Unified normal form active retrieval method and system and storage medium

Abstract

The invention discloses a unified normal form active retrieval method and system and a storage medium. The method comprises the following steps: acquiring an original question of a user; inputting the original problem into the large model, and obtaining hidden layer state data of the last layer of the large model; and inputting the hidden layer state data into a preset classifier combination to determine whether the original problem needs to be retrieved, and obtaining a retrieval judgment result. According to the method, the preset classifier combination is added into the large model, whether the original question needs to be retrieved or not is determined, the number of times of retrieval calling can be optimized, the accuracy and efficiency of answering of the large model are improved, and the response time delay of the large model is shortened.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a unified paradigm active retrieval method, system and computer-readable storage medium. Background Art

[0002] With the continuous advancement of technology, artificial intelligence dialogue systems based on large language models (LLMs) have performed well in various types of instructions such as programming, writing, and reasoning. However, existing large models sometimes generate false information, a phenomenon known as "hallucination", which weakens the credibility of the large models' answers in real-world scenarios.

[0003] Retrieval-Augmented Generation (RAG) is a popular approach to address the problem of large model hallucination. RAG first retrieves documents related to the user query and then uses these documents to enhance the factual correctness of the content generated by the large model. However, retrieval is not always helpful, and calling retrieval for every instruction is a suboptimal strategy. First, for instructions that do not require factual knowledge, such as creation, translation, and rewriting, the additional retrieved documents may interfere with the creation ability of the large model itself and reduce the flexibility of the model. Second, if the retrieved documents are of poor quality, such as containing incorrect factual knowledge, they may interfere with or mislead the model and generate incorrect responses. Finally, the retrieval process increases user waiting time. If retrieval is called for all user queries, it will significantly increase the response time of the system and reduce the user experience. Therefore, it is crucial to decide whether retrieval is needed in RAG. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a unified paradigm active retrieval method, system and computer-readable storage medium, which can optimize the number of retrieval calls, improve the accuracy and efficiency of the large model's answers, and reduce the response delay of the large model by adding a preset classifier combination to the large model to determine whether the original question needs to be retrieved.

[0005] The first aspect of the present invention provides a unified paradigm active retrieval method, including:

[0006] Get the user's original question;

[0007] Input the original problem into the large model to obtain the hidden layer state data of the last layer of the large model;

[0008] The hidden layer state data is input into a preset classifier combination to determine whether the original question needs to be retrieved, and a retrieval determination result is obtained.

[0009] Optionally, the method further comprises:

[0010] When the search result indicates that a search is required, a search is performed from a preset knowledge base according to the original question to obtain N corresponding knowledge texts; wherein N≥1;

[0011] Combining the original question with all the knowledge texts to obtain a prompt text;

[0012] The prompt text is input into the large model to obtain the answer result for the user.

[0013] Optionally, each classifier in the preset classifier combination is used to perform a corresponding binary classification task; wherein any two of the binary classification tasks do not overlap.

[0014] Optionally, the preset classifier combination includes: an intention-aware classifier, a knowledge-aware classifier, a time-aware classifier and a self-aware classifier.

[0015] Optionally, the input of each classifier in the preset classifier combination is the hidden layer state data.

[0016] Optionally, each classifier in the preset classifier combination is enabled according to a preset order; or,

[0017] All of the classifiers are enabled simultaneously.

[0018] Optionally, the activation order of each classifier in the preset classifier combination is specifically:

[0019] Enabling the intent-aware classifier to determine whether the user has an intent to search;

[0020] If there is an intention to search, the search judgment result is output as needing to search; otherwise, the knowledge-aware classifier is enabled to determine whether the original question needs to be answered with factual knowledge;

[0021] If factual knowledge is not required to answer, the search decision result is output as no search is required; otherwise, a time-aware classifier is enabled to determine whether the answer to the original question changes over time;

[0022] If it changes over time, the retrieval judgment result is output as needing retrieval; otherwise, the self-awareness classifier is enabled to determine whether the large model has the knowledge content required to answer the original question.

[0023] Optionally, the training data of each classifier in the preset classifier combination consists of a number of user questions; wherein each of the training data includes a plurality of corresponding positive samples and negative samples.

[0024] A second aspect of the present invention provides a unified paradigm active retrieval system, including:

[0025] Question acquisition module, used to obtain the user's original question;

[0026] A hidden layer data acquisition module, used to input the original problem into the large model and obtain the hidden layer state data of the last layer of the large model;

[0027] The retrieval determination module is used to input the hidden layer state data into a preset classifier combination to determine whether the original question needs to be retrieved and obtain a retrieval determination result.

[0028] An embodiment of the third aspect of the present invention is a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the unified paradigm active retrieval method described in any embodiment of the first aspect above.

[0029] Compared with the prior art, the embodiments of the present invention provide a unified paradigm active retrieval method, system and computer-readable storage medium, the method comprising: first, obtaining the original question of the user; then, inputting the original question into the large model to obtain the hidden layer state data of the last layer of the large model; then, inputting the hidden layer state data into the preset classifier combination to determine whether the original question needs to be retrieved, and obtaining the retrieval judgment result. The present invention adds a preset classifier combination to the large model to determine whether the original question needs to be retrieved, thereby optimizing the number of retrieval calls, improving the accuracy and efficiency of the large model's answers, and reducing the response delay of the large model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flowchart of an embodiment of the unified paradigm active retrieval method provided by the present invention;

[0031] Figure 2 is a schematic diagram of another embodiment of the unified paradigm active retrieval method provided by the present invention;

[0032] Figure 3 is a schematic diagram of an embodiment of the original problem type provided by the present invention;

[0033] Figure 4 It is a structural diagram of an embodiment of the unified paradigm active retrieval system provided by the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this technical field without creative work are within the scope of protection of the present invention.

[0035] See also Figure 1 , is a flow chart of an embodiment of the unified paradigm active retrieval method provided by the present invention.

[0036] The first embodiment of the present invention provides a unified paradigm active retrieval method, including steps S1 to S3, which are as follows:

[0037] Step S1: Obtain the user's original question;

[0038] Step S2: input the original problem into the large model to obtain the hidden layer state data of the last layer of the large model;

[0039] Step S3: input the hidden layer state data into a preset classifier combination to determine whether the original question needs to be retrieved, and obtain a retrieval judgment result.

[0040] Specifically, the input of each classifier in the preset classifier combination in the above step S23 is the hidden layer state data.

[0041] It should be noted that each classifier in the preset classifier combination is a lightweight and plug-and-play component. Figure 2 FIG. 1 is a schematic diagram of another embodiment of the unified paradigm active retrieval method provided by the present invention. Figure 2 In (a), the hidden state data of the last layer of the large model (i.e., the input vector of the language prediction head) is used as the input of each classifier to perform a binary classification task; wherein, the hidden state data of the last layer contains a deep semantic understanding of the user's original question. Therefore, each classifier is designed as a single-layer fully connected network with an input dimension equal to the hidden layer dimension and an output dimension equal to 2. In addition, these classifiers do not need to modify the parameters of the large model during training, thereby avoiding expensive fine-tuning processes and not causing a decrease in the performance of the large model. Based on this, the embodiment of the present invention adds a preset classifier combination to the large model to determine whether the original question needs to be retrieved, thereby optimizing the number of retrieval calls, improving the accuracy and efficiency of the large model's answers, and reducing the response delay of the large model.

[0042] In an optional embodiment, each classifier in the preset classifier combination is used to perform a corresponding binary classification task; wherein any two of the binary classification tasks do not overlap with each other.

[0043] It should be noted that any two binary classification tasks do not overlap with each other, indicating that the judgment criteria of the corresponding two classifiers are orthogonal. In order to better understand the technical solution in the embodiment of the present invention, the following four orthogonal active retrieval enhancement judgment criteria proposed by the inventor are used as reference examples:

[0044] (1) Intent-aware: Determine whether the user expresses the intention to retrieve or use external information;

[0045] (2) Knowledge-awareness: Determine whether the question requires factual knowledge to answer;

[0046] (3) Time-sensitive-aware: Determine whether the question is a time-sensitive question, that is, whether the answer changes over time;

[0047] (4) Self-awareness: Determine whether the large model contains the knowledge content required to answer the original question.

[0048] It is worth noting that the above judgment criteria are only examples and are not limited to the four judgment criteria of intention perception, knowledge perception, time perception and self perception. The embodiments of the present invention can be expanded and refined according to specific application scenarios to better adapt to diverse application requirements.

[0049] In an optional embodiment, the preset classifier combination includes: an intention-aware classifier, a knowledge-aware classifier, a time-aware classifier, and a self-aware classifier.

[0050] like Figure 3 The figure shows a schematic diagram of an embodiment of the original question type provided by the present invention. Figure 3 The various types of original questions can be judged by the following classifiers:

[0051] (1) an intent-aware classifier is used to determine whether the user has an intent to search;

[0052] (2) a knowledge-aware classifier is used to determine whether the original question requires an answer using factual knowledge;

[0053] (3) a time-aware classifier for determining whether the answer to the original question changes over time;

[0054] (4) The self-awareness classifier is used to determine whether the large model has the knowledge content required to answer the original question.

[0055] It is worth noting that existing active retrieval methods usually rely on a single standard and have difficulty handling different types of user questions (user instructions), such as simply judging whether the instruction is related to knowledge, or whether the model knows a specific question. In contrast, the embodiment of the present invention can handle user questions more comprehensively and flexibly through the combined application of four classifiers, namely, intent-aware classifier, knowledge-aware classifier, time-aware classifier and self-aware classifier, that is, it ensures that the large model can quickly and accurately determine whether the user's original question needs to be retrieved when facing queries of different types and high complexity. Based on this, the large model in the embodiment of the present invention can not only provide more accurate and timely responses, but also significantly improve the user experience, reduce unnecessary external retrieval, and thus optimize resource utilization and response time.

[0056] In an optional embodiment, the training data of each classifier in the preset classifier combination consists of a number of user questions; wherein each of the training data includes a plurality of corresponding positive samples and negative samples.

[0057] It should be noted that the output label of each positive sample is that it needs to be retrieved, and the output label of the negative sample is that it does not need to be retrieved. Each classifier in the preset classifier combination is a lightweight and plug-and-play binary classifier, which can avoid modifying the parameters of the original large model. Therefore, there is no need to retrain the large model, and only each classifier needs to be trained.

[0058] The embodiment of the present invention constructs specific training data for each classifier. All training data are composed of binary classification data consisting of positive samples and negative samples. Each classifier is used to determine whether the user question needs to be retrieved in the corresponding scenario. The details are as follows:

[0059] In the intent-aware scenario, we collect a variety of explicit or implicit user expressions for retrieval, information search, or external data verification, and concatenate these expressions with ordinary user input as positive samples, while the input in which the user does not explicitly express the need for external information query is used as negative samples.

[0060] In the self-awareness scenario, an open domain knowledge question answering dataset is used to determine whether the model knows the answer to a given question, and questions that the model knows and does not know are constructed as training data; questions that the model does not know are used as positive samples that need to be retrieved, and questions that the model knows are used as negative samples that do not need to be retrieved.

[0061] In the timeliness-aware scenario, questions whose answers change over time and those that do not change over time are used for training. Questions whose answers change over time are used as positive samples that need to be retrieved, and questions whose answers do not change over time are used as negative samples that do not need to be retrieved.

[0062] In the knowledge perception scenario, factual question and answer data and general instructions that do not involve factual knowledge are used for training. Factual question and answer data serve as positive samples that need to be retrieved, and general instructions that do not involve factual knowledge serve as negative samples.

[0063] Through the above-mentioned training data construction and classifier design, the embodiment of the present invention can ensure that the RAG-based large model can comprehensively and efficiently process different types of user instructions, thereby improving the accuracy of information retrieval and user experience.

[0064] In an optional embodiment, each classifier in the preset classifier combination is enabled according to a preset order; or,

[0065] All of the classifiers are enabled simultaneously.

[0066] It should be noted that, since the judgment criteria corresponding to any two classifiers are orthogonal, the embodiment of the present invention can enable each classifier in the preset classifier combination in two ways:

[0067] The first implementation is to enable in a preset order: all classifiers corresponding to the judgment criteria are integrated into a decision tree, and four classifiers are used to determine the timing of retrieval according to the specified priority order in the reasoning stage.

[0068] The second implementation method is to enable all classifiers at the same time: all classifiers are enabled at the same time, each classifier makes independent judgments, and finally all judgment results are considered comprehensively to determine whether the user's original question needs to be retrieved.

[0069] Compared with the first implementation, the second implementation can evaluate the original problem at one time (i.e., each classifier analyzes the problem from a different perspective), but running all classifiers at the same time consumes more computing resources. Therefore, when it is necessary to balance computing resource consumption, the first implementation is used.

[0070] In the first implementation, the preset order can be selected in practical applications:

[0071] Example 1: For each specific industry field, statistics on various user issues (such as Figure 3 The occurrence frequencies of the classifiers are calculated and the activation order (priority) of the classifiers is arranged from high to low according to these frequencies.

[0072] Example 2: For all industry fields, a common active search discrimination order is set in the reasoning stage, which can adapt to different types of user inputs. The activation order is intent-aware classifier, knowledge-aware classifier, time-aware classifier and self-aware classifier. In specific implementation, the activation order of each classifier in the preset classifier combination is as follows:

[0073] Enabling the intent-aware classifier to determine whether the user has an intent to search;

[0074] If there is an intention to search, the search judgment result is output as needing to search; otherwise, the knowledge-aware classifier is enabled to determine whether the original question needs to be answered with factual knowledge;

[0075] If factual knowledge is not required to answer, the search decision result is output as no search is required; otherwise, the time-aware classifier is enabled to determine whether the answer to the original question changes over time;

[0076] If it changes over time, the retrieval judgment result is output as needing retrieval; otherwise, the self-awareness classifier is enabled to determine whether the large model has the knowledge content required to answer the original question.

[0077] like Figure 2 (b) is a schematic diagram of an embodiment of the classifier activation sequence provided by the present invention. Figure 2 In (b), when the self-awareness classifier is enabled to determine whether the large model has the knowledge content required to answer the original question, if it has the knowledge content required to answer the original question, the output retrieval judgment result is that no retrieval is required; otherwise, the output retrieval judgment result is that retrieval is required.

[0078] In other words, the general active retrieval judgment order (i.e., the activation order of each classifier) ​​in the reasoning stage is as follows: first, determine whether the original question X contains the user's intention to search. If the user expresses the intention to search, the retrieval judgment result is output as need to search; otherwise, continue to determine whether question X requires factual knowledge to answer. If X does not require factual knowledge to answer, the retrieval judgment result is output as no need to search. For question X that requires factual knowledge to answer, continue to determine whether the required factual knowledge will change over time. For factual knowledge that may change over time, the retrieval judgment result is output as need to search. For factual knowledge that does not change over time, finally determine whether the large model itself already has the corresponding knowledge. Only when the large model does not have the corresponding knowledge, the retrieval judgment result is output as need to search.

[0079] It is worth noting that existing active retrieval methods are highly dependent on specialized and differentiated designs and processes. For example, these methods require full parameter fine-tuning, use of uncertainty estimation based on model calibration, or training of external classifiers for specific scenarios. This complex design makes it very difficult to integrate active retrieval of different standards into a unified retrieval enhancement generation system, and may result in higher response delays. In contrast, the embodiment of the present invention uses the design of a multi-level classifier, so that large models can not only respond to various user inputs more intelligently in complex and changeable practical application environments, but also improve retrieval efficiency and reduce response delays.

[0080] In an optional embodiment, the method further includes:

[0081] When the search result indicates that a search is required, a search is performed from a preset knowledge base according to the original question to obtain N corresponding knowledge texts; wherein N≥1;

[0082] Combining the original question with all the knowledge texts to obtain a prompt text;

[0083] The prompt text is input into the large model to obtain the answer result for the user.

[0084] During specific implementation, for user questions (instructions) that require retrieval enhancement, the retrieved external information (knowledge text) is spliced ​​with the user's original input (original question) to obtain a prompt text (Prompt); this can avoid the large model from re-encoding the original instruction (original question) to reduce the overhead of computing resources. Then, the prompt text is input into the large model to guide the large model to generate the answer result expected by the user. In the above process, the retrieved external knowledge text can be the top five candidate documents ranked by similarity. Of course, the embodiment of the present invention can also add some restrictions on the answer content in the Prompt, such as allowing the large model to only use the knowledge text recalled based on the original question to answer, so as to prevent irrelevant information from misleading the large model.

[0085] For example, the user's original question is "Please tell me what famous tourist attractions are there in Beijing?" If the search result is that a search is required, the original input, search enhancement instructions, and retrieved external knowledge are spliced ​​together as the input of the large model, that is, the prompt is "Please tell me what famous tourist attractions are there in Beijing? Here are some references: [References]. Please select information related to the question from the references to answer the question." Since the beginning of the prompt ("Tell me what famous tourist attractions are there in Beijing?") has not changed, the large model does not need to re-encode this part of the content (the original question).

[0086] In order to further reflect the technical effect achieved by the unified paradigm active retrieval method provided by the present invention, the technical effect of the present invention is explained below in combination with the application example of the active retrieval judgment order (i.e., the activation order of each classifier) ​​generally used in the reasoning stage proposed by the inventor.

[0087] Table 1 shows the accuracy of different active retrieval methods in "judging whether user questions need to be retrieved" in various scenarios (intention perception, knowledge perception, time perception and self perception). The "present invention" in Table 1 refers to the method of active retrieval judgment sequence structure commonly used in the reasoning stage (such as Figure 2 (b)); FLARE (Forward-Looking ActiveREtrieval-augmented generation) is a forward-looking active retrieval augmented generation method; Self-RAG (Self-Reflective Retrieval-Augmented Generation) is a self-reflective retrieval augmented generation method; SKR (Self-Knowledge Guided Retrieval Augmentation) is a self-knowledge guided retrieval augmentation method.

[0088] Table 1. Comparison of active retrieval timing accuracy in large models with different parameters

[0089]

[0090] As can be seen from Table 1, no matter under the 7 billion parameter model or the 13 billion parameter model, the performance of the present invention in all perception scenarios is better than that of the other three methods (i.e., FLARE, Self-RAG, and SKR), which shows that the present invention can more accurately determine whether user questions need to be retrieved, thereby improving the response accuracy of the large model when users ask questions.

[0091] Table 2. Comparison of the effects of various active retrieval methods in downstream tasks

[0092]

[0093] It is worth noting that the value "A ( B% ) "A is the accuracy of a retrieval method on the corresponding dataset, and B% is the active retrieval ratio of the retrieval method on the corresponding dataset. For example, 39.17 (5.7%) means that in the Drop dataset, when the Self-RAG method is applied, the accuracy of the model is 39.17%; at the same time, the active retrieval ratio is 5.7%, that is, only 5.7% of the answers are solved with the help of active retrieval, and the remaining answers do not call retrieval.

[0094] Table 2 shows the performance (i.e., accuracy) of different retrieval methods on various data sets, such as reading comprehension, math word problems, question answering, etc. The "present invention" in Table 2 refers to the method of active retrieval judgment sequence formation commonly used in the reasoning stage (e.g. Figure 2 (b)); Drop is a dataset used to evaluate the model's ability to handle complex reading comprehension tasks; GSM8K is a dataset used to evaluate the model's ability to solve mathematical word problems; a dataset used to evaluate the model's ability to answer trivial question-and-answer questions; WQ (Wiki Questions) is a question-and-answer dataset based on Wikipedia; TAQA is a question-and-answer dataset for a specific field; FreshQA is a new and challenging question-and-answer dataset.

[0095] As can be seen from Table 2, no matter in the 7 billion parameter model or the 13 billion parameter model, the present invention shows excellent performance in multiple tasks, especially in terms of overall accuracy, which is significantly better than other retrieval methods (no retrieval, always retrieval, FLARE, Self-RAG and SKR). This shows that the present invention can optimize the number of retrieval calls and improve the accuracy of large model answers.

[0096] See also Figure 4 , is a structural diagram of an embodiment of the unified paradigm active retrieval system provided by the present invention.

[0097] The second aspect of the present invention provides a unified paradigm active retrieval system, which is used to implement the unified paradigm active retrieval method described in any one of the embodiments of the first aspect. The system includes:

[0098] The question acquisition module 11 is used to acquire the original question of the user;

[0099] Hidden layer data acquisition module 12, used for inputting the original problem into the large model and acquiring the hidden layer state data of the last layer of the large model;

[0100] The retrieval determination module 13 is used to input the hidden layer state data into a preset classifier combination to determine whether the original question needs to be retrieved, and obtain a retrieval determination result.

[0101] It should be noted that the unified paradigm active retrieval system provided in the second aspect of the present invention can implement all the processes of the unified paradigm active retrieval method described in the first aspect of the present invention. The functions of each module in the system and the technical effects achieved are respectively the same as the functions and technical effects achieved by the unified paradigm active retrieval method described in the first aspect of the present invention, and will not be repeated here.

[0102] An embodiment of the third aspect of the present invention also provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the unified paradigm active retrieval method described in any embodiment of the first aspect above.

[0103] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A unified paradigm active retrieval method, characterized in that: include: Get the user's original question; Input the original problem into the large model to obtain the hidden layer state data of the last layer of the large model; The hidden layer state data is input into a preset classifier combination to determine whether the original question needs to be retrieved, and a retrieval determination result is obtained.

2. The unified paradigm active retrieval method according to claim 1, characterized in that: The method further comprises: When the search result indicates that a search is required, a search is performed from a preset knowledge base according to the original question to obtain N corresponding knowledge texts; wherein N≥1; Combining the original question with all the knowledge texts to obtain a prompt text; The prompt text is input into the large model to obtain the answer result for the user.

3. The unified paradigm active retrieval method according to claim 1, characterized in that: Each classifier in the preset classifier combination is used to perform a corresponding binary classification task; wherein any two of the binary classification tasks do not overlap with each other.

4. The unified paradigm active retrieval method according to claim 1, characterized in that: The preset classifier combination includes: an intention-aware classifier, a knowledge-aware classifier, a time-aware classifier and a self-aware classifier.

5. The unified paradigm active retrieval method according to claim 1, characterized in that: The input of each classifier in the preset classifier combination is the hidden layer state data.

6. The unified paradigm active retrieval method according to claim 1, characterized in that: Each classifier in the preset classifier combination is enabled according to a preset order; or, All of the classifiers are enabled simultaneously.

7. The unified paradigm active retrieval method according to claim 4, characterized in that: The specific activation order of each classifier in the preset classifier combination is: Enabling the intent-aware classifier to determine whether the user has an intent to search; If there is an intention to search, the search judgment result is output as needing to search; otherwise, the knowledge-aware classifier is enabled to determine whether the original question needs to be answered with factual knowledge; If factual knowledge is not required to answer, the search decision result is output as no search is required; otherwise, the time-aware classifier is enabled to determine whether the answer to the original question changes over time; If it changes over time, the retrieval judgment result is output as needing retrieval; otherwise, the self-awareness classifier is enabled to determine whether the large model has the knowledge content required to answer the original question.

8. The unified paradigm active retrieval method according to claim 1, characterized in that: The training data of each classifier in the preset classifier combination is composed of a number of user questions; wherein each of the training data includes a plurality of corresponding positive samples and negative samples.

9. A unified paradigm active retrieval system, characterized in that: include: Question acquisition module, used to obtain the user's original question; A hidden layer data acquisition module, used to input the original problem into the large model and obtain the hidden layer state data of the last layer of the large model; The retrieval determination module is used to input the hidden layer state data into a preset classifier combination to determine whether the original question needs to be retrieved and obtain a retrieval determination result.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program; wherein, when the computer program is run, it controls the device where the computer-readable storage medium is located to execute the unified paradigm active retrieval method according to any one of claims 1 to 8.

Images