ååä¸ã¯è¨èªå¦ç¿æ¯æ´ã»æ©æ¢°ç¿»è¨³ã®é²æå ±åãããããæ°ããç 究ã§åèªè²ã èªåã§é²ãã¦ããã®ã§ã話ãèãã®ã¯æ¥½ããã®ã ããå¹´å ã«å®é¨çµæãåºãã®ãï¼ãã¨ããä¸å®ãããã¯ãã¨ãªããããæ¬å½ã¯ããããªç· ãåãã«è¿½ããããããªæãã§ã¯ãªãããã£ããã§ããã¨ããã¨æãã®ã ããç· ãåãã«è¿½ããã¦ããæ¹ãé²ãã¨ããä¸æè°ï¼ããããã¾ãä¸æè°ã§ã¯ãªããï¼ã
ãæ¼ã¯è«æç´¹ä»ã
- Tang et al. Document Modeling with Gated Recurrent Neural Network for Sentiment Classification. EMNLP 2015.
ãç´¹ä»ãã¦ããããææ 極æ§åæã§ã¯ãé å¼µã£ã¦ç´ æ§ã¨ã³ã¸ãã¢ãªã³ã°ãã SVM ã¨ãé©å½ã«ä½ã£ãç³ã¿è¾¼ã¿ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ãåããããã®æ§è½ã«ãªããããã®ã ãããªã«ã¬ã³ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ï¼æ®éã® RNN ãããLSTMããã㦠LSTM ã®ç¹æ®ç³»ã§ãã GRNN ã¾ã§ï¼ã使ãã¨ãããè¶ ãããã¾ãããã¨ãã話ã
深層ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã表ç¾å¦ç¿ãä½ãã®ã¿ã¹ã¯ã«é©ç¨ãã¾ãããã¨ããã®ã¯ãããããè ¹ãã£ã±ãã®äººãå¤ããã¨æãã®ã ããæ¦å¨ãæã£ã¦ããã®ã§ãææ 極æ§åæã«ã¯ããããä½ãåããã°ããã®ãï¼ãã¨ãããã¨ãããä¸åº¦èãããã
ãã¨ãã¨ããã®åé㧠state-of-the-art ã®ä¸ã¤ã¨ããã¦ããã®ã¯ãtetsu-na ããã®ãé ãå¤æ°ãæã¤æ¡ä»¶ä»ã確çå ´ã«ããä¾åæ§é æ¨ã®æ¥µæ§åæãã§ã極æ§ï¼ãã¸ãã£ãããã¬ãã£ãï¼ã¯é¨åæ¨ã«ã¤ããæï¼ï¼æ§ææ¨ï¼ã®æ¥µæ§ã¯ãããããè¨ç®ãããã¨ããç´è¦³ã¨ãé¨åæ¨ã«ã¯æ¥µæ§ãä»ä¸ããã¦ããªããã¨ããç¾å®ããã¾ãã¤ãªãã ãé常ã«åªããç 究ã§ãã£ãããã®å¾ãSocher ãã recursive neural network ãç¨ãã¦å帰çã«é¨åæ¨ã®æ¥µæ§ãè¨ç®ãããããã¬ã¼ãºï¼é¨åæ¨ï¼ã«æ¥µæ§ãä»ä¸ããã Stanford Sentiment Treebank ãç¨ãã¦å¦ç¿ãããããããåºæ¬çã«ã¯ tetsu-na ããã®ç 究ã§ããã®è¾ºãã®ä¸é£ã®ç 究ã¯å å«ããã¦ããããã«æãã
ãããã®ç 究ãæ§ææ¨ï¼åã«ç³»åãäºåæ¨ã«å¤æããã ãã®ãã®ããããï¼ã使ã£ã¦æ¥µæ§åæããããã¨ããã®ã¯ãããæå³ãããã§ããã話ã ãªã¼ã¨æãã®ã ããããã¨éåãã®è©±ãã¤ã¾ããæ§ææ å ±ãªã©ä½¿ãããã»ã¼åèªã®æ å ±ã®ã¿ã§åé¡ãããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã®ææ³ã§ãé«ã精度ãåºãããããã®ã¯ã©ããããã¨ã ãããï¼
ãªã«ã¬ã³ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ï¼LSTM çå«ãï¼ã¯ãé·è·é¢ã®ä¾åé¢ä¿ãæãããã¨ãã§ããã®ã§ãããããã¾ããããã¨ããã®ã¯ãæããè¦ãã¦ãããªãã¨ãããªãã®ã¯æ¨æ§é ãã®ãã®ã§ã¯ãªãã極æ§ã®å転ã«å¿ è¦ãªæ å ±ã ãã§ããããã¸ãã£ãããã¬ãã£ãã«å転ãããããããã¯ãã®éãã¾ãã¯ãã¬ãã£ãã»ãã¸ãã£ãããã¥ã¼ãã©ã«ã«ãããã¨ãã£ãæå³ã®è¨ç®ãããã«ã¯ããããªåèªãåºããã¨ãããã¨ããããªã®ãã¨ãããã¨ãããããªãï¼è¨èªã«ãã£ã¦ãå転ã®æ¼ç®åãã¿ã¼ã²ããã®å³ã«åºããå·¦ã«åºããã¯éãã®ã§ãåæ¹åã®ãªã«ã¬ã³ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ããããã¨ãã£ããã¤ãã¼ãã§ã³ã¸ã¯ããããï¼ã
ä¸æ¹ãç³ã¿è¾¼ã¿ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã®ãããªå±æçãªæ å ±ãç³ã¿è¾¼ããã¨ã§ãé«ã精度ãåºãããã¨ããã®ã¯ããããã極æ§ã®å転ã®ããã«è¨æ¶ãã¦ããã¹ããã¨ã¯ãé ãé¢ããã¨ããã«åºã¦ããããã§ã¯ãªãã決ããããé©å½ãªã¦ã£ã³ãã¦å¹ ã®ä¸ã ãè¦ãã°ï¼ããã ãè¨æ¶ãã¦ããã°ï¼ãããããããããã2ã¤ã®åèªã®çµã¿åãããè¦ããããã§ã極æ§æä½ã¯å¯è½ã§ããããã¨ã¯ããããã¡ãã¡è¤æ°è¦ã¦æçµçãªå¤æãã§ããããã«ãªã£ã¦ããã°ãããã¨ãããã¨ãªã®ãããããªãããã¡ã®ç 究室ã§ãä»å¹´ Stacked denoising Autoencoder ãç¨ãã¦æ¥µæ§åé¡ã®ç 究ããã¦ãã¦ãï¼LSTM ãªã©ä½¿ããã¨ãï¼æã£ã以ä¸ã«ç²¾åº¦ãé«ãã®ã§ãæ¬å½ã«ãããï¼ãã¨æã£ã¦ããã®ã ãã極æ§åé¡ã¯ãããªã«æ¨æ§é ãè¦ãã»ã©ã®ã¿ã¹ã¯ã§ã¯ãªããã¨ãããã¨ãªã®ã ãããï¼æ§ææ¨ãèªåã§ä»ä¸ãã¦ãããããã®èª¤ãã®å½±é¿ãåãã¦ãã¾ããï¼ã
極æ§åæãªããã¯ããã«ãæ¨æ§é ãå¿ è¦ãããªã¿ã¹ã¯ã ã¨æã£ã¦ãã¦ããããæ¨æ§é ã使ã£ã¦ç²¾åº¦ãé«ããªã£ã¦ãã話ã ã¨æã£ã¦ããã®ã ããæ¬å½ã«ï¼ç¾å¨ãããã®ç²¾åº¦ã®ï¼æ§æ解æãå¿ è¦ãã¤æå¹ãªã¿ã¹ã¯ã¯ä½ãï¼ãã¨ãããã¨ã¯ãã£ããèããæ¹ãããã®ãããããªãã
åå¾ã¯ãç 究室ã¤ã³ã¿ã¼ã³ã·ãããã¨é¡ãã¦æ¯é±ç 究室ã®åå¼·ä¼ã»ç 究ä¼ã«åå ãã¦ããã¦ãã B3 ã®äººãã¡ã®ç 究çºè¡¨ãMMDAgent ãç¨ããæ¥éãã£ã³ãã¹ã®æ¡å ã¨ã¼ã¸ã§ã³ãã®éçºã«ã¤ãã¦è©±ãã¦ãããä»ãä»è¨èªå¦çå¦ä¼ã®å¹´æ¬¡å¤§ä¼ã«åãã¦ãã¼ã¿ä½æãã¦ããå 容ã«ã¤ãã¦ãå ±åãã¦ããããã½ã¼ã·ã£ã«ã¡ãã£ã¢è§£æåå¼·ä¼ã®äººãã¡ã¯æ¯é±è©±ãèãã¦ããããããã§ã¯ãªã人ãã¡ã«ã¯åã®ãæ«é²ç®ã§ãããB3 ã®äººãã¡ãã©ãèãã¦ãããã¯åãããªãããåå ãã¦ããç 究室ã®ã¡ã³ãã¼ããã¯å²ã¨è©ä¾¡ãé«ãåå¼·ä¼ãªã®ã§ãæ¥å¹´ã©ããããã¾ãèããããåºç¤åå¼·ä¼ã«ããã¨ãB3 ã®äººãã¡ã®ã¢ããã¼ã·ã§ã³ã¯é«ãããç 究室å¦çã®ã¢ããã¼ã·ã§ã³ãä½ãï¼ç 究ãã¹ã¿ã¼ããã¦ããã®ã§ãå½ç¶ã ãï¼ããã¾ã模ç¯ã«ãªããªãã®ã§ããã¯ãé©åãªé£æ度ã»åéã®ãµãã¿ã¹ã¯ã«ãã¦ãã£ã¦ãããã®ãããã®ããªãã