SoftTraderBackend/ServiceHandler/FinBertScoringService.cs

using Microsoft.ML.OnnxRuntime;
using Microsoft.ML.OnnxRuntime.Tensors;
using System.Globalization;
using System.Text;

namespace Backend.ServiceHandler
{
	public sealed class FinBertScoringService : IDisposable
	{
		private readonly InferenceSession _session;
		private readonly Dictionary<string, int> _vocab;
		private readonly string _inputIdsName;
		private readonly string _attentionMaskName;
		private readonly string _tokenTypeIdsName;
		private readonly string _outputName;

		private const int MaxLength = 512;
		private const int ClsId = 101;
		private const int SepId = 102;
		private const int UnkId = 100;

		// Label mapping from config.json: 0=positive, 1=negative, 2=neutral
		private static readonly string[] Labels = ["positive", "negative", "neutral"];

		public FinBertScoringService(IWebHostEnvironment env)
		{
			string modelDir = Path.Combine(env.ContentRootPath, "AIModels", "finbert");
			string modelPath = Path.Combine(modelDir, "model.onnx");
			string vocabPath = Path.Combine(modelDir, "vocab.txt");

			_vocab = LoadVocab(vocabPath);
			_session = new InferenceSession(modelPath);

			var inputNames = _session.InputMetadata.Keys.ToList();
			_inputIdsName = inputNames.First(n => n.Contains("input_id", StringComparison.OrdinalIgnoreCase));
			_attentionMaskName = inputNames.First(n => n.Contains("attention_mask", StringComparison.OrdinalIgnoreCase));
			_tokenTypeIdsName = inputNames.First(n => n.Contains("token_type", StringComparison.OrdinalIgnoreCase));
			_outputName = _session.OutputMetadata.Keys.First();
		}

		public (double Score, string Label) Score(string text)
		{
			var inputIds = Tokenize(text);
			int seqLen = inputIds.Count;

			var inputIdsTensor = new DenseTensor<long>(new[] { 1, seqLen });
			var attentionMaskTensor = new DenseTensor<long>(new[] { 1, seqLen });
			var tokenTypeIdsTensor = new DenseTensor<long>(new[] { 1, seqLen });

			for (int i = 0; i < seqLen; i++)
			{
				inputIdsTensor[0, i] = inputIds[i];
				attentionMaskTensor[0, i] = 1;
				tokenTypeIdsTensor[0, i] = 0;
			}

			var inputs = new List<NamedOnnxValue>
			{
				NamedOnnxValue.CreateFromTensor(_inputIdsName, inputIdsTensor),
				NamedOnnxValue.CreateFromTensor(_attentionMaskName, attentionMaskTensor),
				NamedOnnxValue.CreateFromTensor(_tokenTypeIdsName, tokenTypeIdsTensor)
			};

			using var results = _session.Run(inputs);
			var outputTensor = results.First().AsTensor<float>();
			float[] logits = outputTensor.ToArray();

			var probs = Softmax(logits);

			// Score: positive * 100 + neutral * 50 + negative * 0
			double score = Math.Round(probs[0] * 100 + probs[2] * 50, 2);

			int maxIdx = Array.IndexOf(probs, probs.Max());
			string label = Labels[maxIdx];

			return (score, label);
		}

		private List<int> Tokenize(string text)
		{
			var ids = new List<int> { ClsId };

			text = NormalizeText(text);
			var basicTokens = BasicTokenize(text);

			foreach (var token in basicTokens)
			{
				ids.AddRange(WordPieceTokenize(token));
			}

			ids.Add(SepId);

			if (ids.Count > MaxLength)
			{
				ids = ids.Take(MaxLength - 1).ToList();
				ids.Add(SepId);
			}

			return ids;
		}

		private static string NormalizeText(string text)
		{
			text = text.ToLowerInvariant();

			// Strip accents: NFD normalize then remove combining marks
			text = new string(text.Normalize(NormalizationForm.FormD)
				.Where(c => CharUnicodeInfo.GetUnicodeCategory(c) != UnicodeCategory.NonSpacingMark)
				.ToArray());

			var sb = new StringBuilder(text.Length);
			foreach (char c in text)
			{
				if (char.IsControl(c))
					sb.Append(' ');
				else
					sb.Append(c);
			}

			return sb.ToString();
		}

		private static List<string> BasicTokenize(string text)
		{
			var tokens = new List<string>();
			var current = new StringBuilder();

			foreach (char c in text)
			{
				if (char.IsWhiteSpace(c))
				{
					if (current.Length > 0)
					{
						tokens.Add(current.ToString());
						current.Clear();
					}
				}
				else if (char.IsPunctuation(c) || char.IsSymbol(c))
				{
					if (current.Length > 0)
					{
						tokens.Add(current.ToString());
						current.Clear();
					}
					tokens.Add(c.ToString());
				}
				else
				{
					current.Append(c);
				}
			}

			if (current.Length > 0)
				tokens.Add(current.ToString());

			return tokens;
		}

		private List<int> WordPieceTokenize(string token)
		{
			var ids = new List<int>();
			int start = 0;

			while (start < token.Length)
			{
				int end = token.Length;
				bool found = false;

				while (start < end)
				{
					string sub = start > 0
						? "##" + token[start..end]
						: token[start..end];

					if (_vocab.TryGetValue(sub, out int id))
					{
						ids.Add(id);
						found = true;
						start = end;
						break;
					}

					end--;
				}

				if (!found)
				{
					ids.Add(UnkId);
					start++;
				}
			}

			return ids;
		}

		private static float[] Softmax(float[] logits)
		{
			float max = logits.Max();
			var exps = logits.Select(l => MathF.Exp(l - max)).ToArray();
			float sum = exps.Sum();
			return exps.Select(e => e / sum).ToArray();
		}

		private static Dictionary<string, int> LoadVocab(string vocabPath)
		{
			var vocab = new Dictionary<string, int>();
			int index = 0;
			foreach (string line in File.ReadLines(vocabPath))
			{
				vocab[line.TrimEnd()] = index++;
			}
			return vocab;
		}

		public void Dispose()
		{
			_session.Dispose();
		}
	}
}