Computing N-dimensional sentiment using large-scale language models

Abstract

Techniques are provided for generating predictions associated with a set of input text. The text can be received. A machine learning model can be utilized to determine a set of classification probabilities for the text relative to a set of anchor points. A sentiment score indicative of the emotional content of the text can be determined based at least in part on a convex combination of the set of probabilities for the text. One or more predictions associated with the text can be generated.

Description

[Background technology]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS [1] This application is a continuation of PCT Application Serial No. PCT / US23 / 70896, filed July 25, 2023, entitled "COMPUTING N-DIMENSIONAL SENTIMENT USING A LARGE LANGUAGE MODEL," U.S. Provisional Application Serial No. 63 / 506,447, filed June 6, 2023, entitled "FROM A LARGE LANGUAGE MODEL TO THREE-DIMENSIONAL SENTIMENT," U.S. Provisional Application Serial No. 63 / 392,451, filed July 26, 2022, entitled "PSILOCYBIN THERAPY FOR TREATMENT RESISTANT DEPRESSION," and U.S. Provisional Application Serial No. 63 / 392,451, filed October 10, 2022, entitled "PSILOCYBIN THERAPY FOR TREATMENT RESISTANT DEPRESSION." This application claims priority to and benefit of U.S. Provisional Application Serial No. 63 / 414,769, entitled "PSILOCYBIN THERAPY FOR TREATMENT RESISTANT DEPRESSION," and PCT Application Serial No. PCT / US23 / 070857, filed July 24, 2023, entitled "PSILOCYBIN THERAPY FOR TREATMENT RESISTANT DEPRESSION," the entire disclosures of which are incorporated herein by reference in their entirety for all purposes.

[0002] [2] Recent advances in artificial intelligence are providing the building blocks necessary to quantify the emotions of human language. One form of human interaction important in mental health settings is the dialogue between therapists and patients. Therapeutic administration of psychedelics has shown significant potential in both historical accounts and recent clinical trials for the treatment of depression and other related mental and behavioral disorders. For example, recent studies have shown promising results when using psilocybin preparations in patients with treatment-resistant depression (TRD). However, while promising, such treatments may only work for a portion of the population, and early prediction of outcome is an important goal of treatment. Having the ability to quantify language spoken during therapy sessions on a scale using natural language processing (NLP) methods in combination with other methods could help bring unprecedented precision and rigor to the analysis of human language.

[0003] [3] Psychological descriptions of emotions can generally involve one of two approaches: a categorical approach or a dimensional approach. A categorical approach allows emotions to be sorted among a number of discrete categories with clear boundaries separating emotional states. In contrast, a dimensional approach to emotions attempts to define a continuous space of emotions within which traditional categorical emotions can be placed. While simple NLP methods based on a categorical approach to emotions, such as positive-negative emotion classifiers, are available, NLP models based on an n-dimensional approach to emotions can provide a more complete and nuanced measurement of language. [Brief explanation of the drawings]

[0004] [4] Various embodiments according to the present disclosure are described with reference to the drawings. [Figure 1] [5] Figure 1 illustrates an exemplary diagram of anchor points that may be utilized in accordance with various embodiments. [Figure 2A][6] Figure 2A shows an exemplary therapy session transcript showing the smoothed means of therapist and patient valence, arousal, and confidence scores for the session. [Figure 2B] Figure 2B shows a time series showing the smoothed means of therapist and patient valence, arousal, and confidence scores for the session. [Figure 2C] FIG. 2C illustrates an exemplary distribution for patients and therapists that may be utilized in accordance with various embodiments. [Figure 3] [7] Figure 3 illustrates an exemplary method that can be utilized to implement one or more aspects of various embodiments. [Figure 4] [8] Figure 4 illustrates an exemplary method that can be utilized to implement one or more aspects of various embodiments. [Figure 5] [9] Figure 5 illustrates components of an exemplary computing device that can be utilized in accordance with various embodiments. [Figure 6]

[10] Figure 6 illustrates an example environment for implementing aspects according to various embodiments. [Figure 7]

[11] Figure 7 illustrates components of another exemplary environment in which aspects of various embodiments may be implemented. DETAILED DESCRIPTION OF THE INVENTION

[0005]

[12] The three-dimensional valence-arousal-confidence (VAC) framework can be used to describe a wide variety of emotional states that represent the full range of human responses and also captures essential features of categorical emotions. According to various embodiments, a VAC model can be used to represent the sentiment of a given input text as a convex combination whose weights are obtained from the class scores of a zero-shot classifier built from a large-scale language model (LLM). Approaches using a VAC model can contribute to more accurate sentiment differentiation between sentences that are missed or misinterpreted by previous approaches that use only categorical approaches or low-dimensional models.

[0006]

[13] In the following description, various embodiments are described. For purposes of explanation, specific configurations and details are set forth in order to provide a thorough understanding of the embodiments. However, it will be apparent to those skilled in the art that the embodiments may be practiced without the specific details. Additionally, well-known features may be omitted or simplified so as not to obscure the described embodiments.

[0007]

[14] Major depressive disorder (MDD) is a debilitating illness that may affect one in six adults in their lifetime. MDD may be characterized by at least one depressive episode lasting at least two weeks and accompanied by significant changes in mood, cognition, and the ability to experience pleasure. While MDD can be effectively managed using psychotherapy and / or pharmacological treatment, some MDD patients may not respond to treatment despite multiple attempts. Such individuals are sometimes referred to as patients with treatment-resistant depression (TRD). While various treatment options exist, they are often deemed inadequate. Therefore, there is a need for the development of alternative treatment options with improved efficacy for TRD patients. Furthermore, tolerability of the condition or treatment may remain a significant challenge for such patients.

[0008]

[15] Psilocybin is a tryptamine alkaloid, and its potential as an effective antidepressant has been preliminarily studied in patients with life-threatening cancer, MDD, and TRD. In at least one psilocybin study using a 25 mg dose of COMP360, a set of participants experienced a reduction in depressive symptoms for as long as 12 weeks. While highly promising, these results indicate that a durable psilocybin response occurs in only a portion of the TRD population. Weeks may pass, potentially wasting a potential opportunity for additional treatment. While this example illustrates the application of the model in the therapeutic space, the model could be used for any other application related to quantifying language or in applications requiring understanding the emotion of such language.

[0009]

[16] Figure 1 illustrates an exemplary three-dimensional sentiment cube 100 that may be utilized in accordance with various embodiments. In at least some embodiments, the model may utilize a weighted average of nine "anchor" points within the three-dimensional cube to generate values ​​along three emotional dimensions: valence 110, arousal 120, and confidence 130. For n dimensions, where n is any positive integer, there are 2n anchor points within the n-cube at the vertices, as well as additional anchor points, such as an anchor point at the center of the n-cube. "Confidence," according to one or more embodiments, may be defined as the degree to which a portion of text conveys an emotional state of confidence and certainty, as opposed to a confidence metric provided by a machine learning classifier. The weights in the weighted average may be calculated from the numbers generated by the classifier for a set of classes corresponding to the various dimensions. Such an approach may be useful for predicting whether a participant will respond to a particular week in their treatment timeline. This architecture can be constructed from a geometric operation called a convex combination. According to this example, nine points 140 of a cube 100 centered at the origin can be considered: the eight corners (±1,±1,±1) and the center 150 at (0,0,0). While this example describes three dimensions, an n-dimensional model can be used for any number of emotional dimensions. The ability to quantify language in this way, above a two-dimensional model, can allow results to be more accurate than what a human therapist would come up with. Furthermore, because different people are likely to have different opinions about what a piece of text means, using techniques such as those described herein can help ensure more accurate objective results overall.

[0010]

[17] According to an exemplary embodiment, audio recordings of therapy sessions may be collected and transcribed into dialogue text through the use of natural language processing (NLP) techniques. The transcripts may then be parsed into individual "utterances" that are used to estimate the session emotions of the therapist and patient or participant using an emotion model. Such models may generate valence, arousal, and confidence scores for each utterance.

[0011]

[18] In at least some embodiments, audio may be parsed and tagged with speech. The speech may be processed using a model to predict the patient's tone. In some embodiments, the speech may be correlated with a score indicative of the patient's tone. The speech score may be compared to valence and arousal scores, or the speech score may be considered together with the valence and arousal scores. A zero-shot classifier built on a Bidirectional Autoregressive Transformer (BART) autoencoder and a Multi-Genre Natural Language Inference (MNLI) dataset may be used to calculate valence, arousal, and confidence scores.

[0012]

[19] For a given text u and anchor point p, a number wp(u) (called the weight of u in p) can be defined for each point. These numbers can satisfy the following properties: wp(u) ≥ 0, and

number

[0013]

[20] Using these weights, the 3D VAC emotion can be expressed as:

number

[0014]

[21] This can also be expressed as:

number

[0015]

[22] According to an exemplary embodiment, the quote "With beautiful love, heal yourself and always remember you are the medicine" may have the following anchor weight values ​​for the quote: w(1,1,1)=0.204866 w(1,1,1)=0.223669 w(1,1,1)=0.249116 w(1,1,1)=0.130423 w(1,1,1)=0.023099 w(1,1,1)=0.012893 w(1,1,1)=0.103748 w(1,1,1)=0.011839 w(1,1,1)=0.040327

[0016]

[23] In this case, the sentiment score can be calculated as follows:

number

number

[0017]

[24] Weights can vary depending on the number of anchor points, the number of dimensions, and the type of input being processed. According to an exemplary embodiment, the anchor weights wp(u) for strings of text u and anchor points p are calculated using a zero-shot classifier built using a BART autoencoder fine-tuned on the Multi-Genre Natural Language Inference (MNLI) dataset. The predictions provided by the classifier can be combined to generate the weights used in the model. The class selection itself is difficult to tune the hyperparameters for. The classes can be used to arrive at these classes and group them into "anchor lists."

[0018]

[25] According to an exemplary embodiment, input text may be passed to a BART-MNLI classifier, which may include, but is not limited to, the following classes: accepting, appropriate, aggressive, surprised, awe, comfortable, confident, in control, dazzled, repulsive, disruptive, detached, dejected, enthusiastic, angry, elated, weak, frustrated, submissive, inferior, joy, isolated, neutral, indifferent, distraught, tolerant, strong, relaxed, resistant, safe, serious, protective, soft, stoic, suffering, surprised, fearful, tolerant, peaceful, successful, victorious, and wise. Any number of classes may be selected, including more or fewer classes than those described herein.

[0019]

[26] For a text u, the value Pc(u) = the probability that u belongs to class c. This probability can be calculated under the assumption that c is in one of the classes and that the classes are mutually exclusive. The classes can be separated into lists, one for each of the nine anchor points p, using a frequency-inverse document frequency technique, according to an exemplary embodiment: (1,1,1)-Class=[Exaltation, Joy, Success, Victory, Powerful]. (1,1,-1)-Class=[Astonished, Awe, Dazzled, Enthusiastic, Astonished]. (1,-1,1)-Class=[Comfortable, Confident, Relaxed, Safe, Sensible] (1,-1,-1)-Class=[Accepting, Protective, Soft, Peaceful, Tolerant]. (-1,1,1)-Class=[Aggressive, Controlling, Antipathy, Destructive, Anger]. (-1,1,-1)-Class=[Frustrated, Submissive, Dismayed, Suffering, Frightened]. (-1,-1,1)-class=[appropriate, detached, resistant, severe, stoic]. (-1, -1, -1)-Class = [Dejected, Weak, Outnumbered, Isolated, Indifferent]. (0,0,0) - class=[neutral,tolerant].

[0020]

[27] The weight wp(u) is calculated as follows:

number

[0021]

[28] The choice of classes and how they are separated into anchor lists contributes to how the model scores a piece of text. The basic behavior that guides the intuition behind which classes should appear in an anchor list is that the higher a piece of text scores for a class in a given anchor list, the greater the anchor weight, resulting in a stronger pull of the text toward the anchor. Therefore, the classes that appear in a given anchor's list should reflect the sentiment that the anchor represents.

[0022]

[29] For example, if "joy" is in the anchor list for anchor (1,1,1) and a piece of text has a high score for the class "joy," the model will pull the text toward point (1,1,1) with a weight that includes the "joy" class score as an addend. In other words, a class in the (1,1,1) class list must be (1,1,1) word (or close to it). Similarly for other anchors.

[0023]

[30] Class lists can be compiled by collecting sentences spoken in therapy sessions and scoring them. Once scored, a large number of sentences for each emotional state corresponding to the anchor points can be collected and the most characteristic words for each sentence can be extracted.

[0024]

[31] The sentiment score of an utterance can depend on weights wp(u), and these weights can depend on probabilities Pc(u). For a piece of text u, the number Pc(u) can depend not only on c but on the entire set of classes. If different classes are used in the anchor list, the number Pc(u) will change even for an unchanged class c. At a surface level, it seems like every number Pc(u) must be recalculated from scratch, but by using the BART-MNLI zero-shot classifier, we can significantly accelerate the calculation of Pc(u), regardless of the other classes involved.

[0025]

[32] Inside the BART-MNLI zero-shot classifier, a number Lc(u) is calculated, independent of other classes. This number is sometimes called the logarithm of the implicature probability, also known as the "logit." The implicature probability is the prediction that text u is a member of class c without any information about other classes. Calculating the implicature probability is a task for which BART-MNLI is tailored upstream of its zero-shot application. Logits for all classes c planned for exploration and for all words and utterances u used in evaluating the model can be calculated in advance. In this way, it is possible to test as many choices of classes as possible, or different assignments of classes to anchor class lists, in a computationally efficient manner.

[0026] 2A and 2B illustrate an exemplary therapy session transcript 200 and time series 210 showing smoothed means of therapist and patient valence, arousal, and confidence scores (220, 230, 240), respectively, for the session. FIG. 2C illustrates an exemplary distribution 250 of patients and therapists that may be utilized according to various embodiments.

[0027] 34 In an exemplary therapy session 200, Gloria, a 30-year-old divorced mother of three, discusses her feelings of anxiety and dissatisfaction with her life. Gloria guides the conversation and expresses her feelings in her own words. According to the VAC model, the average emotions of the session are medium-high valence, neutral arousal, and low confidence for both the therapist and the patient. Figure 2C illustrates the overall distribution 250 of the patient's and therapist's valence, arousal, and confidence scores (270, 280, 290) in the exemplary therapy session 200, presented graphically as a probability density 240.

[0028]

[35] Figure 3 illustrates an exemplary method 300 that can be utilized to implement one or more aspects of various embodiments. According to an exemplary embodiment, text can be received 310. For example, an audio recording can be generated at the time of an initial integrative session between a patient or participant and a healthcare provider, such as a therapist. Using one or more NLP techniques, the recording can be transcribed into text.

[0029]

[36] According to one or more embodiments, a trained large-scale language model may be utilized to assist in scoring the utterance. For example, a machine learning model may analyze the text and determine a set of classification probabilities for the text relative to a set of anchor points 320.

[0030]

[37] Text can be scored or rated on three or more dimensions, including but not limited to valence, arousal, and confidence. In this way, sentiment can capture intensity rather than simply positivity or negativity. Sentiment scores for a piece of text can be distinguishable from attempts to infer the speaker's emotional state. For example, text stating "I love broccoli" may be scored as positive by a traditional sentiment model. However, if spoken in a harsh manner, the text may indicate a negative attitude toward broccoli.

[0031]

[38] A sentiment score indicating the emotional content of the text may be determined based at least in part on a convex combination of a set of text probabilities. 330 In some exemplary embodiments, a classifier may be utilized to classify the text. Such a classifier may be built on the BART autoencoder and the Multi-Genre Natural Language Inference (MNLI) dataset. The use of a model built on a larger model may enable the use of nearly unlimited freely available data for smaller data tasks.

[0032]

[39] One or more predictions associated with the text can be generated 340. For example, the predictions may correspond to a predicted response to an administered treatment or any other type of prediction associated with the set of text. For any process herein, unless otherwise stated, it should be understood that there may be additional, fewer, or alternative steps performed in a similar or alternative order, or in parallel, within various embodiments.

[0033]

[40] Figure 4 illustrates an exemplary method 400 that can be utilized to implement one or more aspects of various embodiments. According to an exemplary embodiment, a set of input text may be passed through a BART-MNLI classifier 410. Probability values ​​indicating the probability that a string of text in the set of input text is within a predetermined class may be calculated 420. A set of classes for the set of input text may be generated based on the calculated probability values ​​430. The set of classes may be separated into lists corresponding to one or more anchor points 440. Weights for the strings of text may be generated based at least in part on the lists 450.

[0034] As described in

[41] , input text may be passed to a BART-MNLI classifier that includes, but is not limited to, the following classes: accepting, appropriate, aggressive, surprised, awe, comfortable, confident, controlled, dazzled, repulsive, destructive, detached, dejected, eager, angry, elated, weak, frustrated, submissive, inferior, joy, isolated, neutral, indifferent, distraught, tolerant, powerful, relaxed, resistant, safe, serious, protective, soft, stoic, suffering, surprised, fearful, tolerant, peaceful, successful, victorious, and wise. The classes may be separated into lists, one list for each of the nine anchor points p, according to an exemplary embodiment: (1,1,1)-Class=[Exaltation, Joy, Success, Victory, Powerful]. (1,1,-1)-Class=[Astonished, Awe, Dazzled, Enthusiastic, Astonished]. (1,-1,1)-Class=[Comfortable, Confident, Relaxed, Safe, Sensible] (1,-1,-1)-Class=[Accepting, Protective, Soft, Calm, Tolerant] (-1,1,1)-Class=[Aggressive, Controlling, Antipathy, Destructive, Anger] (-1,1,-1)-Class=[Frustrated, Submissive, Dismayed, Suffering, Frightened] (-1,-1,1)-Class=[Appropriate, Separate, Resistant, Severe, Stoic] (-1, -1, -1)-Class = [Dejected, Weak, Outnumbered, Isolated, Indifferent] (0,0,0)-class=[neutral,tolerant]

[0035]

[42] Any number of classes may be selected, including more or fewer than those described herein.

[0036]

[43] Computing resources, such as servers, that can have software and / or firmware updated in such a manner typically include at least one set of standard components configured for general-purpose operation, although a variety of proprietary components and configurations can likewise be used within the scope of various embodiments. FIG. 5 illustrates components of an exemplary computing device 500 that can be utilized in accordance with various embodiments. As is known for computing devices, a computer has one or more processors 502, such as a central processing unit (CPU), graphics processing unit (GPU), etc., electronically and / or communicatively coupled to various components using various buses, traces, and other such mechanisms. The processor 502 can include memory registers 506 and cache memory 504 for holding instructions, data, etc. In this example, a chipset 514, which in some embodiments can include a northbridge and a southbridge, in cooperation with various system buses, can connect the processor 502 to components such as system memory 516 in the form of physical RAM or ROM, which can contain code for an operating system and various other instructions and data utilized for operation of the computing device. The computing device may also include or communicate with one or more storage devices 520, such as a hard drive, flash drive, optical storage device, etc., for persisting data and instructions similar to or in addition to those stored in the processor and memory. The processor 502 may also communicate with various other components via a chipset 514 and an interface bus (or graphics bus, etc.), which may include a communication device 524, such as a cellular modem or network card, media components 526, such as a graphics card and audio components, and a peripheral interface 530 for connecting peripheral devices, such as a printer, keyboard, etc.At least one cooling fan 532 or other such temperature regulation or reduction component may also be included, which may be driven by the processor or triggered by various other sensors or components on or remote from the device. Various other or alternative components and configurations known in the art for computing devices may be utilized.

[0037]

[44] In some embodiments, at least one processor 502 can retrieve data from physical memory 516, such as dynamic random access memory (DRAM) modules, via a coherency fabric. It should be understood that various architectures, including various selections, numbers, and arrangements of buses and bridges, may be utilized for such a computing device within the scope of various embodiments. Data in memory is managed and accessed by a memory controller, such as a DDR controller, via the coherency fabric. In at least some embodiments, data may be temporarily stored in a processor cache 504. The computing device 500 can also support multiple I / O devices using a set of I / O controllers connected via an I / O bus. There may be an I / O controller for each type of I / O device, such as a universal serial bus (USB) device, a data storage device (e.g., flash or disk storage), a network card, a peripheral component interconnect express (PCIe) card or interface 530, a communications device 524, a graphics or audio card 526, and a direct memory access (DMA) card, among other options. In some embodiments, components such as a processor, controller, and cache may be configured on a single card, board, or chip (i.e., a system-on-chip implementation), while in other embodiments, at least some of the components may be located in different locations, etc.

[0038]

[45] An operating system (OS) running on the processor 502 can help manage various devices that may be utilized to provide input to be processed. This may include, for example, utilizing associated device drivers to enable interaction with various I / O devices, which may be related to data storage, device communications, user interfaces, etc. Various I / O devices typically connect through various device ports and communicate with the processor and other device components via one or more buses. There may be specific types of buses that provide communication according to a particular protocol, such as Peripheral Component Interconnect (PCI) or Small Computer System Interface (SCSI) communications, among other such options. Communication may occur using registers associated with each port, including registers such as data-in and data-out registers. Communication may also occur using memory-mapped I / O, in which a portion of the processor's address space is mapped to a particular device and data is written directly to and from that portion of the address space.

[0039]

[46] Such devices may be used, for example, as servers in a server farm or data warehouse. Server computers often need to perform tasks outside the environment of the CPU and main memory (i.e., RAM). For example, a server may need to communicate with external entities (e.g., other servers) or process data using an external processor (e.g., a general-purpose graphical processing unit (GPGPU)). In such cases, the CPU may interface with one or more I / O devices. In some cases, these I / O devices may be dedicated hardware designed to fulfill a specific role. For example, an Ethernet network interface controller (NIC) may be implemented as an application-specific integrated circuit (ASIC) comprising digital logic operable to send and receive packets.

[0040] In an exemplary embodiment, a host computing device is associated with various hardware and software components and their respective configurations that facilitate the execution of I / O requests. One such component is an I / O adapter, which inputs and / or outputs data along a communication channel. In one aspect, the I / O adapter device can communicate as a standard bridge component to facilitate access between various physical and emulated components and the communication channel. In another aspect, the I / O adapter device can include an embedded microprocessor to enable the I / O adapter device to execute computer-executable instructions related to implementing management functions or managing one or more such management functions, or other computer-executable instructions related to the implementation of the I / O adapter device. In some embodiments, the I / O adapter device may be implemented using multiple separate hardware elements, such as multiple cards or other devices. The management controller can be configured to be electrically isolated from other components in the host device other than the I / O adapter device. In some embodiments, the I / O adapter device is attached external to the host device. In some embodiments, the I / O adapter device is internally integrated into the host device. Also in communication with the I / O adapter device may be an external communication port component for establishing a communication channel between the host device and one or more network-based services or other network-attached or directly-attached computing devices. Illustratively, the external communication port component may correspond to a network switch, sometimes known as a top-of-rack ("TOR") switch. The I / O adapter device may utilize the external communication port component to maintain a communication channel between the host device and one or more services, such as a health check service, a financial service, etc.

[0041]

[48] ​​I / O adapter devices may also communicate with basic input / output system (BIOS) components. BIOS components may contain non-transitory executable code, often referred to as firmware, that is executed by one or more processors and used to cause host device components to initialize and identify system devices such as video display cards, keyboards and mice, hard disk drives, optical disk drives, and other hardware. BIOS components may also contain or host boot loader software utilized to boot the host device. For example, in one embodiment, a BIOS component may contain executable code that, when executed by a processor, causes the host device to attempt to host pre-boot execution environment (PXE) boot software. Additionally, a BIOS component may include or benefit from hardware latches electrically controlled by the I / O adapter device. The hardware latches may restrict access to one or more aspects of the BIOS component, such as controlling modification or configuration of executable code held within the BIOS component. BIOS components may be connected to (or communicate with) numerous additional computing device resource components, such as processors, memory, etc. In one embodiment, such computing device resource components may be physical computing device resources that communicate with other components via communication channels. The communication channel may correspond to one or more communication buses over which the components of the bare-metal host device communicate, such as a shared bus (e.g., a processor bus, a memory bus), a point-to-point bus such as a PCI or PCI Express bus, etc. Other types of communication channels, media, buses, or protocols (e.g., an Ethernet communication protocol) may also be utilized. Additionally, in other embodiments, one or more of the resource components of the computing device may be virtualized hardware components emulated by the host device.In such embodiments, the I / O adapter device may implement a management process in which the host device is configured with physical or emulated hardware components based on various criteria. Resource components of the computing device may communicate with the I / O adapter device via a communication channel. Additionally, the communication channel may connect the PCI Express device to the CPU through a northbridge or host bridge, among other such options.

[0042] Communicating with the I / O adapter device via the communication channel may be one or more controller components for managing a hard drive or other form of memory. An example of a controller component may be a SATA hard drive controller. Similar to a BIOS component, the controller component may include or benefit from hardware latches electrically controlled by the I / O adapter device. The hardware latches may restrict access to one or more aspects of the controller component. Illustratively, the hardware latches may be controlled together or independently. For example, the I / O adapter device may selectively close hardware latches for one or more components based on a trust level associated with a particular user. In another example, the I / O adapter device may selectively close hardware latches for one or more components based on a trust level associated with a creator or distributor of executable code executed by the I / O adapter device. In a further example, the I / O adapter device may selectively close hardware latches for one or more components based on a trust level associated with the component itself. A host device may also include additional components that communicate with one or more of the illustrative components associated with the host device. Such components may include devices such as one or more controllers combined with one or more peripheral devices, such as hard disks or other storage devices. Additionally, additional components of the host device may include another set of peripheral devices, such as a graphics processing unit ("GPU"). The peripheral devices may also be associated with hardware latches to restrict access to one or more aspects of the component. As mentioned above, in one embodiment, the hardware latches may be controlled together or independently.

[0043] As noted above, different approaches may be implemented in various environments according to the described embodiments. For example, FIG. 6 illustrates an example environment 600 for implementing aspects according to various embodiments. As will be appreciated, while a web-based environment is used for illustrative purposes, different environments may be used, as needed, to implement various embodiments. The system includes an electronic client device 602, which may include any suitable device operable to send and receive requests, messages, or information over a suitable network 604 and return information to a user of the device. Examples of such client devices include personal computers, mobile phones, handheld messaging devices, laptop computers, set-top boxes, personal data assistants, e-book readers, etc. Examples of such recipients or users may include healthcare providers, including therapists, or patients. The network may include any suitable network, including an intranet, the Internet, a cellular network, a local area network, or any other such network, or any combination thereof. The components used in such a system may depend at least in part on the type of network and / or environment selected. Protocols and components for communicating over such networks are well known and will not be described in detail herein. Communication over the network may be enabled via wired or wireless connections, and combinations thereof. In this example, the network includes the Internet and the environment includes a web server 606 for receiving requests and providing content in response thereto, although other networks may use alternative devices that serve similar purposes, as will be apparent to those skilled in the art.

[0044] The exemplary environment includes at least one application server 608 and a data store 610. It should be understood that there may be several application servers, layers, or other elements, processes, or components, which may be chained or otherwise configured, that may interact to perform tasks, such as retrieving data from an appropriate data store. As used herein, the term “data store” refers to any device or combination of devices capable of storing, accessing, and retrieving data and may include any combination and number of data servers, databases, data storage devices, and data storage media in any standard distributed or clustered environment. The application server 608 may integrate with the data store 610 as needed to run aspects of one or more applications for client devices and may include any suitable hardware and software for handling much of the data access and business logic for the applications. The application server may cooperate with the data store to provide access control services and generate content, such as text, graphics, audio, and / or video, that is transmitted to users, which in this example may be served to users by the web server 606 in the form of HTML, XML, or another suitable structured language. All request and response processing, as well as delivery of content between client device 602 and application server 608, may be handled by web server 606. It should be understood that the structured code described herein may be executed on any suitable device or host machine, as described elsewhere herein, and thus a web server and application server are not required and are merely exemplary components.

[0045]

[52] The data store 610 may include several separate data tables, databases, or other data storage mechanisms and media for storing data related to particular aspects. For example, the illustrated data store includes mechanisms for storing biomarker data (e.g., production data) 612 and user information 616, which may be used to provide content to the production side. The data store is also shown to include mechanisms for storing log or session data 614. It should be understood that there may be many other aspects that may need to be stored in the data store, such as page image information and access rights information, which may be stored in any of the mechanisms listed above or in additional mechanisms within the data store 610, as appropriate. The data store 610, through its associated logic, is operable to receive instructions from the application server 608 and, in response, retrieve, update, or otherwise process data. In one example, a user may submit a search request for a certain type of item. In this case, the data store may access user information to verify the user's identity and may access catalog details to obtain information about that type of item. The information can then be returned to the user, such as through a patient or therapist portal, including biomarker and diagnostic data accessible through web pages that the user can view via a browser on the user device 602. Information regarding particular items of interest can be viewed in dedicated pages or windows in the browser.

[0046]

[53] Each server typically includes an operating system that provides executable program instructions for the general management and operation of that server, and typically includes a computer-readable medium that stores instructions that, when executed by the server's processor, enable the server to perform its intended functions. Suitable implementations for server operating systems and general functionality are known or commercially available, and are readily implemented by those of skill in the art, especially in light of the disclosure herein.

[0047]

[54] In one embodiment, the environment is a distributed computing environment utilizing several computer systems and components interconnected through communications links using one or more computer networks or direct connections. However, those skilled in the art will appreciate that such a system may operate equally well in a system having fewer or more components than illustrated in Figure 6. Therefore, the depiction of system 600 in Figure 6 should be construed as being exemplary in nature and not limiting on the scope of the present disclosure.

[0048] FIG. 7 illustrates an exemplary environment 700 in which aspects of various embodiments can be implemented. In this example, a user can utilize a client device 702 to submit requests to a multi-tenant resource provider environment 706 across at least one network 704. The client device can include any suitable electronic device operable to send and receive requests, messages, or other such information over an appropriate network and return information to the device user. Examples of such client devices include personal computers, tablet computers, smartphones, notebook computers, etc. The at least one network 704 can include any suitable network, including an intranet, the Internet, a cellular network, a local area network (LAN), or any other such network or combination, and communication through the network can be enabled via wired and / or wireless connections. The resource provider environment 706 can include any suitable components for receiving requests and returning information or performing actions in response to those requests. As an example, the provider environment can include a web server and / or application server for receiving and processing requests and then returning data, web pages, video, audio, or other such content or information in response to the request.

[0049]

[56] In various embodiments, the provider environment may include various types of resources that can be utilized by multiple users for a variety of different purposes. As used herein, computing resources and other electronic resources utilized in a network environment may be referred to as “network resources.” These may include, for example, servers, databases, load balancers, routers, etc., which may perform tasks such as receiving, transmitting, and / or processing data and / or executable instructions. In at least some embodiments, all or a portion of a given resource or set of resources may be allocated to a particular user or allocated for a particular task, at least for a determined period of time. The sharing of these multi-tenant resources from the provider environment is often referred to as resource sharing, web services, or “cloud computing,” among other such terms, depending on the particular environment and / or implementation. In this example, the provider environment includes multiple resources 714 of one or more types. These types may include, for example, application servers operable to process instructions provided by users, or database servers operable to process data stored in one or more data stores 716 in response to user requests. As is known for such purposes, users may also reserve at least a portion of data storage in a given data store. Methods for allowing users to reserve various resources and resource instances are well known in the art, and therefore a detailed description of the entire process and a description of all possible components will not be discussed in detail herein.

[0050] In at least some embodiments, a user desiring to utilize a portion of the resources 714 can submit a received request to an interface layer 708 of the provider environment 706. The interface layer can include an application programming interface (API) or other exposed interface that allows a user to submit a request to the provider environment. The interface layer 708 in this example can also include other components, such as at least one web server, routing component, load balancer, etc. When a request to provision a resource is received against the interface layer 708, information for the request can be directed to a service manager 710 or other such system, service, or component configured to manage user accounts and information, resource provisioning and usage, and other such aspects. The service manager 710 receiving the request can perform tasks such as authenticating the identity of the user submitting the request and determining whether the user has an existing account with the resource provider, and the account data can be stored in at least one account data store 712 within the provider environment. The user can provide any of a variety of types of credentials to authenticate the user's identity to the provider. These credentials may include, for example, a username and password pair, biometric data, a digital signature, or other such information. The provider can verify this information against information stored for the user. If the user has an account with the appropriate permissions, status, etc., the resource manager can determine whether there are sufficient resources available to meet the user's request, and if so, can provision resources for the user's use for the amount specified by the request, or otherwise grant access to the corresponding portion of those resources.This amount may include, for example, capacity to process a single request or perform a single task, a specified period of time, or a repeatable / renewable period of time, among other such values. If the user does not have a valid account with the provider, the user account does not allow access to resources of the type specified in the request, or another such reason is to prevent the user from gaining access to such resources, a communication may be sent to the user to enable the user to create or modify an account or change the resources specified in the request, among other such options.

[0051]

[58] Once a user is authenticated, their account is verified, and resources are allocated, the user can utilize the allocated resources for a specified capacity, data transfer amount, duration, or other such value. In at least some embodiments, the user may provide a session token or other such proof with subsequent requests to enable those requests to be processed on that user session. The user may receive a resource identifier, specific address, or other such information that can enable the client device 702 to communicate with the allocated resources without having to communicate with the service manager 710, at least until such time as relevant aspects of the user account change, the user is no longer authorized to access the resources, or another such aspect changes.

[0052] 59 The service manager 710 in this example (or another such system or service) may also function as a virtual layer of hardware and software components that handles control functions in addition to management actions, such as those that may include provisioning, scaling, replication, etc. The resource manager may utilize dedicated APIs in the interface layer 808, each of which may be provided to receive a request for at least one specific action to be performed on a data environment, such as provisioning, scaling, cloning, or hibernating an instance. Upon receiving a request for one of the APIs, the web services portion of the interface layer may parse or otherwise analyze the request to determine the steps or actions necessary to act on or process the call. For example, a web services call may be received that includes a request to create a data repository.

[0053]

[60] The interface layer 708 in at least one embodiment includes a scalable set of user-facing servers capable of offering various APIs and returning appropriate responses based on the API specifications. The interface layer can also include at least one API service layer, which in one embodiment consists of stateless replicated servers that handle externally facing user APIs. The interface layer can be responsible for web service front-end functions such as authenticating users based on credentials, authorizing users, throttling user requests to the API server, validating user input, and marshalling or unmarshalling requests and responses. The API layer can also be responsible for reading and writing database configuration data to and from a management data store in response to API calls. In many embodiments, the web service layer and / or API service layer will be the only externally visible components, or the only components visible to and accessible by users of the control service. The servers in the web service layer are stateless and can be horizontally scaled as known in the art. The API servers, as well as the persistent data store, can be distributed across multiple data centers within a region, for example, so that the servers are resilient to a single data center failure.

[0054]

[61] Various embodiments may further be implemented in a wide variety of operating environments and, in some cases, may include one or more user computers or computing devices that may be used to operate any of several applications. User or client devices may include any of several general-purpose personal computers, such as desktop or laptop computers running standard operating systems, as well as cellular, wireless, and handheld devices capable of running mobile software and supporting several networking and messaging protocols. Such systems may also include several workstations running any of various commercially available operating systems and other known applications for purposes such as development and database management. These devices may also include other electronic devices, such as dummy terminals, thin clients, gaming systems, and other devices capable of communicating over a network.

[0055]

[62] Most embodiments utilize at least one network, familiar to those skilled in the art, to support communications using any of a variety of commercially available protocols, such as TCP / IP, FTP, UPnP, NFS, and CIFS. The network may be, for example, a local area network, a wide area network, a virtual private network, the Internet, an intranet, an extranet, a public switched telephone network, an infrared network, a wireless network, or any combination thereof. In embodiments utilizing a web server, the web server may run any of a variety of server or middle-tier applications, including an HTTP server, an FTP server, a CGI server, a data server, a Java server, and a business application server. The server may also be capable of executing programs or scripts in response to requests from user devices, such as by running one or more web applications, which may be implemented as one or more scripts or programs written in any programming language, such as Java, C, C#, or C++, or any scripting language, such as Perl, Python, or TCL, and combinations thereof. The server may also include database servers, including, but not limited to, those commercially available from Oracle, Microsoft, Sybase, and IBM.

[0056]

[63] The environment can include various data stores and other memory and storage media as described above. These can reside in a variety of locations, such as on storage media local to (and / or resident within) one or more of the computers, or remote from any or all of the computers via a network. In a particular set of embodiments, information can reside within a storage area network ("SAN") familiar to those skilled in the art. Similarly, any files necessary to perform functions ascribed to a computer, server, or other network device can be stored locally and / or remotely, as appropriate. Where the system includes computerized devices, each such device can include hardware elements that can be electrically coupled via a bus, including, for example, at least one central processing device (CPU), at least one input device (e.g., a mouse, keyboard, controller, touch-sensitive display element, or keypad), and at least one output device (e.g., a display device, printer, or speaker). Such a system may also include one or more storage devices, such as disk drives, optical storage devices, and solid-state storage devices such as random access memory (RAM) or read-only memory (ROM), as well as removable media devices, memory cards, flash cards, etc. Such devices may also include computer-readable storage medium readers, communication devices (e.g., modems, network cards (wireless or wired), infrared communication devices), and working memory, as described above. The computer-readable storage medium readers may be configured to connect to or receive computer-readable storage media representing remote, local, fixed, and / or removable storage devices, as well as storage media for temporarily and / or more permanently containing, storing, transmitting, and retrieving computer-readable information.

[0057]

[64] The systems and various devices also typically include several software applications, modules, services, or other elements located within at least one working memory device, including an operating system and application programs such as client applications or web browsers. It should be understood that alternative embodiments may have numerous variations from those described above. For example, customized hardware may also be used, and / or particular elements may be implemented in hardware, software (including portable software such as applets, etc.), or both. Additionally, other computing devices, such as network input / output devices, may be used. Storage media and other non-transitory computer-readable media for containing code or portions of code may include any suitable media known or used in the art, such as, but not limited to, volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules or other data, including RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVDs) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and that can be accessed by a system device. Based on the disclosure and teachings provided herein, one of ordinary skill in the art will appreciate other ways and / or methods for implementing the various embodiments.

[0058]

[65] The specification and drawings are, therefore, to be regarded in an illustrative rather than a restrictive sense. It will, however, be evident that various modifications and changes may be made therein without departing from the scope and broader spirit of the invention as set forth in the appended claims.

Claims

1. 1. A computer-implemented method comprising: Receiving a text; using a machine learning model to determine a set of classification probabilities of the text for a set of anchor points; and determining a sentiment score indicative of emotional content of the text based at least in part on a convex combination of the set of anchor points using the set of probabilities of the text.

2. The computer-implemented method of claim 1 , wherein the classification probabilities are calculated using the output of a large-scale language model.

3. The computer-implemented method of claim 2 , wherein the large-scale language model is fine-tuned on a multi-genre natural language inference (MNLI) dataset.

4. The computer-implemented method of claim 1 , further comprising: generating one or more predictions associated with the text.

5. The computer-implemented method of claim 4 , wherein the one or more predictions relate to a response to an administered treatment for treatment-resistant depression.

6. The computer-implemented method of claim 1 , wherein the sentiment scores include an arousal score, a valence score, and a confidence score for each portion of the text.

7. 1. A computing system comprising: a computing device processor; a memory device containing instructions that, when executed by the computing device processor, cause the computing system to: Receive a text and using a machine learning model to determine a set of classification probabilities of the text for a set of anchor points; a computing system that enables determining a sentiment score indicative of emotional content of the text based at least in part on a convex combination of the set of anchor points using the set of probabilities of the text.

8. The computing system of claim 7 , wherein the classification probabilities are calculated using the output of a large-scale language model.

9. The computing system of claim 8 , wherein the large-scale language model is fine-tuned on a multi-genre natural language inference (MNLI) dataset.

10. The instructions, when executed by the computing device, cause the computing system to further: The computing system of claim 7 , enabled to generate one or more predictions related to the text.

11. The computing system of claim 7 , wherein the sentiment scores include an arousal score, a valence score, and a confidence score for each portion of the text.

12. The computing system of claim 10 , wherein the one or more predictions are related to a response to an administered treatment for treatment-resistant depression.

13. The instructions, when executed by the computing device processor, cause the computing system to: passing the text through a classifier; calculating a probability value indicative of the probability that a string of text in said text is within one of one or more classes; separating the one or more classes into a set of lists corresponding to the set of anchor points; The computing system of claim 7 , further comprising: a computing system configured to generate weights for individual strings of text based at least in part on the set of lists.

14. A non-transitory computer-readable medium storing instructions that, when executed by at least one processor, cause the one or more processors to: Receive the text, using a machine learning model to determine a set of classification probabilities of said text for a set of anchor points; a non-transitory computer-readable medium for determining a sentiment score indicative of emotional content of the text based at least in part on a convex combination of the anchor points using the set of probabilities for the text;

15. The non-transitory computer-readable medium of claim 14 , wherein the sentiment score is determined using a classifier built from a large-scale language model.

16. 16. The non-transitory computer-readable medium of claim 15, wherein the large-scale language model is fine-tuned on a multi-genre natural language inference (MNLI) dataset.

17. The instructions, when executed by the one or more processors, further cause the one or more processors to: The non-transitory computer-readable medium of claim 14 , further comprising: generating one or more predictions associated with the text.

18. The non-transitory computer-readable medium of claim 14 , wherein the affective scores include an arousal score, a valence score, and a confidence score for each portion of the text.

19. 20. The non-transitory computer-readable medium of claim 17, wherein the one or more predictions relate to a response to a therapy administered for treatment-resistant depression.

20. The instructions, when executed by the at least one processor, cause the one or more processors to further: passing the text through a classifier; calculating a probability value indicative of the probability that a string of text in the set of text is within one of one or more classes; separating the one or more classes into a set of lists corresponding to the set of anchor points; The non-transitory computer-readable medium of claim 14 , further comprising generating weights for individual strings of text based at least in part on the set of lists.

Images