaton, the quantum automaton is considered to have N {\displaystyle N} possible internal states, represented in this case by an N {\displaystyle N} -level qudit | ψ ⟩ {\displaystyle |\psi \rangle } . More precisely, the N {\displaystyle N} -level qudit | ψ ⟩ ∈ P ( C N ) {\displaystyle |\psi \rangle \in P(\mathbb {C} ^{N})} is an element of ( N − 1 ) {\displaystyle (N-1)} -dimensional complex projective space, carrying an inner product ‖ ⋅ ‖ {\displaystyle \Vert \cdot \Vert } that is the Fubini–Study metric. The state transitions, transition matrices or de Bruijn graphs are represented by a collection of N × N {\displaystyle N\times N} unitary matrices U α {\displaystyle U_{\alpha }} , with one unitary matrix for each letter α ∈ Σ {\displaystyle \alpha \in \Sigma } . That is, given an input letter α {\displaystyle \alpha } , the unitary matrix describe
Read more →
Dan Hendrycks

Dan Hendrycks (born 1994 or 1995) is an American machine learning researcher. He serves as the director of the Center for AI Safety, a nonprofit research organization based in San Francisco, California. == Early life and education == Hendrycks was raised in a Christian evangelical household in Marshfield, Missouri. He received a B.S. from the University of Chicago in 2018 and a Ph.D. from the University of California, Berkeley in Computer Science in 2022. == Career and research == Hendrycks' research focuses on topics that include machine learning safety, machine ethics, and robustness. He credits his participation in the effective altruism (EA) movement-linked 80,000 Hours program for his career focus towards AI safety, though denies being an advocate for EA. Hendrycks is the main author of the research paper that introduced the activation function GELU in 2016, and of the paper that introduced the language model benchmark MMLU (Massive Multitask Language Understanding) in 2020. In February 2022, Hendrycks co-authored recommendations for the US National Institute of Standards and Technology (NIST) to inform the management of risks from artificial intelligence. In September 2022, Hendrycks wrote a paper providing a framework for analyzing the impact of AI research on societal risks. He later published a paper in March 2023 examining how natural selection and competitive pressures could shape the goals of artificial agents. This was followed by "An Overview of Catastrophic AI Risks", which discusses four categories of risks: malicious use, AI race dynamics, organizational risks, and rogue AI agents. Hendrycks is the safety adviser of xAI, an AI startup company founded by Elon Musk in 2023. To avoid any potential conflicts of interest, he receives a symbolic one-dollar salary and holds no company equity. In November 2024, he also joined Scale AI as an advisor collecting a one-dollar salary. Hendrycks is the creator of Humanity's Last Exam, a benchmark for evaluating the capabilities of large language models, which he developed in collaboration with Scale AI. In 2024, Hendrycks published the textbook Introduction to AI Safety, Ethics, and Society, based on courseware he had previously developed. == Selected publications == Hendrycks, Dan; Gimpel, Kevin (2020-07-08). "Gaussian Error Linear Units (GELUs)". arXiv:1606.08415 [cs.LG]. Hendrycks, Dan; Gimpel, Kevin (2018-10-03). "A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks". International Conference on Learning Representations 2017. arXiv:1610.02136. Hendrycks, Dan; Mazeika, Mantas; Dietterich, Thomas (2019-01-28). "Deep Anomaly Detection with Outlier Exposure". International Conference on Learning Representations 2019. arXiv:1812.04606. Hendrycks, Dan; Mazeika, Mantas; Zou, Andy (2021-10-25). "What Would Jiminy Cricket Do? Towards Agents That Behave Morally". Conference on Neural Information Processing Systems 2021. arXiv:2110.13136.
Read more →
Weak supervision

Weak supervision (also known as semi-supervised learning) is a paradigm in machine learning, the relevance and notability of which increased with the advent of large language models due to the large amount of data required to train them. It is characterized by using a combination of a small amount of human-labeled data (exclusively used in more expensive and time-consuming supervised learning paradigm), followed by a large amount of unlabeled data (used exclusively in unsupervised learning paradigm). In other words, the desired output values are provided only for a subset of the training data. The remaining data is unlabeled or imprecisely labeled. Intuitively, it can be seen as an exam and labeled data as sample problems that the teacher solves for the class as an aid in solving another set of problems. In the transductive setting, these unsolved problems act as exam questions. In the inductive setting, they become practice problems of the sort that will make up the exam. == Problem == The acquisition of labeled data for a learning problem often requires a skilled human agent (e.g. to transcribe an audio segment) or a physical experiment (e.g. determining the 3D structure of a protein or determining whether there is oil at a particular location). The cost associated with the labeling process thus may render large, fully labeled training sets infeasible, whereas acquisition of unlabeled data is relatively inexpensive. In such situations, semi-supervised learning can be of great practical value. Semi-supervised learning is also of theoretical interest in machine learning and as a model for human learning. == Technique == More formally, semi-supervised learning assumes a set of l {\displaystyle l} independently identically distributed examples x 1 , … , x l ∈ X {\displaystyle x_{1},\dots ,x_{l}\in X} with corresponding labels y 1 , … , y l ∈ Y {\displaystyle y_{1},\dots ,y_{l}\in Y} and u {\displaystyle u} unlabeled examples x l + 1 , … , x l + u ∈ X {\displaystyle x_{l+1},\dots ,x_{l+u}\in X} are processed. Semi-supervised learning combines this information to surpass the classification performance that can be obtained either by discarding the unlabeled data and doing supervised learning or by discarding the labels and doing unsupervised learning. Semi-supervised learning may refer to either transductive learning or inductive learning. The goal of transductive learning is to infer the correct labels for the given unlabeled data x l + 1 , … , x l + u {\displaystyle x_{l+1},\dots ,x_{l+u}} only. The goal of inductive learning is to infer the correct mapping from X {\displaystyle X} to Y {\displaystyle Y} . It is unnecessary (and, according to Vapnik's principle, imprudent) to perform transductive learning by way of inferring a classification rule over the entire input space; however, in practice, algorithms formally designed for transduction or induction are often used interchangeably. == Assumptions == In order to make any use of unlabeled data, some relationship to the underlying distribution of data must exist. Semi-supervised learning algorithms make use of at least one of the following assumptions: === Continuity / smoothness assumption === Points that are close to each other are more likely to share a label. This is also generally assumed in supervised learning and yields a preference for geometrically simple decision boundaries. In the case of semi-supervised learning, the smoothness assumption additionally yields a preference for decision boundaries in low-density regions, so few points are close to each other but in different classes. === Cluster assumption === The data tend to form discrete clusters, and points in the same cluster are more likely to share a label (although data that shares a label may spread across multiple clusters). This is a special case of the smoothness assumption and gives rise to feature learning with clustering algorithms. === Manifold assumption === The data lie approximately on a manifold of much lower dimension than the input space. In this case learning the manifold using both the labeled and unlabeled data can avoid the curse of dimensionality. Then learning can proceed using distances and densities defined on the manifold. The manifold assumption is practical when high-dimensional data are generated by some process that may be hard to model directly, but which has only a few degrees of freedom. For instance, human voice is controlled by a few vocal folds, and images of various facial expressions are controlled by a few muscles. In these cases, it is better to consider distances and smoothness in the natural space of the generating problem, rather than in the space of all possible acoustic waves or images, respectively. == History == The heuristic approach of self-training (also known as self-learning or self-labeling) is historically the oldest approach to semi-supervised learning, with examples of applications starting in the 1960s. The transductive learning framework was formally introduced by Vladimir Vapnik in the 1970s. Interest in inductive learning using generative models also began in the 1970s. A probably approximately correct learning bound for semi-supervised learning of a Gaussian mixture was demonstrated by Ratsaby and Venkatesh in 1995. == Methods == === Generative models === Generative approaches to statistical learning first seek to estimate p ( x | y ) {\displaystyle p(x|y)} , the distribution of data points belonging to each class. The probability p ( y | x ) {\displaystyle p(y|x)} that a given point x {\displaystyle x} has label y {\displaystyle y} is then proportional to p ( x | y ) p ( y ) {\displaystyle p(x|y)p(y)} by Bayes' rule. Semi-supervised learning with generative models can be viewed either as an extension of supervised learning (classification plus information about p ( x ) {\displaystyle p(x)} ) or as an extension of unsupervised learning (clustering plus some labels). Generative models assume that the distributions take some particular form p ( x | y , θ ) {\displaystyle p(x|y,\theta )} parameterized by the vector θ {\displaystyle \theta } . If these assumptions are incorrect, the unlabeled data may actually decrease the accuracy of the solution relative to what would have been obtained from labeled data alone. However, if the assumptions are correct, then the unlabeled data necessarily improves performance. The unlabeled data are distributed according to a mixture of individual-class distributions. In order to learn the mixture distribution from the unlabeled data, it must be identifiable, that is, different parameters must yield different summed distributions. Gaussian mixture distributions are identifiable and commonly used for generative models. The parameterized joint distribution can be written as p ( x , y | θ ) = p ( y | θ ) p ( x | y , θ ) {\displaystyle p(x,y|\theta )=p(y|\theta )p(x|y,\theta )} by using the chain rule. Each parameter vector θ {\displaystyle \theta } is associated with a decision function f θ ( x ) = argmax y p ( y | x , θ ) {\displaystyle f_{\theta }(x)={\underset {y}{\operatorname {argmax} }}\ p(y|x,\theta )} . The parameter is then chosen based on fit to both the labeled and unlabeled data, weighted by λ {\displaystyle \lambda } : argmax Θ ( log ⁡ p ( { x i , y i } i = 1 l | θ ) + λ log ⁡ p ( { x i } i = l + 1 l + u | θ ) ) {\displaystyle {\underset {\Theta }{\operatorname {argmax} }}\left(\log p(\{x_{i},y_{i}\}_{i=1}^{l}|\theta )+\lambda \log p(\{x_{i}\}_{i=l+1}^{l+u}|\theta )\right)} === Low-density separation === Another major class of methods attempts to place boundaries in regions with few data points (labeled or unlabeled). One of the most commonly used algorithms is the transductive support vector machine, or TSVM (which, despite its name, may be used for inductive learning as well). Whereas support vector machines for supervised learning seek a decision boundary with maximal margin over the labeled data, the goal of TSVM is a labeling of the unlabeled data such that the decision boundary has maximal margin over all of the data. In addition to the standard hinge loss ( 1 − y f ( x ) ) + {\displaystyle (1-yf(x))_{+}} for labeled data, a loss function ( 1 − | f ( x ) | ) + {\displaystyle (1-|f(x)|)_{+}} is introduced over the unlabeled data by letting y = sign ⁡ f ( x ) {\displaystyle y=\operatorname {sign} {f(x)}} . TSVM then selects f ∗ ( x ) = h ∗ ( x ) + b {\displaystyle f^{}(x)=h^{}(x)+b} from a reproducing kernel Hilbert space H {\displaystyle {\mathcal {H}}} by minimizing the regularized empirical risk: f ∗ = argmin f ( ∑ i = 1 l ( 1 − y i f ( x i ) ) + + λ 1 ‖ h ‖ H 2 + λ 2 ∑ i = l + 1 l + u ( 1 − | f ( x i ) | ) + ) {\displaystyle f^{}={\underset {f}{\operatorname {argmin} }}\left(\displaystyle \sum _{i=1}^{l}(1-y_{i}f(x_{i}))_{+}+\lambda _{1}\|h\|_{\mathcal {H}}^{2}+\lambda _{2}\sum _{i=l+1}^{l+u}(1-|f(x_{i})|)_{+}\right)} An exact solution is intractable due to the non-convex term ( 1 − | f ( x ) | ) + {\displayst
Read more →
AI Image Generators: Free vs Paid (2026)

Looking for the best AI image generator? An AI image generator is software that uses machine learning to help you get more done — it can save you hours every week by automating repetitive work. Most options offer a generous free tier, with paid plans unlocking higher limits, faster processing, and team features. Whether you are a beginner or a pro, the right AI image generator slots into your workflow and pays for itself fast. This guide breaks down the top picks, their pros and cons, and who each one is best for.
Read more →
Yaron Singer

Yaron Singer is a computer scientist and entrepreneur whose work has focused on algorithms, machine learning, optimization, and artificial intelligence security. He was the Gordon McKay Professor of Computer Science and Applied Mathematics at Harvard University and co-founded Robust Intelligence, an artificial intelligence security company acquired by Cisco Systems in 2024. == Education == Singer received a PhD in computer science from the University of California, Berkeley under the supervision of Christos Papadimitriou. == Academic career == Singer was a postdoctoral research scientist at Google Research. Singer joined the computer science faculty at Harvard John A. Paulson School of Engineering and Applied Sciences in 2013 and became a full professor in 2019. == Research == Singer's research has focused on algorithms and machine learning, including optimization, algorithmic mechanism design, and adversarial machine learning. His doctoral work studied computational limits in algorithmic mechanism design, including truthful mechanisms and budget-feasible mechanisms. In optimization, Singer co-authored work on submodular optimization and parallel algorithms for large-scale data processing. Singer has also worked on adversarial machine learning, including attacks that use small perturbations or noise to affect the behavior of machine learning systems. == Entrepreneurship == In 2020, Singer co-founded Robust Intelligence Kojin Oshiba. Harvard SEAS reported that the company raised $14 million that year, and TechCrunch reported in 2021 that the company raised a $30 million Series B round led by Tiger Global. The company developed tools for testing AI models and detecting failures before or during deployment. TechCrunch described its RIME product as using an "AI firewall" to stress-test models. In 2024, Cisco Systems acquired Robust Intelligence. CTech reported that Cisco had not disclosed the purchase amount when the acquisition was announced, and later reported the deal value as $400 million. In 2025, Cisco launched Foundation AI, a Cisco team focused on AI for cybersecurity. Techzine reported that Singer led the team and was Cisco's VP of AI and Security. == Recognition == Singer has received a Sloan Research Fellowship, an NSF CAREER Award, a Google Faculty Research Award, and a Facebook Faculty Award. As a graduate student, he received Microsoft Research and Facebook fellowships. In 2012, he received the Best Student Paper Award at the ACM International Conference on Web Search and Data Mining for "How to Win Friends and Influence People, Truthfully: Influence Maximization Mechanisms for Social Networks."
Read more →
Top 10 AI Website Builders Compared (2026)

Curious about the best AI website builder? An AI website builder is software that uses machine learning to help you get more done — it combines speed, accuracy, and an interface that just works. Hands-on testing shows real-world results vary, so a short free trial is the smartest way to decide. Whether you are a beginner or a pro, the right AI website builder slots into your workflow and pays for itself fast. Read on for hands-on impressions, pricing tiers, and the standout features that matter.
Read more →
Psychology in cybersecurity

The psychology of cybersecurity (often intersecting with usable security and cyberpsychology) is an interdisciplinary field studying how human behavior, cognitive biases, and social dynamics influence information security. While traditional cybersecurity focuses on hardware and software vulnerabilities, this discipline addresses the "human factor," which is exploited in cyberattacks. Psychology in cybersecurity draws from cognitive psychology and human–computer interaction. == History and evolution == The challenge of human behavior in computing was noted as early as the 1960s with multi-user mainframes like the Compatible Time-Sharing System (CTSS). In 1966, a software error on CTSS caused the system's master password file to be displayed to every user upon login—one of the earliest documented security incidents attributable to a combination of system design and human factors. These behaviors gained broader significance in the 1990s as the Internet became widely accessible. High-profile incidents involving figures like Kevin Mitnick demonstrated how human trust could be exploited through social engineering such as pretexting over the phone. == Cognitive and behavioral factors == Much of the psychology of cybersecurity focuses on decision-making under stress or uncertainty. Researchers apply frameworks like dual process theory to explain why humans fall for phishing or business email compromise. Threat actors design malicious communications to trigger fast, emotional "System 1" thinking—using urgency, authority, or panic, which prompts users to click a link or wire funds before their analytical "System 2" can assess the situation's legitimacy. Industry research has consistently documented the effectiveness of these techniques at scale, pointing to several recurring psychological phenomena that influence daily security practices: Cognitive biases: The optimism bias leads users to believe they are unlikely to be targeted by cybercriminals, resulting in lax password practices or delayed software updates. The availability heuristic causes individuals to focus on highly publicized, sophisticated threats while ignoring common, statistically probable risks like credential reuse. Social influence: Attackers leverage established principles of persuasion, such as those categorized by Robert Cialdini. Impersonating a CEO leverages the psychological trigger of authority, while fake tech support scams use reciprocity (offering to fix a problem before asking for network credentials). == Neurological and pre-cognitive factors == Functional magnetic resonance imaging (fMRI) studies show that neural activation in visual and attentional regions decreases with repeated exposure to the same stimulus, a phenomenon termed repetition suppression. Experiments have confirmed this effect in the context of security warnings: static warning designs produce declines in user attention and adherence. Information processing research on phishing indicates that affective cues, such as artificial urgency or fear, increase cognitive load and elicit automatic heuristic processing, reducing the likelihood of analytical evaluation and facilitating compliance with malicious requests. == Security fatigue and organizational dynamics == Aggressive cybersecurity postures can sometimes lead to mental and emotional exhaustion, a phenomenon known as security fatigue. === Alert fatigue === One example is alert fatigue, which most frequently affects both end-users and security operations center analysts. Continuous exposure to browser warnings or antivirus pop-ups, particularly those that are false positives, conditions users to dismiss alerts automatically due to the volume of notifications rather than their repetitive appearance (see § Neurological and pre-cognitive factors). The scale of this problem is significant in enterprise: SOC teams in large organizations receive thousands of alerts daily, and a survey published in ACM Computer Surveys found that analysts spend over 25% of their time handling false positives, meaning that malicious indicators can be buried in the noise. === Password fatigue === Similarly, password fatigue is the feeling experienced by many people who are required to remember an excessive number of passwords as part of their daily routine, such as to log in to a computer at work. Users cope with the memory burden by making predictable, iterative changes to their passwords (such as updating "Password01!" to "Password02!"), which decreases password security.
Read more →
Heikki Mannila

Heikki Olavi Mannila (born 4 January 1960 in Espoo) is a Finnish computer scientist, the president of the Academy of Finland. Mannila earned his Ph.D. in 1985 from the University of Helsinki under the supervision of Esko Ukkonen and for many years he was a professor at the University of Helsinki himself. From 2004 to 2008 he was Academy Professor at the Academy of Finland. He became Vice President for Academic Affairs at Aalto University in 2009, and was appointed by the Finnish government as president of the Academy of Finland for a term lasting from 2012 to 2017. The appointment was renewed for the period 2017–2022. Mannila is known for his research in data mining, and has published highly cited papers on association rule learning and sequence mining. With David Hand and Padhraic Smyth, he is the co-author of the book Principles of Data Mining (MIT Press, 2001). Heikki Mannila is son to the professor Elina Haavio-Mannila.
Read more →
The Best Free AI Code-review Tool for Beginners

Curious about the best AI code-review tool? An AI code-review tool is software that uses machine learning to help you get more done — it combines speed, accuracy, and an interface that just works. Hands-on testing shows real-world results vary, so a short free trial is the smartest way to decide. Whether you are a beginner or a pro, the right AI code-review tool slots into your workflow and pays for itself fast. This guide breaks down the top picks, their pros and cons, and who each one is best for.
Read more →
Quotient automaton

In computer science, in particular in formal language theory, a quotient automaton can be obtained from a given nondeterministic finite automaton by joining some of its states. The quotient recognizes a superset of the given automaton; in some cases, handled by the Myhill–Nerode theorem, both languages are equal. == Formal definition == A (nondeterministic) finite automaton is a quintuple A = ⟨Σ, S, s0, δ, Sf⟩, where: Σ is the input alphabet (a finite, non-empty set of symbols), S is a finite, non-empty set of states, s0 is the initial state, an element of S, δ is the state-transition relation: δ ⊆ S × Σ × S, and Sf is the set of final states, a (possibly empty) subset of S. A string a1...an ∈ Σ is recognized by A if there exist states s1, ..., sn ∈ S such that ⟨si-1,ai,si⟩ ∈ δ for i=1,...,n, and sn ∈ Sf. The set of all strings recognized by A is called the language recognized by A; it is denoted as L(A). For an equivalence relation ≈ on the set S of A’s states, the quotient automaton A/≈ = ⟨Σ, S/≈, [s0], δ/≈, Sf/≈⟩ is defined by the input alphabet Σ being the same as that of A, the state set S/≈ being the set of all equivalence classes of states from S, the start state [s0] being the equivalence class of A’s start state, the state-transition relation δ/≈ being defined by δ/≈([s],a,[t]) if δ(s,a,t) for some s ∈ [s] and t ∈ [t], and the set of final states Sf/≈ being the set of all equivalence classes of final states from Sf. The process of computing A/≈ is also called factoring A by ≈. == Example == For example, the automaton A shown in the first row of the table is formally defined by ΣA = {0,1}, SA = {a,b,c,d}, sA0 = a, δA = { ⟨a,1,b⟩, ⟨b,0,c⟩, ⟨c,0,d⟩ }, and SAf = { b,c,d }. It recognizes the finite set of strings { 1, 10, 100 }; this set can also be denoted by the regular expression "1+10+100". The relation (≈) = { ⟨a,a⟩, ⟨a,b⟩, ⟨b,a⟩, ⟨b,b⟩, ⟨c,c⟩, ⟨c,d⟩, ⟨d,c⟩, ⟨d,d⟩ }, more briefly denoted as a≈b,c≈d, is an equivalence relation on the set {a,b,c,d} of automaton A’s states. Building the quotient of A by that relation results in automaton C in the third table row; it is formally defined by ΣC = {0,1}, SC = {a,c}, sC0 = a, δC = { ⟨a,1,a⟩, ⟨a,0,c⟩, ⟨c,0,c⟩ }, and SCf = { a,c }. It recognizes the finite set of all strings composed of arbitrarily many 1s, followed by arbitrarily many 0s, i.e. { ε, 1, 10, 100, 1000, ..., 11, 110, 1100, 11000, ..., 111, ... }; this set can also be denoted by the regular expression "10". Informally, C can be thought of resulting from A by glueing state a onto state b, and glueing state c onto state d. The table shows some more quotient relations, such as B = A/a≈b, and D = C/a≈c. == Properties == For every automaton A and every equivalence relation ≈ on its states set, L(A/≈) is a superset of (or equal to) L(A). Given a finite automaton A over some alphabet Σ, an equivalence relation ≈ can be defined on Σ by x ≈ y if ∀ z ∈ Σ: xz ∈ L(A) ↔ yz ∈ L(A). By the Myhill–Nerode theorem, A/≈ is a deterministic automaton that recognizes the same language as A. As a consequence, the quotient of A by every refinement of ≈ also recognizes the same language as A.
Read more →
Shape factor (image analysis and microscopy)

Shape factors are dimensionless quantities used in image analysis and microscopy that numerically describe the shape of a particle, independent of its size. Shape factors are calculated from measured dimensions, such as diameter, chord lengths, area, perimeter, centroid, moments, etc. The dimensions of the particles are usually measured from two-dimensional cross-sections or projections, as in a microscope field, but shape factors also apply to three-dimensional objects. The particles could be the grains in a metallurgical or ceramic microstructure, or the microorganisms in a culture, for example. The dimensionless quantities often represent the degree of deviation from an ideal shape, such as a circle, sphere or equilateral polyhedron. Shape factors are often normalized, that is, the value ranges from zero to one. A shape factor equal to one usually represents an ideal case or maximum symmetry, such as a circle, sphere, square or cube. == Aspect ratio == The most common shape factor is the aspect ratio, a function of the largest diameter and the smallest diameter orthogonal to it: A R = d min d max {\displaystyle A_{R}={\frac {d_{\min }}{d_{\max }}}} The normalized aspect ratio varies from approaching zero for a very elongated particle, such as a grain in a cold-worked metal, to near unity for an equiaxed grain. The reciprocal of the right side of the above equation is also used, such that the AR varies from one to approaching infinity. == Circularity == Another very common shape factor is the circularity (or isoperimetric quotient), a function of the perimeter P and the area A: f circ = 4 π A P 2 {\displaystyle f_{\text{circ}}={\frac {4\pi A}{P^{2}}}} The circularity of a circle is 1, and much less than one for a starfish footprint. The reciprocal of the circularity equation is also used, such that fcirc varies from one for a circle to infinity. == Elongation shape factor == The less-common elongation shape factor is defined as the square root of the ratio of the two second moments in of the particle around its principal axes. f elong = i 2 i 1 {\displaystyle f_{\text{elong}}={\sqrt {\frac {i_{2}}{i_{1}}}}} == Compactness shape factor == The compactness shape factor is a function of the polar second moment in of a particle and a circle of equal area A. f comp = A 2 2 π i 1 2 + i 2 2 {\displaystyle f_{\text{comp}}={\frac {A^{2}}{2\pi {\sqrt {{i_{1}}^{2}+{i_{2}}^{2}}}}}} The fcomp of a circle is one, and much less than one for the cross-section of an I-beam. == Waviness shape factor == The waviness shape factor of the perimeter is a function of the convex portion Pcvx of the perimeter to the total. f wav = P cvx P {\displaystyle f_{\text{wav}}={\frac {P_{\text{cvx}}}{P}}} Some properties of metals and ceramics, such as fracture toughness, have been linked to grain shapes. == An application of shape factors == Greenland, the largest island in the world, has an area of 2,166,086 km2; a coastline (perimeter) of 39,330 km; a north–south length of 2670 km; and an east–west length of 1290 km. The aspect ratio of Greenland is A R = 1290 2670 = 0.483 {\displaystyle A_{R}={\frac {1290}{2670}}=0.483} The circularity of Greenland is f circ = 4 π ( 2166086 ) 39330 2 = 0.0176. {\displaystyle f_{\text{circ}}={\frac {4\pi (2166086)}{39330^{2}}}=0.0176.} The aspect ratio is agreeable with an eyeball-estimate on a globe. Such an estimate on a typical flat map, using the Mercator projection, would be less accurate due to the distorted scale at high latitudes. The circularity is deceptively low, due to the fjords that give Greenland a very jagged coastline (see the coastline paradox). A low value of circularity does not necessarily indicate a lack of symmetry, and shape factors are not limited to microscopic objects.
Read more →
Li Sheng (computer scientist)

Li Sheng (Chinese: 李生; born 1943), is a professor at the School of Computer Science and Engineering, Harbin Institute of Technology (HIT), China. He began his research on Chinese-English machine translation in 1985, making himself one of the earliest Chinese scholars in this field. After that, he pursued in vast topics of natural language processing, including machine translation, information retrieval, question answering and applied artificial intelligence. He was the final review committee member for computer area in NSF China. Born and raised in Heilongjiang province, he graduated in 1965 from the computer specialty of HIT, which is one of the earliest computer specialties in Chinese universities. Then he started to work as a staff in the Computer specialty of HIT, which was finally granted as a department in 1985. Also from 1985, he was appointed to undertake a series administrative positions in HIT, e.g. Dean of Computer Department(1987–1988), Director of R&D Division (1988–1990), Chief R&D Officer and several other key leading positions in HIT. Resigned all his administrative positions in 2004, Li devoted himself as the director of MOE-Microsoft Join Key Lab of NLP& Speech (HIT), making it a leading NLP research group with more than 100 staffs and students working on various aspects of NLP. So far, the lab has already been granted for dozens of technology awards by the ministries of central government and local provincial government of China. Its research progresses are reported annually in top tier conferences including ACL, IJCAI, SIGIR etc. As one of the pioneers in NLP research in China, he contributes NLP in China not only in technology innovations but also in talents education. So far, his research group has graduated more than 60 Ph.D. and almost 200 M.E with NLP major. Most of them are now working as the chief researcher in various NLP groups of universities and companies in China, including several world-known NLP scholars, such as Wang Haifeng of Baidu, Zhou Ming of Microsoft Research, Zhang Min (张民) of Soochow University (China), and Zhao Tiejun (赵铁军) and Liu Ting (刘挺) of HIT. Owing to his contributions in Chinese language processing, Li was elected as the President of Chinese Information Processing Society of China (CIPSC) in 2011. He scaled this top level academic organization in China up to more than 3000 registered members, and promoted NLP into several national projects for research or industry development. In addition, the CIPSC is now enhancing its co-operations with world NLP organizations including ACL. == Machine Intelligence & Translation Laboratory (MI&TLAB) == Originates from Machine Translation Research Group of Computer Science Department, Harbin Institute of Technology, which was started Li in 1985. It is one of the earliest institutions engaged in MT research in China, featured by its investigations into Chinese-English machine translation. It is now running under the Research Center on Language Technology, School of Computer Science and Technology, HIT. Details for staffs and publications can be found at https://mitlab.hit.edu.cn. == MOE-MS Joint Key Lab of Natural Language Processing and Speech (HIT) == In June, 2000, the Joint HIT-Microsoft Machine Translation Lab was founded by MI&T Lab and Microsoft Research (China). It was the third joint lab established by Microsoft Research (China) with Chinese universities, and the only one focusing on Machine Translation. Based on this jointly lab, the cooperation between HIT and Microsoft gradually extended to the areas of machine translation, information retrieval, speech recognition and processing, natural language understanding. In Oct, 2004, the joint key lab was granted as one of the 10 joint key labs supported by the Microsoft Research of Asia and Ministry of Education in China. In July 2006, the Shenzhen extension of the lab was launched. More than 200 staff and students have undertaken research projects, including some sponsored by the National Natural Science Foundation of China and the National 863 program of China. Since 2005, the lab has also been organizing a summer camp in Harbin Institute of Technology, and approximately 150 faculty members and students from universities in China have participated. This summer workshop was organized annually until 2014, when it was organized formally as the summer school series by Chinese Information Processing Society, China. Through the lab, a Microsoft Research of Asia-HIT joint PhD program was implemented in 2012. == CEMT-I MT System == In May 1989, CEMT-I passed the formal project appraisal in Harbin, China. Capable of translating technical paper titles from Chinese to English, it is not only the first MT system completed by Li and his group, but also the first Chinese-English Translation system that passed the technical appraisal by Chinese government according to the public reports. It was then awarded the Second Prize of Ministry Level Technology Innovation by the former National Aerospace Industry Corporation in 1990. == Daya Translation Workstation == Owing to the technical achievements by Li's group in Chinese-English machine translation, the former National Aerospace Industry Corporation of China sponsored a commercial system development of "Daya Translation Station (MT)" in 1993. Designed as a comprehensive English composition aid for Chinese users, this system was finished and put into the market in 1995. And in 1997, this system was awarded the Second Prize of Ministry Level Technology Innovation by the former National Aerospace Industry Corporation. == BT863 MT System == From 1994, the researches in Li's lab were supported by National 863 Hi-tech Research and Development Program. During this period, the BT863 system was explored to employ one engine for both Chinese-English and English-Chinese translation. This system was proved to be the best performance among Chinese-English MT systems in the formal technical evaluation of National 863 program, yielding the Third Prize of Ministry Level Technology Innovation by the former National Aerospace Industry Corporation in 1997. == Next Generation IR == This is a key project granted by NSF China (with a joint sponsorship from MSRA) started form 2008. In contrast to his previous NSF grants for different NLP issues, Li explored in his last PI project on key technologies in personalized IR, together with researchers from Tsinghua University and Institute of Software, Chinese Academy of Science. With impressive publications in top tier journals and conferences (including breakthrough publications in SIGIR of his own group), this projected was approved "A-level" achievements by the NSF China office in 2012.
Read more →
Is an AI Analytics Tool Worth It in 2026?

Curious about the best AI analytics tool? An AI analytics tool is software that uses machine learning to help you get more done — it combines speed, accuracy, and an interface that just works. Hands-on testing shows real-world results vary, so a short free trial is the smartest way to decide. Whether you are a beginner or a pro, the right AI analytics tool slots into your workflow and pays for itself fast. Read on for hands-on impressions, pricing tiers, and the standout features that matter.
Read more →